Not a chip. A crystal. Someone grows a boule of indium phosphide, slices it into wafers, polishes them flat enough that a laser structure can be grown on top, and ships them to the epitaxy houses. Those epiwafers become lasers. Those lasers become 800G and 1.6T optical modules. Those modules are the only reason a GPU in one rack can talk to a GPU in another rack fast enough to matter.

The market has spent two years pricing the modules. It has spent about two days pricing the crystal.

On Thursday July 30, after the close, AXT reported a quarter that broke a three year pattern. Revenue of $47.6 million against roughly $34.8 million of consensus, up about 165 percent year on year. Gross profit of $21.4 million. Operating profit of $10.4 million. The stock rose 27 percent that session and traded up another 24 percent after hours.

I want to do three things. Walk through what the quarter said. Then explain the one mechanism that actually determines AXT's reported revenue, which is not demand and which almost nobody modelling this company has right. Then put the price into the context that a two day chart hides.

What the quarter said

Revenue has moved in a step, not a drift. Six quarters of history and the guided September quarter, which the company put at roughly $66 million on the call:

The margin line matters more than the revenue line, and it is the reason this is worth your time at all.

Fifteen months ago this business sold product below cost. Gross margin was negative 6.4 percent. It printed 44.9 percent in the June quarter. That is a fifty point swing in five quarters and none of it came from a price increase. It came from volume moving through a fixed cost base.

A crystal growing operation is close to a fixed cost business. The furnaces run whether they are full or not. The difference between running them at sixty percent and ninety five percent is most of the income statement. Which is why the bottom line flipped so violently:

One correction to the headline before we go further. Diluted EPS was $0.17, but core EPS is closer to $0.11. About a third of pre-tax income is treasury yield on the cash from the April raise, not operations. Interest on a bank balance is real money and it is not the business, and it will not scale with wafer volume.

A second one. Minority interests take 14.6 percent of everything from here. AXT does not own all of Tongmei, the Chinese subsidiary that grows the crystal. When you read consolidated profit, roughly a seventh of it belongs to somebody else.

And a third thing you cannot find in any filing. AXT's filed segment disclosure has exactly two buckets, Substrates and Raw materials and other. The product level split exists only on the earnings call, and it shows how concentrated this company has become:

Two thirds of AXT is now one product line. If you are reading the 10-Q you cannot see that.

The mechanism nobody is modelling

Here is the part that changes how you read every AXT quarter, including this one.

AXT's reported revenue is gated by Chinese export permits, not by customer demand.

On February 4, 2025, China's Ministry of Commerce issued Announcement No.10, placing indium phosphide substrates under dual-use export licensing as control code 3C004.a, alongside trimethylindium, triethylindium and the associated production technology. Every shipment of indium phosphide out of China to a customer outside China now requires a permit. Permits run roughly sixty business days per order.

Watch what that does to the revenue line:

That is not a demand curve. Demand for AI optical interconnect did not fall 39 percent in the fourth quarter of 2025 and then rise 70 percent in the first quarter of 2026. Permits did.

The company said so directly. Fourth quarter revenue came in at $23.0 million against guidance of $27 to $30 million, and the chief executive explained the miss in these words: "we didn't receive as many export permits in Q4 as we had hoped based on the average processing time we had seen up to that point in October."

So when a $47.6 million quarter lands against $34.8 million of consensus, the honest reading is that it is at least as much a permit release event as a demand event. Permits granted inside a quarter convert to revenue inside that quarter. The backlog they unlock is lumpy by construction.

Two consequences follow and they point in opposite directions.

The bearish one. A single administrative decision in Beijing can produce another fourth quarter style air pocket with zero change in end demand. If you own this stock you are underwriting a permit office.

The bullish one, which is subtler and which I think is underappreciated. If revenue is permit-gated rather than demand-gated, then reported revenue is a floor on demand, not a measure of it. The orders exist whether or not the paperwork cleared. And the air pockets are mechanical, which means they are buyable.

One detail that softens the risk. In the first quarter of 2026, 61.5 percent of revenue went to China-destination customers, which requires no export permit at all. Only 38.5 percent was licence-exposed. Management guides a 40 to 60 percent China mix. The permit exposure is real and it is not the whole company, and the mix is a lever management can pull.

Three things about the trade regime that are not in the coverage

One. There is a 50 percent tariff on this product and it was filed under solar.

HTS 3818.00.00 is the classification for doped indium phosphide and doped gallium arsenide substrates. Effective January 1, 2025, the US Trade Representative moved that entire eight digit subheading from the 25 percent Section 301 list to the 50 percent list. The action was framed publicly as a solar measure covering polysilicon and wafers, but it was executed at the eight digit level rather than the statistical suffix level, so compound semiconductor substrates were swept in as collateral scope.

AXT's product is Chinese origin. Any Chinese origin doped indium phosphide or gallium arsenide substrate entering the United States has carried a 50 percent duty since the start of 2025.

That also explains a number which looks alarming without context. North America revenue was $1.708 million in fiscal 2025, down 77.5 percent. AXT states plainly that it cannot forecast when export permits will allow indium phosphide shipments to the United States, and that no gallium arsenide permit to the US has ever been approved. The American market is functionally closed to them right now.

Two. The November 2025 truce did not touch any of this. Announcement No.10 was deliberately left in force. Only Article 2 of the separate Announcement No.46 was suspended, and that suspension runs to November 27, 2026. The indium phosphide regime is legally distinct from the gallium and germanium regime and it has no expiry date.

Three. The 25 percent Section 232 semiconductor tariff does not apply. Effective January 15, 2026, its scope is logic integrated circuits in HTS 8471.50, 8471.80 and 8473.30. A search of the full Federal Register notice returns zero hits for 3818, wafer, substrate, indium or gallium arsenide. Substrates escape it entirely. That is a bear talking point that turns out to be wrong, and knowing which risks are real is most of the work.

The demand is corroborated, and this is the strongest part of the file

If revenue is permit-gated, you need independent evidence that the underlying orders are real. There is some, and it is unusually hard evidence. Customers are prepaying for capacity.

The Lumentum agreement runs to December 31, 2031 and carries a second $43.5 million tranche in 2028.

That is $91.2 million of customer cash committed in a year when the company will do somewhere near $190 million of revenue. Firms do not wire deposits of that size for capacity they are speculating about. Two of the three counterparties are among the largest optical component makers in the world and have every ability to source elsewhere if they thought they could.

This is the single most persuasive item in the bull case and it is more persuasive than the revenue print, because a prepayment is a decision made by a customer with its own money at risk.

Against it, management's stated 2027 exit target is roughly $130 million per quarter of indium phosphide capacity, about $520 million annualised. That number exists only on an earnings call and requires roughly $100 million of 2027 capital spending to actually happen. Treat it as an aspiration with a funding condition attached, not a forecast.

The price, which is the part the two day move hides

Here is where I have to correct something I wrote earlier this week.

The obvious way to describe this stock is that it rose 57 percent in two sessions and is therefore extended. That is what the tape looks like if you start on Wednesday. It is also wrong.

At $58 this stock is 59.4 percent below its May high and still 9.5 percent below the price at which the company itself placed equity in April. The 57 percent move is a bounce off a deep drawdown, not a breakout to new highs. Anyone who bought the April placement is underwater.

That is a materially different setup from the one the two day chart suggests, and it changes the answer to whether you are chasing. You are not buying a stock making new highs. You are buying one that halved and is trying to come back on a genuine change in the numbers.

What it does not change: insiders sold roughly 266,000 shares between $86.73 and $115.24 about six weeks before this print. They were not selling into the blowout. They were selling before it, at much higher prices. That is not a crime and it is not proof of anything. It is a fact worth having.

What $58 is asking for

Enterprise value at $58.16, adjusted properly for the cash and long-term investments that screeners miss, puts AXT around 25.5 times trailing sales. That is above Lumentum at 21.7 and far above Coherent at 7.5.

On the September quarter run rate of $66 million annualised, the same enterprise value is about 12.1 times sales. At the $46.94 close it is 9.3 times.

So the valuation question reduces to one thing. Does the guided quarter happen, and does it hold. At 25 times trailing you are paying for a business that no longer exists. At 12 times forward you are paying a reasonable multiple for something growing triple digits, if the growth is real and if it recurs.

Given that revenue is permit-gated, whether it recurs is not a question about demand. It is a question about paperwork.

What I am doing

I am starting a position of 15,000 shares. That is a starter, not a full weight.

The plan is to accumulate on weakness, and the permit mechanism is precisely why that plan fits this stock. If reported revenue is gated by an administrative process running on a sixty business day cycle, this company will produce quarters that miss badly for reasons that have nothing to do with its customers. Those are the entries. A business whose air pockets are mechanical rather than fundamental is a business you want to buy on the air pockets.

What I will not do is put full size on before the September quarter tells us whether $66 million was a permit release or a new baseline.

What Part 2 has to answer

Does the guided $66 million land, and what was the permit count behind it. Everything else is downstream of that.

The capacity build. The April raise was $632.5 million against a company doing $27 million a quarter at the time. I want the furnace count, the cost per unit of capacity, the qualification timeline, and when incremental capacity actually produces revenue.

Six inch indium phosphide, which is the cost curve. Where is yield running, what is the customer qualification timeline, and is AXT genuinely ahead of Sumitomo and JX or merely first to announce.

The Tongmei listing on the Shanghai STAR Market. A separate Chinese listing of the subsidiary that owns the crystal growing is the largest structural question a US shareholder faces here. How much would be sold, and what is left for us.

The 2028 supply wave. Sumitomo is targeting 3.1 times its 2024 indium phosphide capacity. San'an has a six and a half billion yuan six inch line. SMART Photonics is bringing an Eindhoven fab to full operation. If all of that lands, the pricing window that produced 45 percent gross margins closes. I want the supply and demand curves side by side with dates on them.

Customer concentration. Three counterparties just prepaid $91.2 million. That is excellent evidence of demand and poor evidence of diversification.

Part 2 carries the model. This one is the map.

New position disclosed above, opened around the time of publication. Not investment advice. Every figure is sourced below and dated.

References

• AXT, Inc. Announces Second Quarter 2026 Financial Results, 30 July 2026 (Business Wire; SEC 8-K EX-99.1, accession 0001437749-26-025061)

• AXT, Inc. Announces First Quarter 2026 Financial Results, 30 April 2026

• AXT, Inc. Form 10-K fiscal 2025, SEC CIK 0001051627, accession 0001437749-26-008612, Note 14 geographical information

• AXT, Inc. Form 10-Q first quarter 2026, accession 0001437749-26-017054

• US Trade Representative, Section 301 Federal Register notice modifying subheadings, effective 1 January 2025, Annex B, HTS 3818.00.00

• Federal Register, Section 232 semiconductor proclamation implementation, 20 January 2026, US note 39

• MOFCOM and General Administration of Customs Announcement No.10 of 2025, 4 February 2025, control codes 3C004.a through 3C004.c and 3E004

• Semiconductor Today, AXT's Q4/2025 revenue constrained by delay in China export permits, 9 March 2026

• Semiconductor Today, AXT's revenue grows 17% in Q1 after greater-than-expected export permits, 5 May 2026

• Capacity reservation agreements, AXT 8-K filings and press releases, Lumentum, Coherent and Nanjing Casela

• Form 4 filings, accessions 0001437749-26-019464 and 0001437749-26-019347

• Price and volume data, 29 and 30 July 2026, NASDAQ consolidated tape

The clean version goes like this. China restricted germanium exports. Germanium is the material every infrared lens is made from. LightPath Technologies makes BlackDiamond, a germanium-free glass, and is therefore the only way for Western defense programs to keep building thermal imagers. Buy the stock.

Two of those sentences are true. The one in the middle is not, and LightPath says so itself, in its own annual report, in writing.

What is left after you remove it is a more interesting company than the story implies, sitting on top of two disclosures made in the last nine days that nobody has written about.

First, the part of the story that is wrong

BlackDiamond BD6 is arsenic selenide, As40Se60. That is not a proprietary formula. It is a well known chalcogenide composition.

VITRON Spezialwerkstoffe in Jena, Germany publishes a glass called IG 6. Its composition is As40Se60. The same material. SCHOTT sells the same family as IRG 22 through IRG 27. Umicore sells GASIR.

You do not have to take my word for it, because LightPath's own fiscal 2025 annual report says it plainly:

"Some of these materials such as BD6 and BD2, are generic, and are produced by at least two other companies. The crown jewels of our materials, the BDNL type, are exclusive to us."

That is the company telling you the material at the center of the popular thesis is generic. Any bull case resting on BD6 being the only germanium substitute is going to meet that sentence eventually, and it will lose.

So what is actually defensible? Four things, and they are better than the chemistry claim because they are harder to copy.

Domestic melt capacity. Making the glass in the United States is not the same as knowing the formula. VITRON is German, SCHOTT is German, Umicore is Belgian. For a US defense program with domestic sourcing requirements, where the glass is melted is the qualifying fact, not what is in it.

A government-funded qualification path. This is the piece I find most persuasive and it is completely absent from coverage. The Defense Logistics Agency has paid LightPath twice, under two sequential contracts, specifically to qualify chalcogenide as a defense material:

Material qualification is the actual barrier to entry in defense optics. Not the recipe. The recipe is in the literature. What takes years is getting a material onto a qualified list and into a program of record, and the US government is underwriting LightPath's path through that process. That is public, dated, and durable in a way a glass composition is not.

Molding yield. Chalcogenide is precision moulded rather than diamond turned. That process knowledge is real and it does not transfer with a datasheet.

Vertical integration into cameras, via the G5 Infrared acquisition, which is where most of the value and most of the accounting complexity now sits.

And the honest technical caveat. Chalcogenide is not a straight drop-in for germanium. Transmission band and thermal drift behaviour differ. Some designs substitute cleanly and some require a redesign. Anyone telling you it is a simple swap has not read a datasheet.

The disclosure from nine days ago that nobody has read

On July 22, LightPath filed twenty two post-effective amendments and seven S-8 post-effective amendments. On July 23, twenty two effectiveness notices came back. Twenty two of anything in one day gets my attention.

Most of it is housekeeping. The cover pages say removal from registration of securities that were not sold.

But buried in it is something live. Registration number 333-262768, filed in February 2022, registered $75.8 million of shelf capacity plus $25.2 million under an at-the-market sales agreement with A.G.P. and Alliance Global Partners. Deregistering it formally terminates that at-the-market facility.

That matters more than it sounds. An at-the-market programme is a machine that sells stock into the market quietly, every day, at whatever the bid is. Turning it off means future capital raises have to be visible, discrete block deals rather than a continuous drip. For a company with LightPath's history of returning to the equity market, replacing silent supply with announced supply is a structural improvement in how shareholders get treated, and nobody has noticed it happened.

The remaining live vehicles are the December 2025 primary shelf and two resale shelves.

The disclosure from eight days ago that is genuinely uncomfortable

On July 23, LightPath agreed to sell one hundred percent of its Chinese subsidiary, LightPath Zhenjiang Optical Instrumentation, to a company called Hengtu Optical Technology in Nanjing.

Read the counterparty description from the filing carefully:

owned by the Purchaser Representative (Mr. Leo Zheng) and certain members of the Company's current management team.

LightPath's own management is on the buying side.

The price is $4.5 million paid over five years, minimum $500,000 per year, with 4 percent financing interest per installment and 7 percent on default, unsecured on the public record. LightPath retains a board observer seat and licenses over the trademarks for five years along with product drawings, tooling, moulds, process documentation, technology and know-how. There is a five year exclusive supply agreement, so LightPath keeps buying from the factory it sold.

Those China operations generated roughly $4.5 million per year of third party revenue. So LightPath is selling a business at approximately one times the revenue it gives up, collecting the money at $500,000 a year, from its own managers, with no disclosed fairness opinion.

Now the clause that made me stop:

"for every payment by the Purchaser of $1,000,000 of the principal... the cost plus markup percentage for products supplied to LightPath shall automatically increase by five percent (5%)... Upon full payment... fixed at cost plus thirty percent (30%)"

The supply agreement starts at cost plus ten percent. Every million dollars the buyer repays raises the price LightPath pays for its own products by five points. Stepped out, it looks like this:

The faster management pays LightPath for the factory, the worse LightPath's gross margin gets.

I want to be fair about this. There is a legitimate strategic reading. A company selling a China-substitution thesis while manufacturing in China has an obvious contradiction to resolve, and selling to the local management team who actually run the plant is a normal way to divest an operation that will not attract an arm's length buyer. Retaining supply, tooling and an observer seat is prudent.

But the ratchet is not standard, the counterparty includes insiders, no fairness opinion is disclosed, and the transaction has not closed yet. Those four facts belong in front of anyone underwriting this equity.

G5 is the company now, and the earnout tells you what it earned

LightPath bought G5 Infrared in February 2025. G5 makes cooled and uncooled thermal cameras. Camera level integration carries better margins than selling lens elements to somebody else who builds the camera.

G5's financial performance is never separately disclosed. But the earnout structure is disclosed, and the earnout can be solved backwards. This is the most useful piece of arithmetic in the file.

Start with the base year. The purchase agreement contained a clawback equal to 50 percent of any shortfall below $17.3 million of calendar 2024 revenue. The annual report shows a revenue clawback of $1,104,471. So $17,300,000 minus twice $1,104,471 gives $15,091,058. G5 did $15.09 million of revenue in calendar 2024. That number appears nowhere in any filing. It falls out of the clawback.

Then year one. The earnout ladder paid $3.5 million above $21 million of revenue, $7.0 million above $23 million, $10.5 million above $25 million and $14.0 million above $27 million. In the quarter ended March 31, 2026, LightPath paid $7.3 million in cash plus 297,445 shares worth $3.2 million. Exactly $10.5 million. That is the third tier.

Then year two, and this is the tell. The ladder pays $4.5 million above $30 million and $9.0 million above $33 million. On April 16, 2026, LightPath signed a side letter fixing year two at the maximum $9.0 million, ten and a half months before the measurement period even ends and four and a half months earlier than the contract required payment.

You do not agree to pay the top tier early unless you already know you are hitting it.

That is de facto forward guidance of G5 revenue above $33 million, disclosed in a contingent consideration footnote instead of a press release.

What it actually cost. All in, G5 is roughly $43.1 million rather than the $27.1 million headline, of which about $32.3 million is cash. The sellers collect $19.5 million of a possible $23.0 million of earnout, or 84.8 percent. That is a business bought well if the growth holds and expensively if it does not.

Why the reported loss is mostly not a loss

The nine month fiscal 2026 net loss is $16,404,698. That is the number a screener shows you and it looks awful for a company this size. Inside it is $12,234,529 of change in fair value of acquisition liabilities, which is the earnout being marked up because G5 keeps beating its tiers.

Strip it out and the underlying loss is about $4.17 million, with an underlying operating loss near $3.12 million. Quarter by quarter, excluding the mark, that loss narrows every single period:

That is a completely different picture from a $16.4 million hole.

Here is the elegant, perverse part. The better G5 performs, the larger the reported GAAP loss becomes, because the earnout liability marks up. Anyone screening this company on GAAP net income is being shown a number that gets worse as the business gets better.

The earnout liability trajectory is the cleanest public scorecard on a subsidiary that is otherwise invisible:

The jump in the December quarter is the moment G5 blew through its tiers.

The thing happening in about 48 hours

The June 3, 2026 offering carried twin sixty day restrictions that expire on or about August 2, 2026.

The first is a lock-up on all directors and executive officers and on North Run Strategic Opportunities Fund I, LP. North Run's residual 6,717,376 shares, 9.6 percent of the company, become saleable.

The second is LightPath's own restriction on issuing stock or convertibles, including the ban on variable rate transactions. The company becomes free to raise again, directly into the fiscal 2026 annual report window, which historically lands in late September.

I am not predicting either happens. I am telling you the doors open this week, and that a 9.6 percent holder becoming free to sell is the kind of thing that later gets attributed to news.

The order book, which is bigger than the press releases and is in the filings

Correction, added 31 July 2026. An earlier version of this piece said the backlog could not be verified from financial statements and that anyone quoting an order figure was quoting a press release. That was wrong, and a paid subscriber who has held the stock over a year and visited the Orlando facility corrected me within hours of publication. LightPath discloses backlog in MD&A as a named key indicator, with a five quarter table. I looked in the wrong part of the document. The corrected section follows, and the real number is much better than the one I used.

From the quarterly report for the period ended 31 March 2026, verbatim:

"Our total backlog at March 31, 2026 was approximately $110.6 million, an increase of 196%, as compared to $37.4 million as of June 30, 2025."

That is $110.6 million of firm orders against a business that did roughly $45 million of revenue in the trailing year. It is not a press release number, it is management discussion in a filed report, and the company defines it tightly: firm orders reasonably believed to remain in the backlog and convert to revenue, including customer purchase orders and qualifying supply contracts.

The composition matters more than the total.

Read the first line again. Fifty eight million dollars of it comes from "a leading global technology customer" for advanced infrared camera systems shipping in calendar 2026 and 2027. Not a defense prime. A technology company. I had assumed the large camera orders were a prime contractor buying for a military programme, and the filing says otherwise. Whoever that customer is, they are more than half the backlog, which is both the strongest fact in the bull case and the single largest concentration risk in the company.

There is also a third acquisition I had not found. LightPath closed on Amorphous Materials LLC on 21 January 2026. Amorphous Materials makes chalcogenide glass. So the company whose own annual report concedes that BD6 is generic went and bought one of the other producers of it. That is a direct answer to the objection I raised earlier in this piece, and it is a better answer than anything in the marketing.

And in the engineering services line, the filing names a Visimid development contract with Lockheed Martin. That is a real prime relationship, disclosed, currently small enough that its timing moves the segment around.

On the federal contract data, the original point still stands and is worth keeping. LightPath has approximately $1.0 million of lifetime federal prime contracts across six awards. It is a subcontractor. Its product goes into somebody else's camera. Prime to subcontractor purchase orders are not public, so USASpending will never show you this book. The mistake I made was concluding that therefore nothing was verifiable. The company reports it directly; you just have to read the management discussion rather than the contract database.

One research note for anyone following this trail. A plain search for LIGHTPATH in federal contract data returns about thirty awards, but most of them belong to a completely different company, the Lightpath fibre business selling internet circuits to the DEA, the FBI and the Maritime Administration. Filter to LIGHTPATH TECHNOLOGIES or you will overstate the defense footprint by roughly sixty percent.

Where I come out

I am not taking a position today. Here is the honest reason.

The company is more interesting than its story. The real moat, domestic melt plus a government-funded qualification path plus moulding yield plus camera integration, is more durable than the germanium narrative that got people here, and buying Amorphous Materials in January strengthens it further. G5 is compounding fast enough that management pre-committed to its maximum earnout. Backlog is $110.6 million against roughly $45 million of trailing revenue and it is disclosed, not asserted. The underlying loss is a quarter of the reported one and shrinking. The at-the-market facility is off.

Against that, a related party sale of the China factory to management with a margin ratchet and no fairness opinion, a 9.6 percent holder coming unlocked this week, a fresh ability to issue equity in the same window, more than half the backlog resting on one unnamed customer, and a central marketing claim the company's own annual report contradicts.

Those are not disqualifying. They are the reason to do the second half of the work before putting money down rather than after.

What Part 2 has to answer

Does the China sale close on these terms. No completion has been filed. Terms can change and the market has priced none of this.

Fiscal 2026 full year numbers. The year ended June 30. The annual report lands around late September and will carry the first clean full year view with G5 and AML in it, plus the June backlog figure.

Whether North Run sells. Ownership filings after August 2 answer that quickly.

Segment margins, camera versus component, separately. The entire vertical integration thesis rests on cameras being structurally better and nobody has shown the split.

Who the customer is. One unnamed leading global technology customer accounts for roughly $58 million of a $110.6 million backlog. Identifying it is the single highest value open question in this name, and the fact that it is a technology company rather than a defense prime narrows the field considerably.

Where the glass gets melted after the divestiture, and what the cost plus ratchet does to gross margin at each repayment step, modelled out.

What Amorphous Materials actually brings. LightPath bought a chalcogenide glass producer in January. Price, capacity, customer list, and whether it closes the generic-material gap or simply adds volume.

And the competitive question that decides the thesis. If VITRON and SCHOTT already make the same composition, what stops a US licensee of either from qualifying domestically. The answer is probably the qualification timeline. I want that timeline quantified in years.

Part 2 carries the model. This one is the map, and the map has a few roads on it that are not on anyone else's.

No position at time of writing. Not investment advice. Every figure sourced below and dated.

References

• LightPath Technologies Form 10-K fiscal 2025, SEC CIK 0000889971, accession 0001654954-25-011130, Item 1 and Note 3

• Form 10-Q, quarter ended 31 March 2026, accession 0001437749-26-015594, Note 4 and segment reconciliation

• Form 8-K, 23 July 2026, accession 0001437749-26-024234, Item 1.01 and Exhibit 99.1, China divestiture

• Form 8-K, 3 June 2026, accession 0001437749-26-019435, Item 1.01, offering and lock-up

• Membership Interest Purchase Agreement, G5 Infrared, Exhibit 10.1 to 8-K accession 0001654954-25-001429, sections 2.8(b) and 2.8(c)

• Post-effective amendments, accessions 0001437749-26-024147 through 0001437749-26-024176, and EFFECT 9999999995-26-002408

• Registration statement 333-262768, A.G.P. and Alliance Global Partners at-the-market facility

• USASpending prime award records, recipient LIGHTPATH TECHNOLOGIES, contracts SP800023P0004 and SP800024C0009

• LightPath BD6 datasheet, Infrared Chalcogenide Glass As40Se60

• VITRON Spezialwerkstoffe GmbH IG 6 specification and SCHOTT IRG series datasheets

• Price and volume data, 30 July 2026, NASDAQ consolidated tape

The answer is on the record. It came on July 23 with the half-year report.

The order did not come.

What came instead was a set of numbers almost nobody has read properly, including the financial press that covered them. LPKF has four million euros of cash. It also claims eighty percent of the market it is trying to win. Both of those sentences come from documents published on the same day.

This post is the longest in the series because the research is. I read the half-year report, the earnings call deck, the annual general meeting voting record, the shareholder counter-motions, the chief financial officer's speech to that meeting, the articles of incorporation, and the full German short-selling register going back to 2013. Several of the things I found have not been reported anywhere. Two of the things that were reported are wrong, and I will show you why.

Let's go.

Read more

This report is free, ungated, and built to be forwarded. Quote anything with attribution to Shawarma Capital, research.shawarmacapital.net. All figures marked to the July 15, 2026 close of $66.31 unless dated otherwise.

Key findings, each one verifiable in a filing:

• Satellite service revenue in Q1 2026 was $1.3 million. The market cap is $20 billion. A pre-revenue multiple is not the indictment, every infrastructure story prices the future. The finding is the race underneath it: the dilution is arriving on schedule and the revenue is not.

• 91 percent of the quarter's revenue was hardware, and more than half of the total was gateway equipment billed to a related party.

• In August 2025 the CEO confirmed a fully-funded plan. Within a year the company priced $2.1 billion in convertible notes across two offerings.

• The company targets approximately 45 satellites in orbit during 2026. It has nine. Of its three named launch providers, Blue Origin blew up its own pad in May and is grounded until December, ULA has yet to fly an AST mission, and the only one actually flying the manifest, SpaceX, competes with AST for the same customers.

• The July convert converts at $79.57. Convert arbitrage hedging is a standing short position above the tape. The stock does not escape the strike zone until the income statement does.

Executive summary, one page

AST SpaceMobile is building a constellation purpose-built to deliver true broadband, not just texting, to unmodified smartphones. Starlink's direct-to-cell reaches the same phones, but at messaging-grade speeds today, while AST's far larger arrays target actual broadband and it holds the first US commercial authorization for it. Over the last ninety days it launched three satellites successfully, won that FCC authorization for commercial US service, signed a $926 million sovereign-network deal in Japan, and grew its partner roster to roughly 60 mobile operators covering more than 3 billion subscribers. The technology case got stronger. The stock still round-tripped from $85 through $133 back to $66, and the reason is arithmetic, not sentiment.

Three exhibits carry the report.

Exhibit A. The $14.7 million question. First-quarter revenue printed $14.7 million against the $24 to 26 million the ramp implied. Inside that number: $13.4 million was hardware, gateway equipment sold to partners, and $7.9 million of that was billed to a related party. Satellite connectivity service, the thing the $20 billion market cap is priced on, produced $1.3 million. Nine percent of a $14.7 million quarter. The 2026 guide of $150 to 200 million is only about half covered by contracted backlog, and the company now points to 2027 for commercial service at scale.

Exhibit B. The manifest math. Management targets approximately 45 satellites in orbit during 2026. Today there are nine functioning spacecraft, five Block 1 and four Block 2, after BlueBird 7 was lost in April. Three more are slated for early August. That leaves roughly 33 satellites to fly in under five months, at a cadence the company has never demonstrated, while its launch options narrow. AST names three providers, SpaceX, Blue Origin, and ULA, but Blue Origin's New Glenn destroyed its own launch complex in a May anomaly and does not return to flight before December, and ULA has yet to fly an AST mission. Every near-term launch that actually flies now routes through SpaceX, a company whose own Starlink direct-to-cell product competes with AST for the same subscribers.

Exhibit C. The dilution ratchet. In August 2025 the CEO told investors, quote, we are confirming our fully-funded plan to deploy 45 to 60 satellites into orbit by 2026. Since that sentence: a $1.075 billion convertible in February 2026 and a $1.0 billion convertible priced July 15, 2026, at 1.625 percent due 2034, converting at $79.57, with $96.9 million of the new money spent on a capped call to push effective dilution to $149.20. That is $2.1 billion of fresh converts inside a year of the words fully funded. Pro forma debt now stands near $4 billion against roughly $4.4 billion of cash, the share count is 299 million and rising, short interest is 21.7 percent of float and at a nine-month high, the stock just left the Russell 2500, and the company president filed sales at $126.64 within a day of the all-time high. The quarter's net loss was $191 million. The capital machine is working exactly as designed. That is the problem: it is designed to run on your shares.

Verdict. The thesis is alive and the price of admission finally reflects the risk. We grade our own April call at the bottom, publish the updated scenario table, and list the three tripwires that would make us buyers. Until service revenue exceeds hardware revenue in a printed quarter, this is a launch-cadence option, not an infrastructure stock.

I. Grading our own April work first

The April 20 piece made five calls. The record, honestly scored.

Right: the satellite loss was noise. We wrote that BlueBird 7 was a launch vehicle failure, not a technology failure, that insurance would cover it, and that the next launch was with SpaceX. All confirmed. BlueBirds 8, 9, and 10 went up clean on a Falcon 9 on June 17, the Block 2 design with 2,400 square foot arrays flying as specified.

Right: the Rakuten overhang resolved bullishly. April's report treated Rakuten's $270 million share sale as portfolio mechanics, not thesis abandonment, and said watch the commercial agreement instead. On July 2 Rakuten committed $926 million alongside AST for a sovereign direct-to-device network in Japan, with government approval following a day later. The seller of April is the anchor tenant of July.

Right, with an asterisk: the technology. The company demonstrated 98.9 megabits per second to an unmodified smartphone, the best direct-to-device figure ever published, and won the FCC's first Supplemental Coverage from Space authorization for commercial service, covering up to 248 satellites. The moat claims are converting into regulatory fact.

Wrong: the revenue trajectory. We relayed an implied Q1 of $24 to 26 million from the Q4 exit rate. The actual print was $14.7 million, and the composition was worse than the size, as Exhibit A details. The Q4 2025 number that anchored the ramp story was itself heavy with gateway hardware and milestone recognition, not recurring service. We passed along the annualized-run-rate framing without decomposing it. That was the error: right numbers, wrong reading.

Wrong: the base-case timing. April's scenario table put base-case 2028 revenue at $800 million on constellation scale arriving through 2026 and 2027. The company has since told investors commercial service at scale is a 2027 event, and the street's Q2 estimate sits at $25 to 40 million. The revenue curve shifted right roughly a year. The scenario table at the bottom reprices accordingly.

Ninety days is a short grading window. It was long enough for the market to reprice the same facts from $133.09, the May 28 all-time high, to $65.62 at the low. What follows is the audit of what actually changed.

II. Exhibit A in full: what the revenue actually is

The Q1 2026 filing, reported May 11, decomposes as follows.

Read that middle line again. More than half the quarter's revenue was gateway hardware billed to a related party. This is not an accusation of anything improper, related-party gateway sales to MNO partners who are also shareholders are a documented part of the model, and gateway deployments are a necessary precondition for service revenue in each market. But an investor pricing a $20 billion company on a revenue line should know that the line is currently dominated by one-time equipment deliveries to affiliated buyers, not by satellites billing subscribers.

The service line, $1.3 million, is the entire current commercial reality of space-based cellular broadband at AST. It will grow, and to be clear about the frame, a $20 billion valuation on nascent revenue is not itself the finding. Amazon traded at absurd trailing multiples for a decade and the future showed up. The finding is the race: the future revenue is now dated 2027 by the company's own offering documents, while the dilution funding it is dated February 2026, July 2026, and, if the pattern holds, whenever the next launch contract needs signing. When a stock prices the future, the only question that matters is which future arrives first. Right now the converts are winning.

The full-year guide of $150 to 200 million was reaffirmed on May 11, with management stating approximately 50 percent is covered by existing contracted backlog. Half of a guide is a plan; the other half is a hope, and the mix inside the contracted half skews toward more gateway equipment and government milestones, not consumer service. Meanwhile the July offering documents moved commercial service at scale into 2027, a slip the market punished within hours, and the street's Q2 consensus of roughly $35 million at the midpoint implies the back half must produce over $100 million to reach even the guide's floor.

Operating costs make the clock audible. Total Q1 operating expenses were $164.1 million, $91.2 million after stripping stock compensation and depreciation. Net loss attributable to common was $191.0 million, one quarter. Gross property and equipment stands at $1.8 billion. Every quarter of 2027-timeline slippage is roughly $200 million of GAAP loss and a proportionate draw against the cash pile, and the cash pile is borrowed.

III. Exhibit B in full: the manifest against the calendar

The deployment record, from the company's own announcements.

From 12 planned in August to the stated 45 requires roughly 33 additional satellites in under five months. At three per Falcon 9, that is eleven dedicated launches, better than one every two weeks, from a standing start of one launch per month at best. The company's Midland and Texas facilities are producing at a claimed rate of phased arrays for more than ten satellites per month, and BlueBirds 11 through 33 are stated to be in various stages of assembly. Production may genuinely be there. Orbit is not production. Orbit is launch slots.

And the launch market just got smaller. Blue Origin's New Glenn, the vehicle that lost BlueBird 7 in April, suffered what its own filings call a devastating anomaly in May that destroyed the launch pad. NASA is assisting the investigation, an engine issue is suspected, and the company has guided return to flight to December 2026 while it scraps and rebuilds the complex. Every near-term AST launch therefore routes through SpaceX.

Sit with that dependency. SpaceX operates Starlink direct-to-cell with T-Mobile, the direct competitor for AST's core product. AST's manifest, its cadence, and therefore its 2026 target and its 2027 service date now run through the pricing desk of its principal rival, with Blue Origin grounded until year-end and ULA yet to fly a mission for it. The July offering language about using proceeds to secure additional access to orbit, and the analyst chatter about buying a rocket company outright, are the company saying the same thing with a billion dollars: launch is the binding constraint, and it is currently rented from the competition.

For completeness, the FCC authorization covers up to 248 satellites, and management holds that roughly 45 to 60 are needed for continuous US coverage. At the demonstrated 2026 cadence, nine satellites in six months including a loss, continuous coverage is a 2028 story without a step-change in launch procurement. That is what the $1 billion is for. Which brings us to what it costs you.

IV. Exhibit C in full: the ratchet

The capital stack, before and after the last ninety days.

Start with the sentence that anchors the exhibit. August 2025, the CEO, on the record: we are confirming our fully-funded plan to deploy 45 to 60 satellites into orbit by 2026 to support continuous service in the US, Europe, Japan, and other strategic markets. Fully funded. Since then the company has priced two converts, $1.075 billion in February 2026 and $1.0 billion on July 15, 2026, roughly $2.1 billion of new paper inside a year of that quote, while the 45-satellite half of the same sentence slipped as Exhibit B shows. Neither raise is scandalous on its own. Together with the quote they are a pattern, and the pattern is that funding statements from this company have a shelf life measured in quarters. Price the next one in.

The July 15 convert: $1.0 billion principal, 1.625 percent coupon, due February 2034, conversion at $79.57, a 20 percent premium to the $66.31 reference. The company spent $96.9 million of the proceeds on a capped call raising the effective dilution threshold to $149.20. An additional $150 million greenshoe is available. Stated use of proceeds: growth initiatives and securing additional access to orbit, including potential partnerships and acquisitions. The stock fell roughly 12 percent around the announcement and pricing.

The mechanics matter as much as the amount. Convertible buyers hedge by shorting the common, which is part of why short interest sits at a nine-month high, and that hedging flow acts as a standing offer above the tape. Two billion dollars of converts struck at $79.57 and, from the February paper, higher, means systematic selling pressure into every rally between here and the strikes until the business generates cash. The retail complaint that the offering caps the stock is not paranoia, it is the arithmetic of how convert arbitrage trades. The cap lifts when the income statement lifts it, not before.

This is the second billion-dollar-class convert in a year, and it will not be the last, because the model consumes capital by design: satellites are capitalized, launches are purchased, gateways are deployed ahead of revenue, and the income statement loses $191 million a quarter while service revenue is $1.3 million. Management calls the balance sheet a fortress. It is, and the fortress is mortgaged: roughly $4 billion of debt against a business whose recurring revenue would not cover one quarter's interest, secured by the promise of 2027.

The supporting cast of signals, each small, all pointing the same way. President Scott Wisniewski filed sales of 25,904 shares at $126.64 on May 27, within a day of the $133.09 top, roughly $3.3 million. CEO Abel Avellan's May filing was a tax-withholding disposition at $113.41, not an open-market sale, and we flag the difference deliberately. The stock exited the Russell 2500 in the July reconstitution, removing passive bid. Short interest sits at a nine-month high with 2.6 days to cover, which cuts both ways: it is fuel for violent rallies on good launches and a standing vote that the arithmetic above is widely understood.

One more structural note. Amazon closed its $11.6 billion acquisition of Globalstar in mid-July, buying Apple's satellite partner outright, days after the FCC signaled an August vote on opening more direct-to-device spectrum. SpaceX flies its own constellation, Amazon now owns one, and both have consumer distribution AST will never match. AST's counter is physics, bigger apertures delivering broadband speeds the others cannot, and carrier alignment, roughly 60 MNOs plus the proposed AT&T, T-Mobile, Verizon joint venture that pools spectrum for exactly this service, with AST positioned as a preferred capacity provider. The competitive story is genuinely two-sided now. It was one-sided in April.

V. What got stronger, stated as plainly as the holes

An audit that only lists holes is a short report, and this is not one. Since April 20:

• Three satellites launched flawlessly on June 17, the largest commercial phased arrays ever flown, designed for roughly 200 megabit peak speeds, nearly double Block 1.

• The FCC granted the first commercial SCS authorization in US history, up to 248 satellites. The regulatory moat is real and it is AST's, not Starlink's.

• Rakuten converted from overhang to anchor: $926 million committed for a sovereign Japanese network, government approval in hand, on top of Japan's earlier equity history.

• The MNO roster reached roughly 60 operators spanning 3 billion subscribers, adding Telus and Axian, with ground integrations under way in 17 countries covering 2.9 billion people.

• Three new US government awards since March, on top of the existing defense relationships, and a $100 million capital advance securing long-term L-band spectrum access through Ligado.

• The 98.9 megabit demonstration to an ordinary smartphone stands as the industry record.

• Vodafone's European joint venture named Spain for its rollout with 2027 service, New Zealand approved a gateway, and European emergency-service demonstrations completed ahead of the August launches.

The asset being built is unique, the demand side keeps signing, and the regulator keeps saying yes. Nothing in this report disputes the destination. Everything in this report is about the fare.

VI. The repriced scenarios

April's table assumed 2026 constellation scale and $800 million of 2028 base-case revenue. Both slipped roughly a year. The update, our model, our probabilities.

Probability-weighted value: roughly $89. Against a $66 tape that is 35 percent of expected upside, materially less asymmetric than the same math looked in April at $85 against a $127 weighted value, because the bear case is fatter and nearer. The option is cheaper and the odds are worse. Both things happened.

What would make us buyers, three tripwires, any one of which converts this from option to position:

1. Service revenue exceeds product revenue in any printed quarter. The business inverts from hardware-and-hope to infrastructure.

2. A sustained launch cadence of six or more satellites per quarter for two consecutive quarters, from any provider mix, proving orbit access is procured, not promised.

3. A definitive capacity contract with the carrier joint venture or a named MNO with committed dollar minimums disclosed, converting the 60-partner roster from memoranda into backlog.

And the standing kill signal, published in advance: a third convertible or equity raise before commercial service revenue passes $50 million in a quarter would confirm the ratchet is the business model, and we would say so in print and stop covering the name as an investment candidate. Management has already spent the benefit of the doubt once, at fully funded. There is no second one.

VII. Verdict

In April we wrote that the market was pricing a thesis break and the math said otherwise. The market then agreed with us for five violent weeks, doubled the stock, and spent the next seven taking it all back as the filings caught up. Both moves were the same mistake in opposite directions: pricing the story instead of the composition.

The composition today: $1.3 million of quarterly service revenue, nine satellites, SpaceX as the only reliable launch path and also the chief competitor, $4 billion of debt, a 2027 service date, and the best direct-to-device broadband technology and regulatory position in the sector. That is a real company building a real thing on rented rockets and shareholder patience.

At $66 you are no longer paying the euphoria tax. You are paying the dilution tax, on a schedule management just published at $79.57 a share. We will pay neither until one of the three tripwires fires. The constellation is fine. The income statement isn't. Watching, no position, and the next hard look is the August 10 print and the BlueBird 11-13 launch in the same week.

Not investment advice. Do your own work.

Methodology note

Every figure in this report traces to a primary source: SEC filings, company press releases, FCC records, and Form 4 filings, with market data from the July 15, 2026 close. Scenario outputs are Shawarma Capital model estimates, not company guidance. We held no position, long or short, at any time during the preparation of this report.

References

1. AST SpaceMobile Q1 2026 business update and results, May 11, 2026. Business Wire and SEC filings. Revenue composition, opex, cash, guide, FCC SCS authorization, MNO roster, 45-satellite target.
   1a. AST SpaceMobile Q2 2025 business update, August 2025, SEC Form 8-K exhibit. CEO quote, confirming our fully-funded plan to deploy 45 to 60 satellites into orbit by 2026. February 2026 convertible notes offering, $1.075 billion, SEC filings and Investing.com coverage.

2. AST SpaceMobile 10-Q, quarter ended March 31, 2026. Products and services revenue split, related-party disclosure.

3. Convertible notes pricing release, July 15, 2026. Business Wire. Terms, capped call, use of proceeds.

4. BlueBird 8-10 launch release, June 17, 2026, and BlueBird 11-13 scheduling release, June 23, 2026. Business Wire.

5. Rakuten-AST Japan sovereign network coverage, July 2-3, 2026. Yahoo Finance and company statements.

6. Blue Origin New Glenn May anomaly and pad rebuild coverage, July 7-14, 2026. MSN, autoevolution, Yahoo Finance. Return to flight December 2026.

7. Form 4 filings, AST SpaceMobile, May-July 2026. SEC EDGAR. Wisniewski sale May 27; Avellan withholding disposition May 30.

8. Amazon-Globalstar acquisition coverage, July 10-16, 2026. Stocktwits, Il Sole 24 Ore syndication. $11.6 billion, Apple stake included.

9. AT&T, T-Mobile, Verizon direct-to-device joint venture coverage, May 14-20, 2026. Reuters, Light Reading, Via Satellite.

10. Service timeline to 2027 and acquisition language, July 15-16, 2026. PCMag, Aviation Week, Yahoo Finance.

11. Russell 2500 deletion, July 10-13, 2026. simplywall.st, Yahoo Finance.

12. Short interest and options data, July 14, 2026. Stocktwits, MarketBeat. Market data from Yahoo Finance, July 15, 2026 close.

13. Prior coverage: One Bird Down, Constellation Intact, Shawarma Capital, April 20, 2026, at research.shawarmacapital.net.

Prices marked as of the July 15, 2026 close, $68.88, with the stock indicated lower premarket July 16. Market cap roughly $3.0B. Fifty-two week range $48.30 to $96.69.

Let's go.

The whiplash, and why it is the setup

Three tapes in three weeks.

On June 30, Rosenblatt named Ambarella one of its eight favorite technology stocks for the second half of 2026 and called it, in analyst Kevin Cassidy's words, a Physical AI pure play. The stock rocketed 28 percent in a session and printed above $90 the next day.

On July 1, the day after that pop, CEO Feng-Ming Wang sold 16,250 shares at $90.08. His Senior VP of Systems sold 10,000 at $88.84 the same day. Hold that thought, we will come back to it with the full table, because I promised you the insider file and it is worse and also better than it looks.

Then the giveback. No earnings, no downgrade, no ruling. The stock bled from above $90 to $68.88 by the July 15 close, down 8.9 percent that day alone on volume 66 percent below average. Fourteen percent came off in a week. Thin-volume drift, not institutional exit. The market handed back almost the entire Rosenblatt move while the average analyst target sits between $95 and $101.

That gap between a $120 top-pick call and a $69 low-volume tape is why this writeup exists. Worth naming the froth too: after the Rosenblatt call, AMBA started showing up in every anonymous physical-AI basket thread on X, the same promo accounts pushing nine tickers at a time. That crowd bought the pop. The drift back to $69 is them leaving. Their exit is the entry.

One correction to the record before we build, because getting the tape right matters more than a clean narrative. When this name first hit my desk I attributed the July drawdown to GoPro's patent win over Insta360, Ambarella's largest end customer. The full docket says otherwise. The ITC sided with GoPro in its February commission opinion, and the stock did take a hit on that news at the time. But the final ITC determination in March went the other way in practice, and the March 11 headline reads that Ambarella's largest customer avoided US import restrictions. There was no new ruling in July. GoPro itself is fighting for survival, taking a $20 million loan from its own founder on July 9. The July drawdown is a sentiment unwind, not a legal event. The distinction matters because sentiment unwinds mean-revert and import bans do not.

What Ambarella actually is

Ambarella is a fabless semiconductor company out of Santa Clara, founded in 2004, that spent its first decade as the video-compression chip inside GoPros and DJI drones and its second decade converting that image-processing franchise into low-power AI inference silicon. The company has shipped more than 39 million SoCs cumulatively.

The product is a system-on-chip that sits directly behind a camera sensor and runs neural networks on the device, at the lens, without sending anything to a data center. Security cameras that recognize instead of record. Car cameras that detect and decide inside the cabin power budget. Drones, delivery robots, industrial inspection systems, port automation, surgical visualization. Any machine where the answer has to arrive in milliseconds on watts of power, not in seconds on a grid connection.

Three product families carry the thesis.

CVflow vision SoCs. The core franchise. The CV5 and CV7 generations run computer vision at single-digit watts. The CV7, launched January 2026 on a 4 nanometer process, cut power 20 percent while multiplying AI throughput 2.5 times against the prior generation.

CV3-AD for automotive. The autonomous-driving line, sold to Tier-1 suppliers, targeting ADAS and higher levels of autonomy. Management called out record automotive revenue in the April quarter, driven by commercial vehicles adopting AI camera systems at a pace, in the company's words, above seasonal.

N1 for edge generative AI. The newest line, built to run multi-modal models on-device. This is the optionality tail, not the core earnings driver yet.

The strategic point is simple. Nvidia owns the data center. Qualcomm owns the phone. The market for inference at the sensor, where power is measured in single watts and latency in milliseconds, is still being divided. Ambarella is one of the only pure plays on that layer at public-market scale, and roughly 80 percent of its revenue is already edge AI product.

The Hanwha agreement, with the ratio and the caveats

On May 28, the same day as earnings, Ambarella announced a long-term edge AI agreement with Hanwha, the Korean industrial and defense conglomerate. Term exceeding ten years. Value in excess of $800 million in potential revenue. Scope covering edge AI SoCs and software across video security, robotics, industrial automation, and life sciences, co-developed on Ambarella's Cooper platform, with the deal sponsored from the Hanwha Group level, the quote coming from Kim Dong-Seon, not just the Hanwha Vision subsidiary.

Sit with the ratio. Ambarella's entire fiscal 2026 revenue was $390.7 million. One customer signed a decade-long agreement worth more than two full years of the company's current revenue.

Now the caveats, because the ratio is doing a lot of work in the bull case and you should know exactly what it is made of.

First, the $800 million is characterized as potential revenue, not committed volume. There are no disclosed take-or-pay commitments and no disclosed minimums.

Second, Hanwha Vision retains independent control of its in-house Wisenet SoC line. This is a transition, not a displacement on day one.

Third, the timing. The street read, per Rosenblatt, is that Hanwha begins contributing meaningful revenue upside next year, as Hanwha moves from a mix of internal and external chip sourcing toward Ambarella supplying its security camera silicon, worth roughly $50 million a year in that segment alone once fully transitioned. Fifty million against a $400 million base is a 12 percent revenue layer from one logo, before the robotics and industrial scope contributes anything.

A signed decade with a chaebol that builds cameras, robots, and defense systems is exactly what an edge AI silicon company's order book is supposed to look like at this stage. It is also, today, a promise with an asterisk. Both things are true.

The print

The April quarter, fiscal Q1 2027, reported May 28.

• Revenue $100.4 million, up 16.9 percent year over year, down half a percent sequentially, slightly above the midpoint of guidance.

• Non-GAAP gross margin 59.9 percent, against 62.0 percent a year ago.

• Non-GAAP net income $5.0 million, 11 cents, versus $3.0 million a year ago.

• GAAP net loss $18.1 million, narrowed from $24.3 million.

• Cash and marketable securities $277.8 million. Debt near zero at $13.3 million.

• Q2 guide $105 to $111 million, which at the midpoint is 8 percent sequential growth and roughly 15 percent year over year.

• New $50 million buyback authorized through June 2027.

The mix is the story inside the story. IoT applications, meaning security cameras plus consumer devices, are still about three quarters of revenue, and consumer IoT declined double digits sequentially. Automotive printed an all-time record on commercial fleet adoption. The growth engine is rotating from the legacy camera base toward auto and enterprise while the consumer tail shrinks. That rotation is the whole investment question, and this quarter it worked, barely, to a flat sequential top line.

The deceleration is real and I am not going to hide it. Fiscal 2026 grew 37.2 percent. The April quarter grew 16.9. The Q2 guide implies mid-teens. The re-acceleration case rests on three named drivers, the Hanwha transition beginning next fiscal year, the commercial-vehicle automotive ramp, and edge inference volume in new categories, robots and industrial systems, that barely existed in the run-rate two years ago.

Stress test one, the GAAP losses and the stock comp

The founder note flagged it and here is the honest math. Ambarella loses money on a GAAP basis, $18.1 million last quarter, roughly $70 million trailing twelve months. The gap between GAAP loss and non-GAAP profit is almost entirely stock-based compensation, running in the low twenty-millions per quarter against $100 million of revenue. That is a real cost. It dilutes you at roughly 20 percent of revenue per year, and the new $50 million buyback, of which just $2.4 million was executed in Q1, offsets only a fraction of it at current issuance rates.

What keeps this from being disqualifying is where the money goes. R&D was $58.1 million in the quarter, 58 percent of revenue. This is a company spending like a large-cap on silicon roadmaps, 4 nanometer vision SoCs, automotive-grade autonomy chips, edge LLM silicon, while running a microcap income statement. If revenue re-accelerates, that R&D line is operating leverage. If it does not, the SBC keeps taxing you while you wait. That is the trade, stated plainly.

Stress test two, the concentration file

This is the section the 10-Q writes for you, and it is the most important table in this piece.

• WT Microelectronics, the Taiwanese distributor that fulfills Ambarella's sales across Asia outside Japan, was 61 percent of total revenue in the April quarter. Fiscal 2026, 70 percent. The trend line runs 53, 63, 70 percent across fiscal 2024, 2025, 2026.

• The top ten end customers were approximately 67 percent of revenue in fiscal 2026 and the April quarter.

• The largest end customer so far in fiscal 2027 is Arashi Vision, Insta360, supplied indirectly through WT to the ODMs that build Insta360's cameras.

• Revenue by bill-to location in the April quarter: Taiwan $61.0 million, Asia Pacific ex-Taiwan $23.2 million, Europe $6.3 million, and the United States $1.9 million. The Santa Clara chip company bills less than 2 percent of its revenue inside the United States, because its customers are the Asian ODMs that build the world's cameras.

• Buried in the subsequent events note: on May 12 the company terminated a development project with a customer and refunded $4.5 million of a $13.5 million deposit. Custom silicon programs die quietly like this. Worth watching whether it was a one-off.

None of this is hidden and none of it is fraud. It is the standard anatomy of a fabless vision-chip company whose demand lives in Asian supply chains. But be precise about what you own: a US-listed IP company whose revenue collects through one Taiwanese distributor and whose single largest end customer is a Shenzhen action-camera maker. Which brings us to the third test.

Stress test three, the Insta360 file, corrected

The record, in order. GoPro filed ITC and district court actions against Insta360 in 2025. In February 2026 the commission opinion reaffirmed GoPro's patent claim, and Ambarella sold off on the headline the same week it printed record Q4 results. In March 2026 the final ITC determination resolved in a way that let Insta360 keep importing, and the coverage read that Ambarella's largest customer avoided US import restrictions. There has been no adverse ruling since. GoPro, meanwhile, is being outsold by DJI and Insta360 globally, and on July 9 disclosed a $20 million loan from its own founder to stay liquid.

So the patent overhang, as of today, is resolved in Insta360's favor at the ITC, with the residual risk being appeals and the district court track rather than an import ban in force.

The real Insta360 risk is simpler and bigger than the lawsuit. It is the concentration itself. Insta360 is a private Shenzhen company, exposed to US-China trade policy, in a consumer category that just printed a double-digit sequential decline for Ambarella, and it is the largest single end customer of the company you would be buying. If Insta360 stumbles, gets tariffed, or dual-sources its next generation, the IoT base takes the hit long before Hanwha and automotive fill it. That risk does not require a courtroom.

Stress test four, the insider file

Every disclosed sale from the last sixty days, from the Form 4 flow.

Zero open-market purchases against them, insider ownership at 4.6 percent, and the CEO's sale landed the day after the Rosenblatt pop, within 20 cents of the six-month high.

Now the other side of the ledger. The amounts are small. The CEO sold a million and a half dollars of a three-billion-dollar company, almost certainly through a pre-set 10b5-1 plan that happened to trigger into strength, and executives at chip companies with heavy SBC sell mechanically because equity is most of their pay. This is not a management team dumping a secret. It is a management team that does not buy its own dip, which tells you something softer: nobody inside believes the stock is stupidly cheap here. They believe it is fairly paid compensation. Read it as an absence of conviction, not the presence of a warning.

What the street sees

Fourteen analysts, consensus buy, average target $95 to $101 depending on the aggregator, against a $69 tape. The recent tape of calls: Rosenblatt reiterating buy at $120 on the physical AI thesis, Susquehanna to $110, Stifel to $106, BofA at $96. Next print August 27, Q2 fiscal 2027.

The consensus case is the simple one. Edge AI inference is where AI deployment physically lands, in cameras, cars, robots, and Ambarella is the only clean public-market vehicle for that layer. The high-conviction versions of the thesis, Rosenblatt's version, add the Hanwha decade and the automotive record as evidence the design-win pipeline is converting.

The market's counter, expressed in the tape rather than in any note, is that 16.9 percent growth at 60 times forward earnings requires the re-acceleration to actually show up, and July's drift is the market refusing to prepay for it twice in three weeks.

The scenarios

My model, my numbers, not guidance. Fiscal 2028 basis, roughly 18 months out.

Probability-weighted value roughly $95, which is 38 percent above the July 15 close and, not coincidentally, right where the analyst consensus sits. The market is pricing the bear-to-base corridor. The Rosenblatt case requires believing the bull column, and the asymmetry at $69 is that you are paid to wait for the evidence, because the bear case at 5 times sales for a net-cash company with 60 percent gross margins and a decade-long Hanwha agreement is a shallow bear, not a zero.

The balance-sheet floor does real work here. $277.8 million of cash, $13 million of debt, a buyback in hand. This is not a story stock that dies in a downturn. It is an R&D machine that becomes cheap or becomes right.

Kill criteria, published in advance

I will exit or refuse entry, and say so in print, if any of these trip.

1. WT Microelectronics exceeds 75 percent of revenue for two consecutive quarters. Concentration ratchets are how single-distributor stories end.

2. Insta360 is confirmed dual-sourcing or replaced in a flagship generation, or an appellate ruling reinstates an import ban on its US products.

3. The Hanwha agreement produces no disclosed design win, NRE milestone, or revenue attribution by the fiscal 2028 first half.

4. Non-GAAP gross margin prints below 55 percent in any quarter. The moat claim dies there.

5. Automotive fails to grow year over year for two consecutive quarters after this record base.

The verdict

The founder note said this name graduates to a full writeup if it survives its own risk list. It survived, with scars, and the scars are the entry.

The GAAP losses are stock comp funding a 58-percent-of-revenue R&D program, acceptable while growth holds. The concentration is real, structural, and the single best reason this trades at 7 times sales instead of 15. The patent scare is resolved and was never the right thing to be afraid of. The insider selling is compensation mechanics without a conviction signal either way. Against those four accepted risks you get the only pure-play edge AI vision silicon franchise on a US exchange, a record automotive quarter, a net-cash balance sheet, a fresh buyback, a decade-long agreement with one of Asia's largest industrial groups signed at group level, and a tape that just retraced an entire top-pick rally on no news at all.

Watching for entry in the high 60s to low 70s. August 27 is the next hard data point. If commercial-vehicle automotive prints another record and consumer IoT stabilizes, the market will not leave this at 7 times sales for long. If the quarter shows the rotation stalling, the kill criteria are above and they are not decorative.

The eyes of physical AI are for sale at a discount to the average analyst target with the promoter risk already rinsed out. That is the setup. Now it has to earn the next leg with a print.

Not investment advice. Do your own work.

References

1. Ambarella Q1 FY2027 financial results, May 28, 2026. investor.ambarella.com and SEC Form 8-K.

2. Ambarella Form 10-Q for the quarter ended April 30, 2026, filed June 2, 2026. SEC EDGAR, accession 0001193125-26-253198. Customer concentration, WT Microelectronics percentages, Arashi Vision disclosure, geographic revenue, subsequent events.

3. Hanwha and Ambarella long-term edge AI agreement, May 28, 2026. Business Wire.

4. Ambarella FY2026 results, February 26, 2026. Business Wire.

5. Rosenblatt Securities, second-half 2026 top picks, June 30, 2026, and reiterated buy, $120 target, Kevin Cassidy. CNBC, Yahoo Finance, Investing.com coverage.

6. Form 4 filings, Ambarella insiders, June 17 through July 9, 2026. SEC EDGAR, summarized via Finviz insider table.

7. US ITC, Certain Cameras and Camera Systems, GoPro v. Arashi Vision. Commission opinion February 26, 2026, PR Newswire; final determination coverage March 11, 2026, Yahoo Finance.

8. GoPro founder loan disclosure, July 9, 2026. Digital Camera World.

9. Q1 FY2027 earnings call coverage, May 28-30, 2026. Motley Fool transcript, Globe and Mail, GuruFocus highlights.

10. MarketBeat, July 15, 2026, AMBA down 8.9 percent, volume analysis.

11. Price and market data as of July 15, 2026 close, Yahoo Finance and StockAnalysis.com.

Marked to the July 16, 2026 tape. Common market cap roughly $374 million on 96.5 million shares. Cash $122.8 million against zero debt as of the March 31 balance sheet. Note the word common, because the preferred stack below changes that arithmetic more than I gave it credit for in the first cut of this piece.

The split screen

In the last ten days, Merlin printed a fresh all-time low at $3.88, down 14 percent in a week on no filing, no downgrade, and no news event of any kind.

In the same ten days:

• July 8: the trade press profiled Merlin's Air Force work and confirmed the expansion path from AFSOC to Air Mobility Command and Air Combat Command, on top of the KC-135 tanker program that runs parallel to the C-130J work. Matt George, on the record: we cannot move at the speed of the defense primes, we have to move at the speed of our adversaries.

• July 9: Cantor Fitzgerald initiated coverage. Overweight, $11 target.

• July 13: TD Cowen, which initiated at buy on the military contract base, cut its target from $11 to $8 and kept the buy. At $3.88 even the cut target is more than a double.

• July 13: Cantor published a second note flagging the milestone almost nobody covered: Merlin became the first company in history to receive an approved certification basis from New Zealand's CAA in a joint project with the FAA for an autonomous flight system. The certification thesis stopped being a promise and produced its first regulatory artifact.

• July 14: the Missile Defense Agency's Golden Dome program, the $151 billion SHIELD vehicle Merlin is qualified inside, awarded $1.75 billion in new tracking-satellite contracts. The money in that pipe is moving from authorization to obligation.

Two buy-rated shops with targets between $8 and $11. A certification first. A budget vehicle disbursing. And a $3.88 print. One of those things is wrong, and the next six weeks are the referee.

Why the tape is here, stated plainly

The drawdown is supply, and it has been supply since April. The PIPE resale registrations went effective April 17 and May 13, releasing stock into a float of roughly 29 million shares. Short interest sits near 5.7 million shares, about a fifth of the float. There has been no operating disclosure since the Q1 print, no 8-K, and a company that files nothing gives an illiquid tape nothing to price except the sellers. Revenue remains certification-gated, roughly $1 million in Q1 against a fiscal-year framework that always depended on second-half task orders and milestones. That is not a broken thesis. It is a thesis with a long quiet stretch between proof points, trading in a float small enough that patience gets marked to market every afternoon.

The balance sheet is the floor under the argument, and I am going to state it more honestly than I did in the first cut. Merlin holds $122.8 million of cash against zero debt as of March 31. But the common is not the whole capital structure. Sitting above it is $260.5 million of Series A preferred at stated value, carried at $180.3 million in mezzanine, plus an $87.8 million warrant liability. Add those to the $374 million common market cap and net out the cash, and the market is paying roughly $520 million for the operating company, not the quarter of a billion the naked market-cap math implies. That is the honest number to argue with. Cantor's $11 still implies a real re-rate from there, but anyone quoting a $374 million market cap on this name is understating what they are actually paying by about half. The tape says the burden of proof is Merlin's. Fair. Here is exactly when the proof is scheduled.

The catalyst calendar, dated

July 20 to 26. EAA AirVenture, Oshkosh. Merlin aircraft on site at the largest aviation gathering in the world. Watch for demo flights, partner announcements, and whether the Condor family gets a public airframe moment. Soft catalyst, hard photographs.

Early August. The SOCOM task-order reveal window. The FY26 C-130J task order, roughly $12.5 million, sits under the standard SOCOM disclosure clause that holds awards from public view for 60 days. The window for that hold to expire lands in August. One verifiable marker for it: federal spending records show $15.96 million already obligated against the ceiling for the PDR and CDR work, last modified February 10, and nothing new obligated since. The next draw has not touched federal systems yet, which means nobody has front-run it, and USASpending is the exact-day tripwire when it lands. This is the single most important line on this list, because it converts the $105 million IDIQ ceiling from authorization into the second funded drawdown, and it is the exact catalyst the entire ceiling-is-not-revenue framework from Part 7 was built around.

August 13. The Q2 print. Street revenue estimates run from $1.3 million to $9 million, an absurd spread that tells you nobody outside the building has the task-order timing modeled. The number that matters more than revenue: backlog, disclosed for the first time alongside whether the loss narrows on plan. Second most important line on the list.

October 15. The preferred conversion reset. I had this date wrong until I went back to the Certificate of Designation, so here it is corrected. The Series A carries a $12.00 stated value and an initial $12.00 conversion price. On the twenty-first trading day after the six-month anniversary of the March 16 close, which lands around October 15, the conversion price resets to the greater of the twenty-day VWAP and a hard floor of $5.00. At $3.88 that reset to the floor is effectively locked in. It roughly doubles the eventual dilution, from about 23 million shares at the old strike to about 55 million at the floor, against 96.5 million common outstanding. The floor is also the protection: the price cannot reset below $5.00, so this is not a death spiral. And conversion is at the holder's option, not management's, which means the preferred simply does not convert while the stock sits under five. It accrues at 12 percent instead, and that accrual is the real cost of waiting.

Second half. The certification cadence. Merlin cleared the SOI-2 audit with New Zealand's CAA on October 28, 2025, so the stage-of-involvement work was already well underway and did not begin with this month's news. What the July certification basis adds is the agreed regulatory foundation, jointly with the FAA, that the remaining audits get measured against. The path from here is the rest of the SOI cycle, verification and conformity, then the supplemental type certificate that unlocks the commercial model. Each audit is a dateable event on a regulator's calendar, and Merlin's New Zealand subsidiary at Kerikeri is the operational proving ground. The 2027 commercial framework lives or dies on this cadence.

Second half. KC-135 program milestones. The tanker program is real, parallel to the C-130J work, aimed at the roughly 300-aircraft fleet, with the July trade profile confirming customer pull beyond AFSOC into Air Mobility Command and Air Combat Command. One caution I flagged before and will keep flagging: the separate Air Force KC-135 cockpit-refresh solicitation making the rounds is an avionics obsolescence program that favors the incumbent, not an autonomy award. Do not let anyone sell it to you as a Merlin catalyst. The Merlin tanker catalyst is task orders and demonstrations on the autonomy program itself.

Second half. Golden Dome and DAWG flow-through. The SHIELD IDIQ is a $151 billion ceiling with over two thousand qualified vendors, so qualification alone is worth little. What is worth something: the July 14 awards prove obligations are flowing, and the FY27 defense appropriations cycle prices the DAWG autonomy accounts Merlin's programs draw against. Watch for task orders, not press releases.

Standing tripwires, no dates. The GM Maritime seat is still posted and open, and a fill there is the org-chart signal that the maritime line is spooling. The GM Tactical Autonomy seat, vacant since last November, has come off the careers page entirely, which means it was either filled quietly or shelved. Worth knowing which. The first New Zealand line-operations or Part 135 posting still marks the commercial 2027 spool-up beginning. And one correction to my own prior framing: the venture debt is gone. The March 31 balance sheet carries zero long-term debt and zero convertible notes, both extinguished at the de-SPAC. There is no 2027 refinancing wall. The senior claim that matters on this cap table is the preferred, not a loan.

What the bear needs and what the bull needs

The bear case at $3.88 needs the August task order to slip past its window, the Q2 backlog disclosure to disappoint, and the October reset to land with the stock still far under five, locking in the wider dilution. That is a real parlay and it pays the shorts, a fifth of the float, if it hits.

The bull case needs exactly what is already scheduled: one funded task order, one clean backlog number, one more certification stage. No heroics, no new contracts, no narrative change. The difference between $3.88 and the $8 to $11 the two covering analysts publish is not a story the company still has to invent. It is a calendar the company has to execute.

None of the seven kill criteria have tripped. The cash covers the burn deep into 2027 before the preferred conversation gets hard. I added to the position into this drawdown and the add is underwater with the rest of it. The calendar above is the list of dates on which I find out whether that was conviction or stubbornness, and you will read the verdict here either way, in print, against the same criteria I published on the way in.

Not investment advice. Do your own work.

References

1. Cantor Fitzgerald initiation, Overweight, $11 target, July 9, 2026, and certification note, July 13, 2026. Investing.com, Daily Political, Ticker Report coverage.

2. TD Cowen, buy rating, target cut from $11 to $8, July 13, 2026. MarketBeat, Moomoo coverage.

3. CAANZ and FAA joint certification basis, first approved certification basis for an autonomous flight system, per Cantor Fitzgerald note, July 13, 2026.

4. Inside Unmanned Systems, Merlin Labs: Bringing Autonomy to the Air Force's Bread-and-Butter Fleet, July 8, 2026. C-130J program history, KC-135 parallel program, AMC and ACC expansion, Matt George quotes.

5. Golden Dome tracking-satellite awards, $1.75 billion, July 14, 2026. Washington Times, Benzinga, GovCon Wire. SHIELD IDIQ program scope, Missile Defense Agency, December 2025 qualification lists.

6. USSOCOM C-130J IDIQ, $105 million ceiling, and task-order disclosure mechanics. Prior coverage, Parts 7 and 10, research.shawarmacapital.net.

7. Series A preferred terms: $12.00 stated value, 12 percent cumulative semi-annually compounding dividends paid in kind, holder-optional conversion, conversion price reset to the greater of the 20-day VWAP or a $5.00 floor on the twenty-first trading day after the six-month anniversary of the March 16 close, company call at 150 percent of accrued value, holder put after the fifth anniversary. Merlin, Inc. Form 10-Q for the quarter ended March 31, 2026, filed May 15, 2026, Note 10, Mezzanine Equity. Balance sheet: cash $122.8 million, zero debt, warrant liabilities $87.8 million, Series A carried at $180.3 million. Corrects the mid-August framing in the first cut of this piece.

8. PIPE resale registration effectiveness, April 17 and May 13, 2026. SEC filings.

9. Market data as of July 16, 2026, Yahoo Finance. Short interest and float, exchange-reported.

10. Prior series: MRLN, The Complete Thesis, research.shawarmacapital.net/p/mrln-the-complete-thesis.

Closed this cycle

Three positions came off the book, one a rotation and two fast round-trips.

PGY, +37 percent, rotated. In mid-June I sold Pagaya at $15.68, up 37 percent from my $11.47 basis, and rotated the proceeds into building Ouster and Amprius. Pagaya did nothing wrong. It printed its first clean GAAP profit and the tape finally repriced it. I took the win and moved the capital into two names with nearer catalysts.

BGDE, +14 percent. I bought Big Digital Energy at the first-post level on June 17 and sold at $11.10 into the re-rate a week later. A failed bitcoin miner sitting on scarce, grid-connected PJM power, insiders buying from four dollars to over eight, and a shell repricing from sixteen to forty million dollars. The trade was the re-rate, and it came fast, from the mid-eights through eleven and change and straight back down. That swing is what I stepped aside from. I still cover the name and shipped a Part 2 on the financing tell, and I am watching for a cleaner re-entry.

CCXI, +19 percent. I bought Agility Robotics through the CCXI trust at $16.40 on June 29, as it broke off the ten-dollar floor, and sold at $19.50 into the spike two sessions later. The stock ran from ten to nearly twenty in two weeks and gives back three and four dollars in a day. I booked the move and I am letting the volatility cool before I decide on re-entry.

The held book, by return

Nine positions now. Four up double digits, from a Rubrik near a double to an AmpliTech at plus forty-nine. Ouster roughly at cost. Four underwater, including one of the two June adds and the flagship I just sized up into weakness.

RBRK, +98 percent. Rubrik nearly doubled from my basis. The Q1 fiscal 2027 print beat every metric and raised guidance across the board, subscription ARR up 32 percent to $1.57B, revenue up 39 percent, and the subscription contribution margin widened to 13.2 percent from 8 percent a year ago. The AI control plane thesis is playing out in the numbers. No exit.

IQEPF, +81 percent. IQE signed the multi-year indium phosphide epiwafer supply agreement with Tower Semiconductor for AI data-center optics and settled the old IP dispute in the same stroke. The stock spiked on the print and gave some back into the semis pullback. Position grew through price action, no fresh capital added.

LPKFF, +80 percent. The activist standoff went to a vote on June 4 and management won. The German listing gave back part of the pop into the small-cap pullback and trades near seventeen and a half euros, still a double from my basis. The glass-substrate AI packaging monopoly thesis is intact and Part 4 is live.

AMPG, +49 percent. AmpliTech is a ten percent position I opened on June 1, on its own and separate from the Pagaya rotation. It ran through the seven-dollar target I underwrote it against and pulled back to the high sixes with the microcaps. A twenty-four-year-old RF house turned 5G open-RAN radio OEM with a quantum tail, 48 percent gross margins and zero debt. The first conversion is on the record, FCC and ISED cleared the radio, and the BEAD program money is moving. Return on my June 1 basis, still the widest asymmetry in the book. Note it is a different company from Amprius below.

OUST, +6 percent. Ouster is now a full ten percent position, built through June and funded by half the Pagaya rotation. The Pentagon widened the China lidar ban to the top two suppliers, the Rev8 factory qualified for federal infrastructure dollars under Build America Buy America, and the company raised two hundred million into strength. Blended basis around forty-one dollars, trading in the low forties. A conviction position now.

BMM, flat. Blue Moon Metals sits right on my basis after the consolidation. Springer in Nevada is still the only permitted US-domiciled ammonium paratungstate hub, and Nussir in Norway carries the EU Critical Raw Materials strategic designation. It is one of the two largest positions in the book, roughly a fifth of it, and I am still building toward the target weight. The probability-weighted expected value against a six-and-a-half-dollar print keeps the asymmetry wide.

MRLN, -17 percent, doubled down. The one I owe you the straight version on, and the biggest position change this month. Merlin is my highest-conviction name, and when the PIPE resale registrations went effective and pushed the stock into a thin float down to a fifty-two-week low near four-thirty, I treated the supply as a gift and added about a hundred thousand shares into the forced selling. That takes Merlin back to the top of the book at around thirty percent and pulls my blended cost down from $6.03 toward the low fives, so the mark is nearer minus seventeen than the minus thirty it would show on the old basis. The move was supply, not story. The C-130J critical design review cleared with USSOCOM, the program moved into aircraft integration on the hundred-million-plus IDIQ ceiling, the defense budget the thesis points at keeps inflecting, and none of the seven kill criteria have been met. I built Merlin World so you can walk the fleet yourself.

CODA, -17 percent. Coda Octopus is still below my basis after the microcap drawdown. Real-time 3D sonar for defense and offshore energy, profitable and debt-free, with defense primes including Anduril moving into the channel. Thesis intact, and the plan is to size back in on dips.

AMPX, -29 percent. Amprius is the other half of the Pagaya rotation, and it is underwater. I built it to a ten percent position in June around sixteen dollars, and the silicon-anode complex has since sold off hard with the rest of the microcaps, taking it to the low elevens. The operating case has not broken. Amprius is the only Western silicon-anode pure-play, Q1 revenue was up two and a half times year over year to $28.5M, it printed its first twenty percent gross margin, and it raised the FY26 guide to a $130M floor with positive adjusted EBITDA. Defense and drone customers include Anduril, Skydio, and Switchblade. Q2 earnings on August 6 are the next read. I show it red because it is red.

How the book is weighted

After the add, Merlin re-anchors the book near thirty percent, with Blue Moon Metals and IQE right behind it around a fifth each. AmpliTech, Ouster, and Amprius each sit near ten percent from the June builds. The winners keep their weight, the losers stay sized to conviction, and nothing is dressed up.

What I am watching into August

• MRLN, the next SOCOM C-130J task order window and the Q2 print with backlog, the first hard read on whether the loss narrows on plan now that I am sized up.

• AMPX and OUST, both report Q2 on August 6, the first look at each balance sheet after their raises and the first checkpoint on the two new adds.

• BGDE and CCXI, both on re-entry watch once the near-term volatility settles.

• BMM, continued build toward the target weight as the Springer and Nussir permitting and financing steps land.

The full theses

Every position, in one place:

• MRLN: The Complete Thesis

• IQE: The Epiwafer Monopoly

• BMM: Blue Moon Metals

• LPKF: The Glass Monopoly

• RBRK: The AI Control Plane

• CODA: The Underwater Monopoly

• AMPG: AmpliTech

• OUST: Ouster

• AMPX: Amprius

• Every portfolio update

See you at the end of the month.

Long all held positions. The three above are closed. Not investment advice.

That is the setup. In Part 1 we said the single re-rating event was the first AI offtake, and that you were getting 129 megawatts of scarce PJM power at a liquidation price because the market assumed the offtake never comes. Nothing since has changed that. What changed is that the team just assembled the pieces you assemble right before you close one: the site, the power partner, the banker, and the cash. You do not spend that money and sign that paper unless a counterparty is close. We read the financing as the tell, and we are still buyers.

This is Part 2. It covers what happened, why the partner matters, how we read the dilution, and what has to land next.

Read more

Agility Robotics builds Digit, a bipedal humanoid robot that already works paid shifts moving totes in real warehouses, which by itself puts it ahead of a field that is still mostly demos. It is going public through Churchill Capital XI, ticker CCXI, in a deal that values it at two and a half billion dollars, a fraction of what the market marks its closest private rival. But the reason this name rewards actual work is that the robot is the least interesting thing about it. Underneath sits a rented global service and software channel that solves the hardest problem in the sector, a cap table that reads like a supply chain with every major backer owning a different link, a government and defense leg that every model carries at zero, and a coordinated regulatory campaign to wall its Chinese competition out of the American market, all timed to the listing. It is also a pre-revenue company whose flagship product has slipped two years and whose real numbers do not appear until the S-4 prints. This series holds both sides at once, the unpublished edges and the honest risks, and updates as each new filing and each new post lands.

Last updated: 2 July 2026. This page is a living document and is updated as each new post goes live.

THE SERIES

Part 1 - The Robot That Already Clocks In

Shawarma Capital

$CCXI / $AGLT The Robot That Already Clocks In - Agility Robotics

Long $CCXI. Everything here is built from public filings, the deal announcement, and public market data. Nothing here is investment advice. For educational purposes only. I may hold positions in names I discuss. Do your own research…

Read more

a month ago · 3 likes · 2 comments · Shawarma Capital

The setup. Digit already works paid shifts at GXO, Schaeffler and Toyota, and CCXI is the only liquid, pre-close way to own it before it relists as AGLT.

The deep file. A three-front campaign to wall Chinese humanoids out of the US market, the rented go-to-market, the consortium cap table, the listed Taiwan read-through, and the honest bears.

HOW TO READ THIS

Read Part 1 first if you are new to the name. It carries the full setup. Read the most recent post if you are already caught up, since the thesis moves with the tape and the filings. Every part is written to stand on its own, but they compound in order.

Disclosure. Everything here is built from public filings and public data. Research synthesis for educational purposes, not investment advice. I am not a registered investment advisor and I do not owe you a fiduciary duty. Positions and price scenarios discussed in the individual posts are illustrative, not forecasts, and the inputs can be wrong. Do your own due diligence.

Every expensive chip built today has to survive an oven before it is allowed to ship. The part is mounted on a specialized board and run hot, at elevated temperature and voltage under electrical load, for hours or days, to force the weak units to fail in the lab instead of inside a customer's data center or car. The screen is called burn-in, and for an AI accelerator that sells for tens of thousands of dollars or an automotive controller where a field failure is a safety event, it is not optional. Trio-Tech International sells the boards that go in the oven, the reliability equipment around them, and the oven time itself as a service, from its own floors across Asia. It is a sixty-seven-year-old microcap that almost nobody covers. Its one public pure-play cousin, Aehr Test Systems, trades near seventy-five times sales while losing money, because the market has decided burn-in levered to artificial intelligence is worth paying for. Trio-Tech trades near two times sales while making money and growing faster. This series is the argument that the distance between two and seventy-five, for a profitable company on the same rung of the same ladder, is the widest mispricing in the book, and a running account of the prints that either confirm it or break it.

Last updated: 2 July 2026. This page is a living document and is updated as each new post goes live.

THE SERIES

Part 1 - Every AI Chip Has to Survive the Oven First

Shawarma Capital

Every AI Chip Has to Survive the Oven First - $TRT

Long $TRT. Everything here is built from public filings and public data. Nothing here is investment advice. For educational purposes only. I may hold positions in names I discuss. Do your own research…

Read more

2 months ago · 6 likes · Shawarma Capital

The thesis. Burn-in as the mandatory screen for AI silicon, the Aehr comp at 75 times sales versus two, and the four-scenario tree on ten million shares.

Part 2 - The Margin Verdict

Shawarma Capital

$TRT - The Margin Verdict

As I write this on 2 July 2026, Trio-Tech trades near $10.26, down about 7.6 percent on the day and down roughly 30 percent from the $14.82 where I wrote it up two weeks ago. The fiscal year that will settle the central question of this thesis ended two days ago, on 30 June. In the same fifteen sessions that took the stock down a third, the business did not shrink. The order book kept filling, the Penang buildout kept staffing, and the trailing revenue base grew. The stock is now cheaper, on a bigger company, than it was when I said it was mispriced. It trades at roughly 1.8 times trailing sales against a comp the market still pays about 75 times sales. The gap I described did not close. It widened…

Read more

a month ago · 2 likes · Shawarma Capital

The stock is thirty percent cheaper and the business is bigger. The one line on the coming full-year print that decides whether the re-rate is now or later.

HOW TO READ THIS

Read Part 1 first if you are new to the name. It carries the full setup. Read the most recent post if you are already caught up, since the thesis moves with the tape and the filings. Every part is written to stand on its own, but they compound in order.

Disclosure. Everything here is built from public filings and public data. Research synthesis for educational purposes, not investment advice. I am not a registered investment advisor and I do not owe you a fiduciary duty. Positions and price scenarios discussed in the individual posts are illustrative, not forecasts, and the inputs can be wrong. Do your own due diligence.

Ouster makes digital lidar, the laser eyes that let machines see in three dimensions, for robots, trucks, factories, and defense. For years the story was a commodity price war against a Chinese champion that could always undercut on cost. Then the Pentagon changed the rules. It moved to ban the Chinese leader from United States government and defense work on national-security grounds, and the ban has only widened since. That is the pivot this series tracks. A hardware company fighting a losing price war becomes the domestic beneficiary of a policy that fences off its largest competitor from the highest-value end market, while its own factory scales into the gap. Each part below marks the widening of the ban and the scaling of the supply, and the read on what it does to a name the market still prices like the old commodity story.

Last updated: 2 July 2026. This page is a living document and is updated as each new post goes live.

THE SERIES

Part 1 - China's Lidar Champion Just Got Banned

Shawarma Capital

$OUST: China's Lidar Champion Just Got Banned From The Pentagon. Ouster Is The Western Pure-Play Left Standing

Position disclosure: I do not own OUST as of this writing. I am opening a starter position this week, sized off stable holdings. Shawarma Capital, June 7, 2026. OUST closed near $39.68 on June 5, down almost sixteen percent on the day, the session after it touched a fifty-two-week high near $47. This post is research synthesis, not investment advice. Th…

Read more

2 months ago · 10 likes · Shawarma Capital

The ban thesis. The Pentagon shut the door on the Chinese lidar leader and Ouster is the domestic beneficiary.

Part 2 - The Ban Widened, The Factory Scaled

Shawarma Capital

$OUST: The Ban Widened, The Factory Scaled, And Here Is The Number Built Driver By Driver

Position disclosure: I own OUST. I opened the starter I flagged in Part 1 and I am sizing it as described in Section XV below. Shawarma Capital, June 2026. OUST closed near $42 on June 28, after running to a $51.50 fifty-two-week high in mid-June from the $39.68 close on June 5 that Part 1 was written against. The pullback matters, and I treat it as a gift rather than a warning. This is research synthesis, not investment advice. The full disclaimer is at the bottom…

Read more

a month ago · 1 like · Shawarma Capital

The ban widened, the factory scaled, and here is the updated read on the setup.

HOW TO READ THIS

Read Part 1 first if you are new to the name. It carries the full setup. Read the most recent post if you are already caught up, since the thesis moves with the tape and the filings. Every part is written to stand on its own, but they compound in order.

Disclosure. Everything here is built from public filings and public data. Research synthesis for educational purposes, not investment advice. I am not a registered investment advisor and I do not owe you a fiduciary duty. Positions and price scenarios discussed in the individual posts are illustrative, not forecasts, and the inputs can be wrong. Do your own due diligence.

AmpliTech spent twenty-four years as a quiet radio-frequency component house before it did something most small caps never manage. It reinvented itself as an original equipment manufacturer of 5G Open-RAN radios, the boxes that carriers bolt to towers and rooftops to actually move the traffic, and it kept a cryogenic amplifier line pointed at quantum computing on the side. The reason the name matters now is Buy-America. The federal BEAD broadband program routes tens of billions of dollars into domestic network buildout, and it favors equipment made outside of the incumbent Chinese and European supply chain. AmpliTech is one of the few American radio OEMs certified to catch that money. The through-line of this series is conversion. A thesis is only worth the paper it prints on until the letters of intent turn into purchase orders, the certifications turn into deployments, and the government dollars turn into revenue. Each part below marks another step of that conversion in real time, from the first look at a five-dollar stock nobody trusted to the full bull discounted-cash-flow model on a name where the hardware is now cleared to ship.

Last updated: 2 July 2026. This page is a living document and is updated as each new post goes live.

THE SERIES

Part 1 - Forty-Eight Percent Gross Margins, Zero Debt

Shawarma Capital

$AMPG: Forty-Eight Percent Gross Margins, Zero Debt, And A Five Dollar Print Nobody Trusts Yet

Shawarma Capital. May 25, 2026. AMPG is $5.04 against a Friday close. I have no AMPG position as of this writing. I am opening a starter position via daily buys Tuesday May 26 through Friday May 29, 2026, sized off stable holdings, with the conditional add line set below. This post is research synthesis, not investment advice…

Read more

2 months ago · 11 likes · 1 comment · Shawarma Capital

The setup. A 5G Open-RAN radio OEM with a quantum tail, 48 percent gross margins and a clean balance sheet, before I owned it.

Part 2 - The Starter Becomes a Ten Percent Position

Shawarma Capital

AMPG Stock: AmpliTech’s 48% Gross Margins, Open-RAN Radios, and a $7 Target at $4.42

I told you in Part 1 I would dollar-cost-average into AmpliTech. I did not. The plan was four tranches across late May. I skipped all of them. The average of the four closes I would have bought was about $4.76. This morning, June 1, the stock opens near $4.42. I am opening the full position in one move and sizing it at ten percent of the book…

Read more

2 months ago · 3 likes · 1 comment · Shawarma Capital

The stock fell on no news, never tripped my add line, and the thesis did not change. Opening a ten percent position and explaining the size.

Part 3 - The First Conversion

Shawarma Capital

$AMPG: The First Conversion. FCC And ISED Cleared The Radio, And The BEAD Money Is Already Moving

This post is research synthesis, not investment advice. The full disclaimer is at the bottom…

Read more

2 months ago · 6 likes · Shawarma Capital

FCC and ISED cleared the indoor 5G radio, the BEAD program signed award agreements with fifty-two states, and I built the full bull DCF.

HOW TO READ THIS

Read Part 1 first if you are new to the name. It carries the full setup. Read the most recent post if you are already caught up, since the thesis moves with the tape and the filings. Every part is written to stand on its own, but they compound in order.

Disclosure. Everything here is built from public filings and public data. Research synthesis for educational purposes, not investment advice. I am not a registered investment advisor and I do not owe you a fiduciary duty. Positions and price scenarios discussed in the individual posts are illustrative, not forecasts, and the inputs can be wrong. Do your own due diligence.

For this one I did not read the press release. I read the 118,000 word merger agreement and its exhibits. I pulled every quarterly filing on the Senate lobbying database. I read the patent recordation ledger at the reel and frame level. I read the actual Section 232 tariff comment Agility filed with Commerce. I found the one bill of lading in the customs record. I read the footnotes in the investor deck that carry the numbers the headline never shows. Almost none of what follows is in the coverage, and a few of the loudest things being said about this deal are simply wrong. I will show you both.

Here is the thesis in one sentence. Agility Robotics is not just building a humanoid. It is quietly building a wall around the American market for humanoids, on three separate fronts at once, and it is timing that wall to its own public listing. The robot is the thing standing behind it.

The three-front campaign, and the gate built into it

Start with the piece that took the longest to find, because it is spread across three federal agencies and no single filing spells it out.

Front one is a bill. The GUARD Act extends the mechanism the government used to ban Huawei and ZTE, the FCC Covered List, to humanoid and quadruped robots. The bill text names no company, but the sponsors' letter names Unitree, the Chinese leader, and calls for it to be run through the Covered List, the Defense Department's 1260H list, and the Commerce Entity List. Agility is a named endorser, and it is the only robot manufacturer among the six endorsers. This is not theoretical. Unitree was actually added to the DoD 1260H list on the eighth of June, five days after the bill dropped, banning it from Defense contracts.

Front two is a standard. There is no finished safety standard in the world today for a humanoid working around people. The first one, ISO 25785-1, is being written right now, and the person holding the pen as its Project Leader is an Agility engineer named Kevin Reese. The same Reese is the inventor on Agility's cooperatively-safe fall patent. One person is building the moat in the patent office and in the standards body at the same time. When that standard publishes, certified compliance becomes the license to put a robot next to a worker.

Front three is a tariff. Commerce opened a Section 232 national-security investigation into imported robotics last September, and I pulled Agility's comment. Its Chief Business Officer wrote to the Secretary, in these words, that tariffs on imported finished humanoid robots such as the ones Agility makes in the United States would be helpful, and backed it with the claim that only twenty percent of its supply comes from overseas and no more than one percent from China. A Section 232 clock runs 270 days, which puts the decision at roughly the end of May 2026, inside this deal's closing window.

Now the detail that makes the whole thing sharper, and that I have seen nowhere. The GUARD Act's rule of construction carves out NATO and major non-NATO allies. It is drafted to wall out China and to spare Taiwan, which is where Foxconn, the lead investor in this very deal, lives. The fence has a gate, and the gate is sized exactly for the money at the table. Even the standards body splits on the tariff: A3, which co-authors the ISO standard with Agility, filed against the tariff. Agility is the one actor holding the standards pen, filing pro-tariff, and endorsing the ban, all at once.

The lobbying that front-ran the policy

Agility retains two Washington firms. New Lantern works the executive branch, and it is effectively the CEO's personal shop, run by her former Microsoft chief lobbyist. SIO Advocacy, registered in the same quarter the deal was signed, works Congress through a fresh Hill exit with an armed-services background. Two vectors, deliberately.

The money points at the Defense Logistics Agency, the Maritime Administration, the Defense Department, the Office of Science and Technology Policy, and the Office of the Vice President of the United States. About 110,000 dollars in disclosed spend, but a target list that tells you Agility sees military warehousing and shipyard logistics as its market. And the tell that it works: Agility lobbied Maritime in the back half of last year, and in February the White House released a Maritime Action Plan that funds shipyard automation. They lobbied the door before it opened. Agility is also a named supporter of a second bill, the National Commission on Robotics Act. Tariff, ban, and national strategy, all at once.

They rented the hardest part of the business

The single most common objection to any robotics company is service. A robot breaks, needs parts, needs a human to show up. How does a company of four hundred people keep thousands of machines running across a continent? For most robot startups the honest answer is that they cannot, and it kills them.

Agility rented the answer, from names you already know. Field service goes to Ricoh, the company whose technician already fixes the copier in your office. The software integration goes to Manhattan Associates, the warehouse-management system most large warehouses already run. Selling and installing goes to two national integrators, Tompkins and Zion. Agility built the distribution and support layer of a mature robotics company while everyone stared at the robot. This is not in a single piece of the coverage. And the software that ties it together reaches past its own machines. Through Agility Arc, Digit picks an item and then dispatches a Zebra or MiR mobile robot to carry it, a handoff that is live today at the GXO site in Atlanta, which makes Arc a control layer sitting above other companies' robots and not only its own.

The cap table is a supply chain, and one of its own suppliers is public

Every article lists the same investors. None explains what they are doing, which is the point. This is a value chain, and every major name owns a different link. NVIDIA sells the compute, provides the simulation, holds equity. Schaeffler makes humanoid actuators for the market, holds equity, and buys Digits for a US plant. Amazon runs a pilot and holds equity. Sony and SoftBank are on the roster. And on the manufacturing side sit two Taiwanese strategics, Foxconn, which leads the financing, and the parent of an optics maker I will come to.

The reason it matters is that supply, demand, and compute are all de-risked at once, because the suppliers, the customers, and the chip vendor are the owners. That said, read the PIPE honestly. Of the roughly 200 million dollar PIPE that everyone calls institutional validation, more than 60 million is existing insiders re-upping. The genuinely new outside money is under about 140 million. It is a real vote, just a smaller one than the headline reads.

Here is the full funding history, reconciled from the filings, which has never been laid out cleanly in one place.

The order book is not what it looks like

This is the sharpest new thing I found, and it changes how you read the whole story. The headline order everyone cites, more than 300 million dollars, over a thousand Digit v5 robots. Read the footnote in the deck and it is not what it sounds like. It is total contract value, not revenue. It is for a robot, v5, that has not shipped, so zero of the thousand are deployed. It is milestone-gated. And the contract issues the customer Agility warrants that vest as robots roll out. A customer that gets paid in the seller's equity is not a customer in the normal sense, it is a strategic, which points straight back at the circular cap table. Nobody in the coverage has flagged the warrant structure. The order is real, but it is a bookings-and-alignment instrument, not a backlog you can multiply by a margin.

You can go further on who signed it, because the arithmetic boxes it in. A thousand robots at eight thousand five hundred dollars a month across three years is three hundred and six million dollars. That is not thirty customers adding up to a number. It is one customer, and the entire disclosed backlog is that single contract. The fingerprints point at Amazon. Paying a supplier in warrants that vest as the machines deploy is Amazon's own documented playbook, and Amazon is the one named customer whose deployment the company has openly paused until v5 ships, which is exactly the robot this order is for. The name most people would reach for, Schaeffler, is the one I can most nearly rule out, because Schaeffler's own public thousand-robot program is with a different manufacturer and its stake in Agility is only about ten million dollars. The tell will not wait for the S-4. If the customer is Amazon, an ownership filing on those warrants can surface before the merger proxy does. That is the thing to watch.

And the price that sits under it finally leaked, in the same deck. Digit rents for 8,500 dollars a month, plus a one-time deployment fee, against a five-year life. Call it a hundred thousand a year per robot. That is the number that makes the payback math checkable for the first time, against a fully loaded human at 30 dollars an hour, which is roughly 60,000 a year for a single shift. The robot pays only at high utilization or across multiple shifts. Now you can argue the thesis with real numbers instead of vibes.

The listed ways to play it that nobody has spotted

There are two, and both are unpublished. The first is Ability Enterprise, a Canon-lineage Taiwan optical maker, ticker 2374 in Taipei. Its parent invested in Agility, its Canon partnership co-develops Digit's cameras, and its stake is doubling to about 20 million dollars into the listing. Ability guides humanoid optics from a single-digit share of its sales toward twenty percent by 2027, and it is positioned to displace Intel's RealSense as Digit's vision supplier. A public company, in a different market, whose revenue is levered to Agility. Be honest about the tape, though. The stock has not actually moved on the Agility news and its turnover cooled after the deal, so this is a supply-chain story the market has not paid for yet, not a proven catalyst.

The second I found in the customs record. There is exactly one ocean bill of lading under Agility's name, a shipment of battery chargers from Delta Electronics, ticker 2308 in Taipei, a Taiwanese power giant. Two things fall out of that. Delta is a second listed supplier tied to Digit by an actual shipping document. And the fact that the entire visible import footprint is a single container of chargers is itself proof of the domestic-assembly claim, because the cameras, cells, and compute do not enter as ocean freight under Agility's name. Be precise, because precision is the game here. The big checks came from the private parents, and some of the Taiwanese press has the valuation wrong by four or five times, so trust the SEC dollars, not the local headlines.

The Taiwanese tie runs deeper than optics and chargers. The same group behind the camera work has also taken the rights to sell Agility's robots across parts of Asia and has told its own investors it will handle some of the manufacturing. Its listed holding arm already booked a revenue jump of almost four hundred percent late last year that it credits to the robot business, before v5 has shipped a single unit. So when these machines start selling into Asia, the camera inside the robot and the channel that sells it both sit inside Agility's own cap table. That is a related-party loop, and it is exactly the kind of thing the first audited filing has to break out.

The government leg is a zero in every model

Every model carries defense at zero. That is a blind spot, because there is already a contract. The Naval Research Laboratory bought Digits, 510,000 dollars, described as bipedal humanoid robots for interaction. It is the only federal award to date and there has been no follow-on, so do not oversell it. But pair it with the lobbying targets, the Defense Logistics Agency and the Maritime Administration, and with the Defense Department publicly working robots into military logistics, and you have the outline of a defense leg worth exactly zero in consensus. Any dollar of it is upside.

The backend is real infrastructure, not a slide

I mapped Agility's cloud out of public DNS and certificate records, and this is a genuine modern production system, not a demo. Every robot authenticates through an Auth0 machine-to-machine identity. There is a separate European identity tenant, the concrete mechanism for serving a customer like Schaeffler in Germany under European data rules. There is a live SOC 2 compliance program on a Vanta trust center, which is what enterprise buyers demand before they let a vendor's software touch operations. And there is an invite-only partner sandbox that mirrors the robot-telemetry stack, a real third-party integration program at the plumbing layer. None of it is marketing.

The brain is rented, and the reflexes are the only part it owns

Here is the piece the deck works hardest to blur. Agility calls itself a proprietary physical AI platform. Read its own engineering pages and the intelligence splits into three layers by speed. The slow layer, the part that understands a task and plans the steps, is a large language model, and Agility does not own it. The fast layer, the sub-millisecond balance and whole-body coordination, is a reinforcement-learning controller with fewer than a million parameters, and that one is Agility's. So the smarts above the neck are bought, and only the reflexes below it are built.

Be precise about the Google tie, because it is thinner than it sounds. Agility is one of roughly sixty outside testers of Google's robotics reasoning model, not a build partner, and the company that got the real Google build partnership is a competitor, Apptronik. Agility's own AI page does not even name it. The reasoning brain is a swappable slot the company deliberately keeps multi-sourced, because that layer is not its edge. The whole thing runs on an NVIDIA computer inside the robot.

Which sets up the dependency nobody has drawn. NVIDIA is not one supplier here, it is four at once. It sells the onboard compute. It supplies the simulator the robot is trained in. It holds equity through its venture arm. And the deepest hook, the safety certification that the entire enterprise-sales case rests on, now routes through NVIDIA, because Agility is the first company to fold NVIDIA's Halos safety system into its own, and the certification runs through an inspection lab NVIDIA itself accredits. The procurement moat I described in Part 1, the safety story that lets a robot stand next to a worker, is real, and it is built on a rail one vendor owns. NVIDIA owns the rails. Agility owns the train.

One last thing to kill, because it is the loudest wrong thing circulating about this company. The viral clip of Claude running a robot is a different robot, a four-legged dog from Unitree, out of Anthropic's own lab. Agility is an enterprise Claude customer for its back office. Claude does not run Digit. Anyone trading this on an intelligence-in-the-robot story is trading a fact that does not exist.

The IP is not the moat the deck sells, and I had that wrong the first time

I flagged the Oregon State IP as a bear in the first draft. Having dug the university's own records, I am softening it, because the risk is smaller than it looks. The foundational leg patents are Oregon State's, DARPA-funded, and Agility licenses them, and the university holds equity. But the equity was non-dilutable only until an early round, so it was diluted through more than 600 million dollars of raises to something immaterial, so immaterial that Oregon State's audited financials do not even itemize it and the deal's own IP risk factor never mentions the license. The royalty on a university spinout of this kind runs below the standard rate by policy. And the DARPA manufacturing preference aligns with Agility assembling in the United States, which its founder testified to the Senate. So the borrowed-IP bear is mostly a footnote.

Two real things survive. First, Oregon State is a policy-mandated forced seller. It must convert its equity to cash within about 180 days of the shares trading, which is a certain, dated supply overhang from a holder with no reason to stay. Second, and separately, the real moat is not the university patents at all. The humanoid competitors do not even cite them. The moat is Agility's own drivetrain, its in-house actuators, and a next-generation cycloidal-gear actuator program its engineers have been patenting since early 2024. The story is in the joints, not the pedigree.

The deal mechanics, read from the actual agreement, and the flow already forming

The float is not retail's. Seven institutional funds, including a Millennium stake I have not seen reported anywhere, hold about forty-five percent of the public shares. They are passive holders and at a price well above trust none will redeem, so the concentration is the fact, not control of a vote.

The deal is close to un-terminable. No break-up fee, no earnout, no sponsor forfeiture in the entire agreement. The fairness opinion, unusually, was signed not by a bank but by an intellectual-property appraisal boutique, and it is fair only to the shareholders other than the sponsor, which tells you the value is being appraised as intellectual property. The sponsor economics are their own tell. The Klein fee runs on four separate streams, a retainer, a cut of any financing, a cut of any strategic investment he introduces, and an advisory lane, and the same person signs for both the Klein side and the company as its CFO. And there is a cash tell in the fine print, an authorized pre-close bridge of up to 12.36 million Agility shares, carved out of the minimum-cash test, which is not the sign of a company swimming in money.

The flow is already being packaged before the stock even trades. A two-times leveraged single-stock AGLT ETF is in registration, filed two days after the deal. A First Trust fund already carries pre-IPO Agility alongside Anthropic, xAI, and Anduril. When the market wants the humanoid trade, there is one ticker for it, and the vehicles to lever it are being built in advance.

Two structural facts I have found nowhere else. SoftBank, the most active robotics investor on earth, tried to buy Agility outright at about 900 million dollars in early 2025, then walked, took only a minority stake, and put its real money into a rival generality bet instead. A sub-billion walk-away from that buyer is a hard anchor sitting under the 2.5 billion mark. And the chief executive of the only US-listed pure-play humanoid sits on the board of BlackRock, the largest ETF sponsor in the world, a director of the exact structural buyer the scarcity thesis leans on. Read both however you like, but neither is nothing.

And the tell that should keep you honest. The private secondary market prices Agility below its own deal. Forge marks it around 71 dollars a share and Hiive around 72, roughly twenty percent under the deal's own 88 dollar per-share figure and barely half of where the SPAC trades in the high teens on a look-through basis. The last private mark before the deal, from Nasdaq Private Market, was lower still, below the price of the company's own most recent funding round. The people closest to the company are paying less than the deal implies, not more. The counterweight is that the structure makes it almost impossible to bet against, there are no listed options on the name and the borrow is thin, while a levered long ETF is already drafted, so the only easy trade at listing is up. Hold both of those at once. The setup is asymmetric, and the asymmetry is exactly why the discipline is in the sizing and the calendar, not in the conviction.

Three more facts about the flow, none of which I have seen printed. A fund is already long this stock before it has closed. Roundhill's humanoid exchange-traded fund bought about a hundred and fifty thousand shares, a three percent weight, after the deal was announced, so the thematic bid did not wait for the ticker to change. The forty-five percent I credited to those seven desks is a floor and not a ceiling, because at least one of them holds its position as a swap that never shows up on an ownership filing, so the real arbitrage hand is bigger than the public register admits. And the passive bid that the whole scarcity case leans on is structurally late. The index rules only consider a de-SPAC after its stated lockup expires, and the stated lockup reads a hundred and eighty days even though the real one dies in about three weeks. So the insider supply arrives more than a year before the first index fund is allowed to buy a share. Demand and supply are mismatched on the calendar, and the mismatch runs against the stock.

The pattern trade, and its trapdoor

This is a Michael Klein SPAC, and his record is bimodal. Some deals popped and held, Oklo and Infleqtion among them. Others broke, MultiPlan on a short report, Skillsoft into a reverse split, Lucid over a longer horizon. So draw no single line through it.

The mechanical risk is a calendar, and it is far earlier than almost everyone assumes. I read the bylaw. The insider lockup does not run a clean 180 days. It releases the moment the stock trades above twelve dollars for fifteen days, with no seasoning period after the close, and at a tape already in the high teens that trigger completes in roughly the third week. Read that again. The supply cliff most people pencil in at six months, roughly three times the public float in rolled-over and sponsor stock, is really at about three weeks. The honest way to hold this is to be long into the catalysts and very awake into that early open.

The honest part, in full, and it got sharper

The flagship product has slipped about two years. The uncaged Digit that carries the entire order book was pitched to ship in the fall of 2024 and is now guided to broad availability in 2027, in the CEO's own words. That is the cleanest bear tell in the file.

There is a free way to check whether v5 is real hardware yet, and so far it says no. Any robot that carries radios has to clear the Federal Communications Commission before it can ship, and the only certification on record under Agility's name is for a v4 engineering unit. There is no v5 filing at all. Until one appears, the version that carries the entire order book has not passed the one gate every shipping product has to pass, and anyone can watch for it.

There is no audited revenue anywhere, because the S-4 is not filed. Every revenue-multiple argument you have read is built on estimates. The real number resolves only when the S-4 prints.

The factory is a nameplate, and the fleet behind it is smaller than anyone writes. I counted it site by site. Two Digits at the GXO flagship, and GXO's own automation chief is the one who said two, which means the famous hundred-thousand-tote headline is two robots working at roughly a third of a single shift. Three at Schaeffler. Seven at Toyota, which is the single largest deployment on earth. A pair at Amazon, a few more scattered. Add it up and the identifiable active fleet is around seventeen robots, call it low twenties once you count the ones sitting on the charger. That entire rental base throws off maybe a million and a half dollars a year against a hundred and eleven million in operating expense. The factory is rated for ten thousand a year and has been open since late 2023, and the company is adding six to ten net active robots a year. The product is real. The scale is not there, not yet.

The management bench is the risk I would weight most. This is a Magic Leap reunion at the top, and it is a trio, not a pair, the CEO, the chief business officer, and the general counsel all came from the company that raised billions and never scaled hardware. Look across the rest of the C-suite and the pedigree is a run of names that hit demand cliffs, Magic Leap, SunPower, Peloton, Fitbit. The finance seat should give you the most pause going into a listing. The person carrying the chief financial officer title has no public-company CFO experience, and her last operating job was chief operating officer of SunPower, which she left shortly before it went bankrupt. The company is still hiring its financial-planning staff in the middle of the deal and has no investor-relations hire at all. There is also a departure that matters more than any arrival. The executive who ran the Salem factory, the only person on the team who had stood up a first-of-its-kind production line before, left three months before the deal was announced, for a defense-drone company, and his replacement came from building products. The factory that anchors the entire scale story lost its most credentialed builder right before the public was asked to fund it. This bench is optimized for fundraising and narrative, not for capital-disciplined hardware. And the earliest backer, Playground Global, was co-founded by Andy Rubin, who is now building a competing humanoid company of his own.

Smaller honest notes. The robot is more human-in-the-loop than the pitch admits. On the deal webcast the chief executive said Digit is fully autonomous, with no teleoperation, and on the same recorded call the co-founder described the teleoperation data the company gathers to train it. Both statements are in the same filing, and the company is actively hiring teleoperators. The sharpest doubt comes from inside the house. Agility's own former chief product officer, asked whether a warehouse needs thousands of these machines, said in print that she does not think anyone has yet found an application that would require several thousand humanoids in a single building. That is the person who used to run the product, questioning the exact number the order book is built on. There are two OSHA worker-complaint inspections at the headquarters from 2023, not the clean sheet the pitch implies, though both closed without public citation. And organic sentiment is inverted, the euphoria is Stocktwits-only, while the real Reddit is skeptical.

What actually decides this, and how I am playing it

Strip it down. A real deployed product, a rented go-to-market that solves the hardest problem in the sector, a strategic consortium that de-risks supply and demand and compute at once, a three-front regulatory campaign built to wall out China and spare its own Taiwanese money, and the packaged flow already forming before the ticker trades. Against that, a flagship that slipped two years, no audited revenue until the S-4, a factory making dozens not thousands, a bench of executives whose prior companies mostly failed, and a supply calendar that punishes the impatient.

The single event that resolves most of it is the S-4. It prints the first audited revenue, the identity of the warrant customer, the Oregon State license terms, the real share count, and the post-close board. Read it the day it drops. Until then, the unpublished edges are the reason to own it, the honest bears are the reason to size it carefully, and the roughly three-week lockup cliff is the reason to respect the calendar. I would rather be early and small on something this asymmetric than wait for the S-4 to make it obvious and pay up.

I told you this would be the deepest file on this name anywhere, and I went back and made it deeper. Everything above is from public records, none of it is priced in, and most of it is not written anywhere else yet. When the S-4 makes it obvious, it will not be a secret anymore.

Disclosure. This is research synthesis built entirely from public filings, public databases, and public web sources, for educational purposes. It is not investment advice, and I am not a registered investment advisor. Positions and scenarios discussed are illustrative, not forecasts, and the inputs can be wrong. The company has not yet filed its S-4, so several figures here are drawn from secondary sources or are estimates and are labeled as such in the text. Do your own work.

Read more

Written 2 July 2026. As I write this, Churchill Capital Corp XI trades near $17.31 on heavy volume, against a 52-week range of $10.07 to $19.62. Eight days ago, on 24 June, that ticker stopped being a blind pool of cash and became the only way a public-market investor can buy a leading humanoid-robotics company before it lists. Churchill XI signed a definitive agreement to merge with Agility Robotics, the maker of Digit, at a $2.5 billion pre-money equity value. The combined company will trade as AGLT. Foxconn led the roughly $200 million financing that comes in alongside the deal. Days after the announcement, BlueCrest Capital Management, the hedge fund built by Michael Platt, crossed above 5 percent of the stock. This letter is about why a robot that already moves totes in a Georgia warehouse and loads parts in a South Carolina factory is worth your attention at $17, and it is equally about the risk you take to own it there.

Let's go.

The Klein Machine Points at Robots

Start with who built the vehicle, because the sponsor is not incidental. Churchill Capital is Michael Klein's SPAC franchise. Klein is the former Citigroup banker who has raised and merged a long series of Churchill vehicles, and the one that matters for context is Churchill Capital Corp IV. CCIV was the shell that took Lucid Motors public in 2021. That deal became the reference point for the entire SPAC era, both for the size of the run into the merger and for the volatility that followed it. The lineage tells you two things at once. Klein can source large, name-brand technology targets, and a Klein deal can move violently on sentiment. Both are true here.

The CCIV precedent is worth holding in mind for a specific reason. When Klein's fourth vehicle announced Lucid, the stock had already run to multiples of the $10 trust on speculation, and the people who bought the fundamentals early were rewarded before the deal even closed. The lesson is not that every Klein deal repeats that path. The lesson is that a Klein vehicle with a marquee technology target attracts real capital, real liquidity, and a real re-rate before the merger completes, and that the pre-close window is where the position gets established. CCXI is in that window now.

Churchill XI is a Cayman-domiciled blank-check company listed on Nasdaq. It carries three securities. CCXI is the Class A ordinary share. CCXIU is the unit. CCXIW is the warrant. The trust that backed the shares was funded at $10 per share, the standard SPAC construction, and it sits near $420 million assuming no shareholder redemptions. Every SPAC gives its holders a redemption right at the vote, the ability to hand back shares for the trust value in cash rather than roll into the merger. That right is why the trust acts as a floor for holders who bought near $10. It is also why, once a stock trades far above the trust, the character of the instrument changes completely, a point that governs the entire risk section below. Until 24 June, the honest description of CCXI was a pile of cash searching for a target. Now it is a signed transaction with a real operating company on the other side.

The structure is clean and easy to hold in your head. Agility is valued at $2.5 billion before the new money arrives. More than $620 million of gross cash is set to flow onto the balance sheet, built from the $420 million trust and roughly $200 million of new financing led by Foxconn. Both boards approved it unanimously. The deal is expected to close during 2026 after a shareholder vote, SEC review of the registration statement, and customary regulatory clearances. When it closes, the shell disappears and you own AGLT, a listed humanoid-robotics company with a working product and a Foxconn-sized bank account.

That is the vehicle. The reason to care about the vehicle is what is inside it.

Digit Already Clocks In

The single most important fact in this entire letter is that Digit is not a stage demo. It is a bipedal humanoid robot that already does paid work for named, blue-chip customers, and it has done so long enough to generate the kind of operating history no other humanoid company can match on the public tape.

Agility deployed Digit at GXO Logistics, the largest pure-play contract logistics provider in the world. The relationship started as a pilot and moved to a commercial Robots-as-a-Service agreement, the first of its kind for a humanoid. At GXO's facility in Flowery Branch, Georgia, Digit units moved more than 100,000 totes by late 2025. That is not a scripted thirty-second clip. That is a fleet of humanoids performing a repetitive material-handling task, shift after shift, measured in six figures of completed moves. When people argue about whether humanoids can do real warehouse labor, this is the counterexample that has already happened.

At Schaeffler, the German industrial and automotive supplier, Digit has been working inside a live factory since early 2025. At the Cheraw, South Carolina plant, Digit loads and unloads parts on a production line. Schaeffler did not stop at renting the robot. It made a minority equity investment in Agility and signed an agreement to purchase Digit units for deployment across its global plant network. A customer that writes an equity check and a multi-plant purchase order is telling you something a press-release pilot never does. It is telling you the robot earns its keep.

Beyond those two anchor accounts, Agility has disclosed deployments with Toyota Motor Manufacturing Canada and with Mercado Libre, the dominant e-commerce and logistics operator in Latin America. Automotive manufacturing and cross-border e-commerce fulfillment are two of the hardest, highest-volume physical environments in the economy. Digit is in both.

Sit with the 100,000-tote number, because it is doing more work than a casual reader gives it credit for. Moving totes in a warehouse is the archetypal task the logistics industry cannot hire enough humans to do. It is repetitive, it is physically taxing, turnover is brutal, and the labor is chronically short in exactly the markets where fulfillment volume is growing fastest. A humanoid that can walk to a shelf, pick a tote, carry it, and place it, over and over, without a facility being re-engineered around a fixed conveyor, is aimed at the largest and most persistent labor gap in the physical economy. The reason the tote count matters is that it converts a theoretical capability into a demonstrated one. Agility has not shown that a humanoid could theoretically do warehouse work. It has shown that a fleet of them did warehouse work, at a real customer, tens of thousands of times.

The commercial model matters as much as the customer logos. GXO is a Robots-as-a-Service arrangement, which means recurring revenue for uptime rather than a single hardware sale. That is the model that turns a robot company into a software-margin business over time, because the fleet keeps paying after it is installed. Agility is also technically anchored to the right partner. Digit's whole-body control runs on an NVIDIA foundation-model stack, and NVIDIA has featured the work as a case study in physical AI. When the company that supplies the compute layer for the entire AI build-out puts your robot in its own marketing, that is third-party validation you cannot buy with a slide.

Hold onto the distinction, because it is the crux of the bull case. Most of the humanoid field is selling a future. Agility is selling a present that already exists in four named accounts, backed by a compute partner that matters and a services model that compounds. When you buy CCXI, you are buying the most deployed humanoid platform that a public investor can currently touch.

The Largest Labor Market On Earth

Step back from the single company and look at the theme it sits inside, because the size of the prize is what justifies paying up for the leader.

Humanoid robotics is the physical expression of the AI wave. The last three years priced the digital side of artificial intelligence. The models, the chips, the memory, the data centers, the networking. The market paid enormous multiples for the layer that generates text and images and code. The physical side, robots that can perceive a messy real-world environment and act inside it, is the next leg, and it is aimed at something far larger than any software market. It is aimed at human labor.

The reason humanoids specifically, rather than purpose-built machines, is that the world is already built for the human form. Doorways, shelves, totes, staircases, vehicles, tools, and workstations are all sized and shaped for a person. A robot with roughly human proportions can slot into that existing infrastructure without a facility being rebuilt around it. A fixed automation cell requires the process to be redesigned around the machine. A humanoid adapts to the process that already exists. That single difference is why the form factor unlocks the enormous middle of the economy, the millions of manual tasks that were never worth the capital and the downtime of purpose-built automation. That is why the largest technology companies on the planet are pouring capital into the form factor at once. Tesla is building Optimus. The physical AI push runs through NVIDIA's entire robotics platform. The venture market has funded a cohort of well-capitalized challengers. The reason so much money is arriving at the same time is that the addressable market is not a category of software. It is a slice of global labor, and estimates from major Wall Street houses for the eventual humanoid opportunity run into the hundreds of billions and, over a long enough horizon, into the trillions of dollars.

The demographic backdrop is what makes the timing structural rather than speculative. The developed world is aging, the manufacturing and logistics workforce is shrinking, and the industries that most need physical labor are the ones that most struggle to staff it. Reshoring of manufacturing into higher-wage economies compounds the pressure, because the factories coming back online need workers those economies do not have in sufficient numbers. A machine that can do human-shaped physical work, at a cost that falls every year as the AI stack improves and manufacturing scales, is not a novelty in that world. It is the answer to a problem the labor market cannot solve on its own. The demand side of the humanoid thesis is not a bet on consumer adoption. It is a bet that the industrial economy will buy labor capacity wherever it can find it.

What changed to make this the moment, after decades of robotics promises, is the arrival of capable AI. The old barrier was never the hardware. Bipedal machines that walk and balance have existed for years. The barrier was the intelligence to perceive an unstructured environment and act in it without a human scripting every motion. Foundation models trained on vast data are what finally give a robot a general sense of how to grasp an unfamiliar object, recover from a stumble, and adapt to a task it was not explicitly programmed for. Digit running whole-body control on an NVIDIA foundation-model stack is the physical embodiment of that shift. The digital AI wave and the physical AI wave are the same wave, one rung apart, and the second rung is where Agility sits.

You do not have to believe the most aggressive of those forecasts to see the setup. You only have to believe that the leaders in a market this large will be worth many multiples of a $2.5 billion starting valuation, and that the company with the longest live operating history has an advantage that is hard to replicate. Deployment data is the moat that compounds. Every shift Digit works generates the operating experience, the failure modes, and the customer trust that a newer entrant has to earn from zero. Agility has a head start measured in tens of thousands of hours of real work.

The Comparison That Defines the Trade

Now the part that turns a good story into a mispricing. To value the leader, you do not need a discounted cash flow you can argue about for an hour. You need the comparison to what the market already pays for the same theme.

Figure AI is the most richly valued pure-play humanoid company in the world. In its Series C, which closed in September 2025, Figure raised more than a billion dollars at a post-money valuation of $39 billion, backed by a roster that includes NVIDIA, Intel Capital, Brookfield, Salesforce, Qualcomm Ventures, and others. Figure has real progress, with Figure 02 units delivered to customers including BMW and a third generation designed for mass manufacturing. It is a serious company. It is also priced at $39 billion in the private market while it is still early in commercial deployment.

Tesla's Optimus is the other reference point, and it is not separately priced at all. Optimus sits inside Tesla, a company valued above a trillion dollars, and any investor who wants exposure to Optimus has to buy the entire car and energy business around it. There is no clean, liquid way to own the humanoid alone.

Against those two, look at where Agility is being brought public.

Read the bars carefully, because they are the whole letter in one picture. The market has already decided that a leading humanoid platform is worth paying for. It paid $39 billion for Figure while Figure was still ramping. It embedded Optimus inside a trillion-dollar company. And it is bringing Agility, the humanoid with the longest live commercial operating history, public at $2.5 billion. Agility does not have to catch Figure's valuation to reward you. It has to close a fraction of that gap while carrying the strongest deployment record in the group.

There is a second signal buried in the same chart, and it is a quiet one. Agility's last private round valued the company near $2.1 billion. The de-SPAC brings it public at $2.5 billion pre-money. That is a modest step up, not the stretched mark-up that soured investors on so many 2021 SPACs. The sponsor did not slap a fantasy number on the target to juice the trust. The transaction is priced close to where the private market last cleared, which is exactly what you want to see when you are buying the leader at the start of a public re-rate rather than at the end of a promotional one.

This is the place to address the SPAC skepticism directly, because it is earned. The 2021 vintage produced a graveyard of pre-revenue companies taken public at valuations disconnected from anything real, with projections that never arrived and sponsors who cashed promote shares while retail holders absorbed the losses. That history is a reason for caution, not a reason to dismiss every de-SPAC reflexively. The tells that separated the bad deals from the survivable ones were consistent. A stretched valuation far above the last private round. No real product. No real customers. No strategic capital, only a sponsor and a hopeful PIPE. Measure this deal against that checklist. The valuation is a modest step up, not a moonshot. The product is deployed at four named blue-chip customers. The revenue is early but real. The lead PIPE investor is the largest electronics manufacturer on earth, not a passive fund reaching for a discount. This does not make the deal riskless. It makes it a different animal from the 2021 shells, and the market's willingness to pay 73 percent above trust reflects that the buyers on the tape have drawn the same distinction.

The deeper point about the comparison is that deployment history is a moat that compounds and cannot be bought with a funding round. Figure has more capital. Tesla has more capital. What Agility has that money cannot immediately replicate is tens of thousands of hours of Digit doing real, paid, repetitive work inside real customer facilities. Every one of those hours produced data on failure modes, edge cases, maintenance, and the unglamorous operational realities that only show up when a robot is on a real floor for months. A competitor with a bigger balance sheet still has to earn that operating history from zero, customer by customer, hour by hour. In a market where the buyer is an operations manager deciding whether to trust a machine with a production line, a track record is worth more than a spec sheet. Agility has the longest one in the field.

Foxconn Wrote the Check

A valuation is a claim. A committed check is a fact. The most important fact in the financing is who led the roughly $200 million that comes in alongside the trust. It is Foxconn.

Foxconn, formally Hon Hai Precision Industry, is the largest contract electronics manufacturer on earth. It is the company that builds the iPhone at scale. Its participation as the lead of the PIPE is not passive portfolio allocation. It is strategic capital from the one company in the world with the most credible claim to manufacture humanoids at volume and to deploy them across its own vast factory footprint. Foxconn is simultaneously a potential at-scale producer of Digit and a potential at-scale customer for Digit. When a manufacturer of that size anchors the money, it is underwriting both the product and the path to mass production.

That distinction is worth sitting with. Retail SPAC enthusiasm is cheap and it comes and goes. A strategic lead investor with manufacturing capacity and factory floors is the kind of validation that survives a bad tape. Foxconn is not betting on a quarter of momentum. It is betting on the industrialization of the humanoid form, and it chose Agility as the vehicle for that bet in the public market.

Manufacturing is the part of the humanoid race that gets the least attention and decides the most. A robot that works in a demo is an engineering achievement. A robot that can be built by the tens of thousands, reliably, at a falling unit cost, is a business. The gap between those two is where most hardware companies die, and it is precisely the gap Foxconn exists to close. Foxconn built the global supply chain that turned the smartphone from a marvel into a commodity produced at a billion units a year. If humanoids follow the same industrial curve, from artisanal to mass-produced, the company that knows how to run that curve is the most valuable partner a robot maker can have. Agility already operates its own robot-manufacturing capability, a facility it has described as purpose-built to produce Digit at scale. Pair that in-house capability with the world's foremost contract manufacturer as a strategic backer, and the production question, the one that quietly kills most hardware dreams, has a credible answer here that its rivals cannot easily match.

The proceeds structure has a feature worth understanding, because it protects part of the downside case even if the trust shrinks. The $200 million PIPE is a committed subscription at $10 per share. It is separate from the trust. If shareholders redeem heavily at the vote, the $420 million trust contribution falls, but the Foxconn-led PIPE money is a distinct commitment. The company can be recapitalized with strategic money even in a high-redemption scenario. That is a different quality of balance sheet than a SPAC relying entirely on a trust that can walk out the door.

Who Else Sees It

Foxconn is the strategic buyer of the story. The financial smart money is the other tell, and it showed up fast.

On a Schedule 13G with an event date of 26 June 2026, two days after the deal was announced, BlueCrest Capital Management disclosed a stake in Churchill XI. The filing reports 2,354,233 Class A shares, equal to 5.6 percent of the class, with sole voting and sole dispositive power. The reporting persons are BlueCrest Capital Management, the investment manager, and Michael Edward Platt, its principal. BlueCrest is one of the most respected macro and multi-strategy hedge funds in the world, and Platt is among the sharpest allocators in the business. When a fund of that caliber crosses 5 percent of a de-SPAC within days of the announcement, it is a signal that serious money did the work and decided the setup was worth a disclosed position.

Be precise about what the filing is and is not. A 13G is a passive, monitoring filing under Rule 13d-1(c). It is not an activist 13D, and it is not an endorsement of the stock or of this letter. It tells you that BlueCrest accumulated more than 5 percent and intends to hold passively. That is a data point about who is in the trade, not a promise about where it goes. Read it as smart money accumulating pre-close, and read the disclaimer at the bottom of this letter about what a 13G means.

The rest of the backer list is deep. Agility's private capital came from a group that reads like a map of who is serious about robotics. The $400 million Series C in 2025 was led by WP Global and drew in SoftBank, Amazon's Industrial Innovation Fund, DCVC, and Playground Global. Amazon's presence is its own kind of signal, given that logistics at Amazon scale is the ultimate proving ground and the ultimate market for warehouse humanoids. Add NVIDIA as the compute and foundation-model partner, and the picture is a company that the strategic and financial elite of the robotics world have already chosen to back.

Put the two groups together. A strategic lead in Foxconn, a financial validator in BlueCrest, and a private cap table anchored by SoftBank, Amazon, and NVIDIA. That is not a crowd that gathers around a science project. It is the crowd that gathers around the company it thinks wins the category.

The Company Behind the Ticker

It helps to know the operating shape of what you are buying, so the valuation has a real body under it rather than a logo.

Agility Robotics has raised roughly $640 million across its life, with the $400 million Series C in 2025 as the largest and most recent private round. The company employed around 406 people as of 31 May 2026. It designs, manufactures, and deploys Digit, and it operates its own robot-manufacturing capability, a facility the company has described as built to produce humanoids at scale. The installed base of Digit units is estimated near 75, which sounds small until you compare it to the rest of the field, where most competitors are counted in single-digit or low-double-digit real-world deployments. On deployed, revenue-generating units doing repetitive commercial work, Agility is at or near the front of the entire industry.

Agility is not a company that appeared to catch the humanoid wave. It grew out of a decade of legged-robotics research and spent years on the hard, unglamorous problem of building a bipedal machine that could walk, balance, carry, and work in the messy geometry of a real facility. That head start on the physical problem is why Digit was ready to do commercial work while much of the field was still perfecting a demo reel. The company reached the deployment stage first because it started on the hardest part first. When the AI layer arrived to give the hardware a brain, Agility already had the body and the customer relationships to put it to work.

The honest read on the financials is that this is an early-commercialization company. Revenue exists and it is real, because GXO and Schaeffler and the others are paying, but it is early and the unit economics at scale are still being proven. That is the correct frame for a humanoid company in 2026. You are not buying a mature earnings stream. You are buying the leader in deployment at the start of a market that the largest companies in the world believe will be enormous, at a valuation a fraction of the nearest pure-play peer, funded by strategic capital. The financials are early on purpose. The whole point of the transaction is to put more than $620 million behind the ramp. The registration statement, when it lands, will replace this estimate-driven picture with audited numbers and management's own projections, and it is the document that will let you underwrite the ramp with precision rather than inference.

Seventeen Dollars Is Not the Trust

Now the number that governs everything about how you own this, and the single most important risk in the letter. CCXI trades near $17.31. The trust behind it is $10 per share. The stock is roughly 73 percent above the cash in the trust. That fact defines both the opportunity and the danger, and any honest bull has to put it on the table before anything else.

Understand what this rules out. This is not a trust arbitrage. In a classic SPAC arb, you buy near $10, you have a redemption right that returns your $10 plus interest if you do not like the deal, and your downside is protected by the trust. That trade does not exist here anymore. The market has already re-rated Churchill XI on the strength of the Agility deal. At $17.31 you are paying $7.31 of premium over the trust for the humanoid fundamentals. You are underwriting the Agility thesis, not clipping a protected coupon.

That premium is the whole game, and it cuts hard in both directions. If the deal closes, that $17 is a claim on a public humanoid leader with a Foxconn balance sheet and, potentially, a re-rate toward the multiples the private market pays for the category. If the deal breaks, the reason to hold the stock at $17 evaporates, and the shares gravitate back toward the trust value near $10, plus whatever accrued interest sits in it. That is meaningful downside, on the order of 35 to 42 percent from here, and you must size the position for it.

For an investor who wants the upside with more torque and who accepts the risk of a total loss on the instrument, the warrants, CCXIW, are the leveraged expression. Warrants give geared exposure to a successful close and re-rate, and they can go to near zero if the deal breaks or the stock lingers below the exercise price. They are not for the faint of heart and they are not the core position. The core position is the common, and the common already carries the premium described above. Know which instrument you hold and why.

The reason to accept the premium is the asymmetry in the second and third rows of that table. The downside is bounded by the trust, near $10. The upside, if Agility closes and the public market decides a deployed humanoid leader deserves anything close to what the private market pays Figure, is not bounded by anything close by. You are risking roughly 40 percent to own the option on a category re-rate in the leader. That is the trade. State it plainly and size it honestly.

What Would Break This

You get the bear case in full, because a bull case is only worth reading if it has survived the argument against it. Here is what makes this wrong.

The first risk is the one just described, and it is the largest. This is not a trust arb. At $17.31 you are 73 percent above the $10 trust, and a broken deal takes you back toward $10. De-SPAC transactions fail. Shareholders redeem, financing falls through, regulators object, targets miss a closing condition, or sentiment turns and the sponsor pulls back. If any of that happens here, the premium is gone and you are left holding trust value. Every other risk in this section is smaller than this one. Do not own this in a size where a reversion to $10 hurts you more than you can accept.

The second risk is redemption mechanics. The $420 million trust figure assumes no redemptions. In practice, when a SPAC trades well above trust into a vote, arbitrage holders often redeem for the $10 rather than roll into the deal, because they can capture the trust and keep the warrants or simply recycle the cash. Heavy redemptions shrink the trust contribution and reduce the cash that lands on Agility's balance sheet. The Foxconn PIPE cushions this, because it is a separate committed check, but a high-redemption close still means less growth capital than the headline more than $620 million suggests. Watch the redemption number at the vote. It is the truest measure of how much conviction the existing holders have.

The third risk is that Agility is early. The deployments are real and the customers are named, but the revenue is early-stage and the unit economics at scale are not yet proven. A humanoid that works in four accounts is a long way from a humanoid that is profitable across thousands of units. The path from here to durable, high-margin, at-scale revenue runs through manufacturing yield, reliability, service cost, and customer expansion, and none of that is guaranteed. You are paying $2.5 billion for a company whose deployed revenue today is a small fraction of that number. The valuation is a bet on the ramp, not a multiple of current sales.

The fourth risk is the hype cycle. Humanoid robotics is one of the most exciting themes in technology, and exciting themes attract capital, promotion, and timeline optimism in equal measure. Timelines in robotics slip. The history of the field is full of demonstrations that were years away from commercial reality. Agility is further along than most, which is precisely the bull case, but the whole category is vulnerable to a sentiment reset if a high-profile peer stumbles or if the market decides physical AI is overpriced. A stock trading at a 73 percent premium to its trust is exposed to exactly that kind of reset.

The fifth risk is competition and capital. Figure raised more than a billion dollars at $39 billion. Tesla is pouring resources into Optimus. The field is deep and well-funded, and Agility, even after this deal, will be capitalized at a small fraction of its largest rivals. Being first in deployment is an advantage. It is not a guarantee that a better-funded competitor cannot out-engineer or out-scale you. The moat is real but it is not deep enough to be complacent about.

The sixth risk is dilution, lockups, and structure. A de-SPAC brings a PIPE priced at $10, sponsor promote shares, and warrants, all of which sit in the capital structure and can pressure the stock as lockups expire and holders manage positions. The share count that supports a $2.5 billion equity value is not the same as a clean, founder-and-public float. Understand that the post-close AGLT will have a cap table shaped by the SPAC mechanics, and that supply can weigh on the price in the months after close even if the fundamentals are intact.

None of these kills the thesis. All of them shape the size. The correct posture on CCXI is conviction in the direction and humility about the path, expressed as a position sized to survive a reversion to the trust.

The Path to the Bell

If you accept the trade, the next question is what to watch, because the value between here and close is created and destroyed at specific milestones.

The near-term catalyst that matters most is the registration statement. When the S-4 and proxy hit, the market gets its first full look at Agility's actual financials, its revenue base, its projections, and the precise deal mechanics. That filing can move the stock hard in either direction, because it replaces the current mosaic of press releases and estimates with audited numbers and management's own forecast. Read it the day it lands.

After that, the shareholder vote and the redemption election set the final shape of the balance sheet. A low redemption number is a vote of confidence and preserves the growth capital. A high redemption number is a warning even if the deal closes. Then comes the close itself, when the PIPE funds, the trust converts, and CCXI becomes AGLT. That conversion is the moment the vehicle stops being a SPAC and becomes a humanoid-robotics company on a major exchange, and it is the cleanest catalyst for the re-rate the bull case is built on. Beyond close, the story is carried by the fundamentals. New customer deployments, unit ramp, and the first public financials as AGLT are what turn a deal trade into a compounding investment.

The Bottom Line

Churchill Capital Corp XI is the only liquid, public way to own a leading humanoid-robotics company before it lists. Behind the ticker sits Agility Robotics, the maker of Digit, a bipedal robot that already moves more than 100,000 totes for GXO, works a live line at Schaeffler, and runs at Toyota and Mercado Libre, backed by an NVIDIA foundation-model stack and one of the largest deployed humanoid fleets in the world. The deal values Agility at $2.5 billion pre-money, a modest step up from its last private round and a fraction of the $39 billion the private market pays Figure. Foxconn, the largest electronics manufacturer on earth, is leading roughly $200 million of new money and validating both the product and the path to scale. BlueCrest and Michael Platt crossed 5 percent within days. The private cap table runs through SoftBank, Amazon, and NVIDIA. Humanoid robotics is one of the largest secular markets in technology, and the leader in real-world deployment is being brought public at a starting valuation that leaves enormous room to re-rate.

The honest catch is the price. At $17.31, CCXI trades 73 percent above its $10 trust, so this is not a protected trust arb, and a broken deal takes the stock back toward $10, a decline in the range of 35 to 42 percent. The revenue is early, the category carries hype and timeline risk, the redemption and dilution mechanics matter, and the competition is deep and better funded. Those are real, and they set the position size.

But understand the asymmetry, because it is the point. Your downside is bounded by a trust near $10. Your upside is the leader in the physical-AI wave, priced today at a fraction of its nearest peer, closing onto a public exchange with a Foxconn-sized balance sheet. To reach the upside in this letter, Agility does not have to become Figure. It has to close, deploy, and be worth a fraction more than the market pays for it today. That is what makes the asymmetry real rather than rhetorical.

The robot already clocks in. The question this letter puts in front of you is whether you own the company before the rest of the market can.

The full valuation build, the redemption scenario model, and the deployment-by-deployment case go out in Part 2. Subscribe so the next one lands in your inbox before the vote.

Disclosure. Long $CCXI. This is research synthesis for educational purposes, not investment advice. You should not buy or sell securities based on anything written here. I am not a registered investment advisor and I do not owe you a fiduciary duty. This is a pre-close de-SPAC, and the single largest risk is completion. The transaction requires a shareholder vote, SEC review, and regulatory clearances, and it can fail. At $17.31 the stock trades far above the $10 trust, so a broken deal, heavy redemptions, or a change in terms can send the shares back toward the trust value, a meaningful loss from current levels. The BlueCrest Capital Management and Michael Platt Schedule 13G is a passive, monitoring filing under Rule 13d-1(c). It is not an activist position, it is not an endorsement of this stock or this letter, and it can change at any time. The price scenarios above are illustrative, built on stated assumptions, and are not forecasts or guarantees. Valuations of private peers such as Figure Robotics are private marks and are not directly comparable to a public price. Figures are drawn from public filings, the deal announcement, and public market data as of 2 July 2026 and may contain errors. Warrants carry the risk of total loss. Position sizing in a pre-close de-SPAC trading well above its trust is your responsibility. Do your own due diligence.

Nothing here is investment advice. For educational purposes only. I may hold positions in names I discuss. Do your own research. No liability assumed.

Executive Summary

Since Part 11, two things moved in opposite directions, and the gap between them is the entire setup.

The price went down. On June 3, when Part 11 published, Merlin Labs (NASDAQ: MRLN) closed near $7.60. As I write at the end of June it trades near $5, a touch above its all-time low and roughly at the cash-floor bear case in my own model. Re-anchor to the live price when you read this.

The company went up. In the same window it completed the critical design review on the C-130J autonomy program, the single largest tripwire in this whole series, and it held. Congress kept legislating the demand. And on June 30 the Wall Street Journal reported the FAA is preparing a mandate that would require nearly all aircraft in civilian airspace to use ADS-B, the post-crash safety push turning into rulemaking.

So the business cleared its biggest gate and the addressable airspace started being upgraded into the exact data Merlin's autonomy consumes, while the stock printed its worst month. That is not a contradiction to resolve. It is the trade.

The architecture of this series has not changed. The trade was never the story, the contracts, or the technology. It is one variable, the execution-risk discount, and that variable compresses on milestones the company is already contracted to hit. Part 11 showed two conservative discounted-cash-flow models, mine and TD Cowen's, landing within pennies of an $11 floor, with Cowen's own undiscounted model implying roughly $27. The stock is now near $5. The discount did not narrow. It went extreme.

One thing I am updating in the open, because a paid reader deserves it and because the September and October mechanics are exactly what Part 11 said this Part would mark. The fully diluted share count is higher than the 135 million I used in Part 11, because the 12 percent convertible preferred is large and is likely to convert near a floor price this autumn. On the fully diluted, post-conversion count, my conservative floor is no longer $11.52. It is closer to $8 to $9.50, depending on how you treat the conversion. That is the honest number. The stock is near $5. The floor came down, the stock came down more, and the discount to a conservatively diluted floor is wider today than it was in Part 11, not narrower.

This is what this document covers. The CDR and what it removed. The demand becoming law. The patent moat the incumbents are citing. The four autonomy pools and the clean balance sheet. Then the part that matters most right now, the September lockup and the October preferred reset, modeled straight, with no flinching. Then the competitive clock, honestly. Then scenarios and the kill criteria.

Let's go.

I. The Setup Got Better And The Price Got Worse

Start with the price, because it is the objection.

Merlin came public in a de-SPAC that closed March 16, 2026 at a $10 reference. It ran to about $17, missed a quarter that was always going to be a near-zero revenue print, and bled to roughly $5. That is a 50 percent drawdown from the deal price and about 70 percent from the high. I am not going to pretend that is fun. It is the cost of owning a pre-revenue, cert-gated, supply-overhung small cap in a tape that has no patience for any of those three words.

But read what the drawdown is and is not. It is not a fundamental break. Across the entire public record there is no design-review failure, no contract cancellation, no FAA suspension, no management exodus that touched the certification bench, no runway-breaking cash event, and no adverse action on the preferred. Those are the seven things from Part 10 that would actually end the thesis. As of today, none has happened. What happened instead is a calendar problem meeting a supply problem. The revenue ramp is a 2027 event, the lockup is a September event, and the market is discounting both into a thin float right now.

The series mechanic still holds, and at a lower price it is simply more extreme. If the entire distance between the tape and fair value is one input, the execution-risk discount, then a lower tape on an unchanged business is a bigger discount, not a worse company. TD Cowen's published model is the cleanest illustration. They reached an $11 target two ways that agree, a 15 times calendar-2027 enterprise-value-to-sales multiple and a discounted cash flow of $10.94. And the $10.94 is not their fair value. It is their fair value after a 60 percent execution-risk haircut. Strip the haircut and the same model implies about $27, which is almost exactly where Roth's initial $25 target sat. A skeptical defense desk thinks Merlin is worth roughly $27 if it executes and roughly $11 once you discount for the risk that it does not. The stock trades near $5, below even the haircut number.

The discount is not a permanent property of the company. It is a function of uncertainty, and milestones remove uncertainty. A cleared design review removes it. A first flight removes it. A funded task order removes it. Each one gives a careful analyst a reason to take the haircut from 60 toward 50 toward 40, and the per-share number climbs toward the undiscounted case on every step. You are being paid to wait for a discount to melt while the company does what it is already under contract to do. The melt just got cheaper to buy.

II. The CDR Cleared, And The Contract Under It Is Sole-Source

On June 4, Merlin announced it completed the critical design review on the C-130J autonomy program for U.S. Special Operations Command. Preliminary design review was March 5. CDR is the gate where the design is judged mature enough to build and integrate. It is the milestone that, in Part 10, I named the single biggest near-term tripwire, because a failed CDR would have been a genuine kill signal. It cleared. The stock popped about 25 percent on the day and then gave it all back in the broader bleed, which is the discount mechanic working in reverse, the milestone landing and the market refusing to re-rate. That refusal is the opportunity.

Here is the part the deep work this month sharpened, and it is genuinely new to this series. The $105 million USSOCOM vehicle is not an ordinary contract. It is structured as an SBIR Phase III. That matters for two concrete reasons. Phase III is the one SBIR mechanism with no dollar ceiling and no competition requirement, reserved for work that derives from a prior Phase I or II. So it is sole-source justifiable, which means a competitor cannot simply underbid Merlin off this vehicle. And it legally requires a predicate, which Merlin has, an earlier AFWERX Phase II effort. The chain is AFWERX Phase II into a USSOCOM Phase III. That is a clean, defensible procurement posture, not a press release.

Be precise about the money, because the bull case is not served by inflating it. Of the $105 million ceiling, roughly $17.7 million is obligated to date, most of it on the C-130J task order. The money that funds it has come as congressional adds, $10.5 million in fiscal 2024 and a separate $12.5 million in fiscal 2026, the latter visible in SOCOM's own budget justification as funds being added to an existing contract with Merlin. Those adds are non-recurring. They have to be re-won each cycle, which is exactly why the lobbying spend pivoted to fiscal 2027 in the first quarter. So the bull statement is narrow and true. The flagship contract is a sole-source Phase III with about $89 million of unfunded headroom that SOCOM can draw against without a recompete, the program just cleared its design gate, and the customer briefs it as a real line of effort. The bear statement is also true. The dollars in hand are modest and the next tranche is an appropriations event, not a sure thing. Both are in the price.

And the customer confirms the program independent of Merlin's own marketing, which is the validation a skeptic should demand. The USSOCOM Program Executive Office for Fixed Wing publicly briefs MC-130J automation and reduced crew workload as a funded line of effort in its cleared conference materials, and it formalized a Mission Autonomy and Applied AI line this year. The slides do not name Merlin. The budget justification does. Put those two public documents together and you have a real customer requirement with Merlin's name on the funding line. That is more than most pre-revenue defense names can show.

III. The Demand Just Started Becoming Law

Now the ADS-B headline, read correctly, because the lazy version of this is wrong and the correct version is better.

The lazy version says a new ADS-B mandate means thousands of planes need upgrades, so Merlin wins. That is wrong. Merlin does not sell ADS-B hardware. The equipage dollars from any ADS-B rule go to the avionics box makers, Garmin and L3Harris and uAvionix and Honeywell. The market knew this instantly, which is why the original report carried Garmin and L3Harris cashtags and not Merlin's. ADS-B Out has also been mandatory in U.S. controlled airspace since January 1, 2020, so a post-crash rule is really about ADS-B In, receiving traffic, and about closing the transmit exemptions that let the January 2025 Washington midair happen with a helicopter that was effectively not squawking. This is an equipage story for the hardware vendors, full stop.

The correct version is the one that should interest a Merlin owner, and the patent work this month makes it concrete. A nationwide mandate that forces a denser, more complete ADS-B traffic picture is upgrading the airspace into the exact data Merlin's autonomy already eats. Merlin's detect-and-avoid intellectual property does not just consume one sensor. Its directed-perception patent application explicitly fuses ADS-B, air-traffic-control audio, and vision into a single cued detect-and-avoid pipeline, refining the picture with ADS-B and even generating a negative-contact report when it confirms no traffic is present. An autonomous aircraft is only as safe as the traffic picture it can build. A federal rule that makes that picture richer, more universal, and more reliable is a structural tailwind to every system that has to see and avoid without a human, and Merlin holds granted and pending IP precisely on fusing that richer picture into a flight decision.

So state it cleanly and do not oversell it. ADS-B equipage revenue is not a Merlin line item. The denser mandated traffic layer is free fuel for Merlin's autonomy, it strengthens the safety case the FAA will judge Merlin's certification against, and it puts a federal regulator on record that the post-crash answer is more automation and more shared situational awareness, not less. That is the rising tide. It is real, it is on-thesis, and it is not a contract.

The demand that is a contract is the federal budget. The fiscal 2027 request carries large autonomy lines, and Congress has been writing reduced-crew and autonomous-systems language into draft defense bills. The honest caveat is that most of the biggest autonomy dollars, the Collaborative Combat Aircraft money and the missile-defense pools, are aimed at uncrewed fighters and kill chains, not crewed transports, and Merlin has to compete inside the narrow lane that fits its product. But the lane exists, it is funded, and the lobbying machine pushing the fiscal 2027 add is a sophisticated one. Watch whether the C-130J add re-ups. That single line is the cleanest revenue-reality test in the whole thesis.

IV. The Moat Is The IP The Incumbents Are Citing

The strongest single fact this month did not come from a press release. It came from the patent office.

Merlin holds fifteen granted U.S. patents. The core of the moat is the air-traffic-control audio family, seven granted patents plus two fresh continuations filed late in 2024, on parsing the controller's spoken instruction and obeying it, hearing the tower and flying the clearance. The root patent on this is forward-cited by the people who would know whether it matters, and the list is the tell. Honeywell is the single heaviest citer of the ATC-audio IP, and Honeywell is also a stated Merlin partner. Airbus, Boeing, Collins Aerospace, and Thales all cite the family as well. A public avionics peer, Innovative Solutions and Support, cites Merlin's dual-assurance certifiable-compute patent. When the legacy primes who dominate the cockpit are citing your patents in their own filings, you own something real and specific in the one place autonomy has to live, the certified flight deck.

There is a second piece of IP that separates Merlin from its nearest competitor in one sentence, and the CEO put it on the record. Merlin's architecture is true onboard autonomy, with the decision-making software running on the airplane and requiring no link to the ground. That is the structural difference from Reliable Robotics, whose model keeps a remote pilot in the loop. A remote-pilot architecture inherits a datalink as a single point of failure and a regulatory and electronic-warfare vulnerability. An onboard architecture does not. For a defense customer worried about contested, comms-denied environments, and for a certification authority worried about link integrity, onboard autonomy is the harder problem and the more defensible answer. Merlin chose the harder problem, and the patent estate plus the dual-assurance, deterministic-core design is how it intends to make that choice certifiable in a way a black-box system cannot be.

I will mark the honest edges of the moat, because a real one survives them. The patents are commercially untested. There has been no challenge at the patent office and no litigation, so the moat has neither been validated by surviving an attack nor monetized by asserting it. And big technology companies cite the same root ATC-audio patent as generic natural-language prior art, which means the method is broad and the lane around pure airspace autonomy is the defensible part, not language parsing in general. With those caveats stated, the structural read stands. The incumbents who could build this instead chose, in Honeywell's case, to partner, and they are citing Merlin's filings. That is the moat.

V. Four Pools, One Brain, Priced For One

The framing that has anchored this series holds. Merlin is not a cargo company. It is one autonomy core, the Merlin Pilot, that sits inside four separate demand pools, and the market is paying for roughly one of them.

The first pool is military fixed-wing transport, the C-130J under the sole-source Phase III, with explicit scope to extend to other special-operations aircraft. This is the funded, customer-confirmed pool, and it is the one Wall Street's $11 target is mostly built on.

The second pool is the tanker, the KC-135 Center Console Refresh, where Merlin works through GE Aerospace. Be honest about this one. CCR is a real, budgeted Air Force program, with proposals due in mid-September, but its stated scope is a cockpit obsolescence refresh, fuel system and flight displays and the flight computer, not autonomy. Autonomy is the differentiator GE and Merlin are adding on top. And GE has to win it full and open against the entrenched incumbent, Collins, who already protested the prior version into a restart. The GE relationship is a credible door, not a signed distribution deal, and the right way to value it is as optionality on a 2027 award, not as revenue today. The tell to watch is a sub-award showing up under GE's name with Merlin inside it.

The third pool is commercial cargo, branded Condor, the largest pool by far over time and the one Wall Street excludes from its model entirely. This is the free optionality. It is also the longest-dated, because the commercial certification clock is the slow one and the regulator is deliberately cautious. The first revenue here is reduced-crew, a co-pilot's workload handled by software with a human still in the seat, not an empty cockpit. Read the New Zealand 2027 milestone the same way. It is achievable as a reduced-crew, safety-pilot first-of-type, which is genuinely on track, and it is not a pilotless revenue flight, because no jurisdiction has a settled rule for that yet. The optionality is real. The timeline is long. Both are true.

The fourth pool is international, anchored by the exclusive teaming agreement with the Remah International Group to scope autonomy opportunities in the United Arab Emirates. Treat this as a soft, framework-stage memorandum with a genuine, well-connected counterparty, not as booked revenue. It carries no value today and an export-control path that is multi-year. It is a call option on a Gulf market, nothing more, and I am not modeling a dollar of it.

One brain, four pools. The Street pays for the first, hopes for the second, excludes the third, and ignores the fourth. That is the structural mispricing the whole series has argued, and a $5 tape only widens it.

VI. The Balance Sheet Is Clean, And That Is Underrated

The most underrated fact about Merlin right now is that it is boring on the balance sheet, in the way a pre-revenue company almost never gets to be.

It is debt-free. At the merger close, the prior secured loan was repaid in cash from the trust, the older facility was cleared in 2025, and the New Zealand government loan was prepaid in February. There is no maturity wall and no covenant risk. Cash was about $122.8 million at the end of the first quarter, and the company added an $80 million PIPE on May 1, taking pro-forma liquidity to roughly $183 million. Against an operating cash burn near $90 million a year, that is a runway into about mid-2028, comfortably past the 2027 certification and revenue milestones that define the thesis.

The accounting under it is mostly conservative, which I checked line by line and which matters for a name this many investors assume must be playing games. Research and development is fully expensed, with zero capitalized software development, the cleanest possible treatment. Deferred revenue is zero, so there is no channel-stuffing or revenue-pull-forward. There is no related-party revenue. There has been no restatement and no non-reliance filing. The scary headline numbers are mostly optics. The $90 million first-quarter net loss is roughly 70 percent non-cash, driven by the fair-value remeasurement of instruments that have since converted to equity, and the real cash burn was about $23.6 million. The $641 million accumulated deficit is dominated by a one-time, non-cash deemed dividend from the pre-merger preferred exchange, a recap artifact, not cash that was incinerated.

There is one piece of conviction in the capital structure worth naming because it is non-conflicted. Baillie Gifford, an existing holder, did not just hold through the deal. It bought into the PIPE at $12 a share against a trust redemption value near $10.39, a deliberate premium. That is the cleanest single insider conviction signal in the file, a blue-chip long-duration manager paying up rather than redeeming. It does not make the stock go up. It tells you who was buying when it was easy to walk away.

I will state the offsetting accounting flag in the same breath, because a clean section that hides it is not clean. The first quarter printed a positive gross profit only because of a favorable release of a contract-loss reserve. Strip that release and the underlying gross margin is still negative. That is the one number an analyst must normalize out, and I am normalizing it out in front of you.

VII. The September Lockup And The October Preferred, Modeled Straight

This is the section Part 11 promised, and it is the one that matters most between now and the autumn. I am going to model the supply overhang without flinching, because it is the real cost of owning this today and because the discount only makes sense if you can see clearly what it is discounting.

Two mechanics dominate the next four months.

The first is the lockup. Roughly 180 days from the March 16 close lands around September 16. The shares that open are the rollover and sponsor blocks, and the honest count is larger than this series previously carried. Adding the holder I under-counted before, the founder, the sponsor, the venture leads, and the rollover blocks total roughly 56 million shares, against a public float that is a fraction of that. The most motivated seller is the sponsor, holding about 8.8 million shares at a basis near four-tenths of a cent. At any positive price that block is deeply in the money and structurally inclined to sell. The lockup has an early-release trapdoor on a subsequent change of control, but absent that, mid-September is the largest single supply event on the calendar.

The second is the preferred, and this is where I am correcting my own prior model in public. Merlin has about 21.7 million shares of a 12 percent cumulative convertible preferred outstanding, a stated value near $260 million, carried on the balance sheet around $180 million. It accrues at 12 percent paid-in-kind, which is roughly $32 million a year, an amount that is about equal to the entire 2026 revenue plan and several times the actual revenue run-rate. It is senior to the common. And its conversion price, currently $12, is subject to a one-time reset this autumn, on about the twenty-first trading day after the six-month anniversary of the close, so roughly mid-October. If the trailing twenty-day volume-weighted price is below the conversion price, it resets down to the greater of that price and a $5.00 floor. With the stock near $5, that reset is likely to trigger near the floor, and a preferred that converts at $5 instead of $12 converts into far more common, on the order of 50 to 75 million shares rather than 22 million.

Put those together and the fully diluted share count is not the 135 million I used in Part 11. It is closer to 185 to 210 million once the preferred resets and converts and the warrants are counted. That is why my conservative DCF floor is no longer $11.52. Spread the same equity value across the larger count and the floor is closer to $8 to $9.50, depending on whether you net out the preferred claim before or after conversion. I would rather hand you the lower, honest floor than defend the prettier one.

Now the part that keeps this in the bull column rather than out of it. First, the dilution is not new information being sprung on you. It is in the filings, it is the reason the stock trades where it does, and a $5 tape already embeds a market that assumes the reset triggers. You are not going to be surprised by it. The surprise risk runs the other way, if the stock is above $5 into October the reset is less punitive than feared. Second, the conversion is not pure destruction. When the preferred converts it stops being a $260 million senior claim that compounds at 12 percent ahead of you, and it stops draining $32 million a year of seniority. Swapping a senior, compounding claim for common is dilutive to the share count and accretive to the common's position in the stack at the same time. Third, the company is debt-free with a runway into 2028, so it is not a forced seller of equity to survive, which means the supply is mechanical, not desperate. And fourth, every one of these is a dated, known event. Mechanical supply on a known date into a thin float is the textbook setup for a price that is artificially depressed before the event and re-rates after it clears, the same way the lockup expiry that everyone fears so often marks the bottom rather than a new leg down.

So I will say it plainly. The single best reason the stock is at $5 and not $8 is the September and October supply, and the single best reason to be buying the discount is that the supply is known, dated, mechanical, and largely already in the price, in front of a 2027 ramp that is not.

VIII. The Competitive Clock, Honestly

A bull case that pretends there is no competition is a pump. There is competition, and the honest read of it is that Merlin is behind on one race and effectively alone on another.

Reliable Robotics is the credible rival and it is ahead on the nearest visible prize, the first paid uncrewed cargo flight. It raised about $160 million at roughly a $1 billion valuation, it holds the only city-led, uncrewed-cargo slot in the FAA pilot program, and it plans first paid uncrewed freight out of Albuquerque this summer, on the same Cessna Caravan airframe Merlin flies. Merlin is not in that pilot program, and its product is reduced-crew, not uncrewed, so it will not own the "first paid uncrewed" headline this year. That is a real competitive fact and I am not hiding it. The capital gap is real too. Shield AI is valued near $13 billion, Reliable near $1 billion, and the sponsor's next de-SPAC, Elroy Air, is coming at the same $800 million stamp Merlin carried. Better-funded peers are raising at scale while Merlin trades at half its deal price.

Here is why Merlin is not in those races and why that is fine for the thesis. No rival occupies Merlin's exact lane, which is crewed, large, multi-engine aircraft, the C-130J and the KC-135 and the commercial freighters, reduced first and autonomous later, using onboard autonomy and the ATC-audio moat. Reliable is a single-airframe uncrewed Caravan story. The eVTOL field is passenger air taxis. Shield AI is fighters and drones. Those are different, often earlier, often smaller prizes. Merlin's prize is later and larger, the multi-crew heavy-aircraft fleet where the labor savings per tail are the highest and where no one else is running the same play. The KC-135 CCR award is the concrete, funded entry point into that lane, and field-wide certification slippage, even the leader's uncrewed type certificate has slipped toward 2029, means Merlin's 2027 reduced-crew timeline is not uniquely behind. It is on a different, defensible clock.

IX. Scenarios, And The Kill Criteria That Still Stand

The scenario ladder that has anchored this series holds, re-anchored to a tape near $5 and a fully diluted count that now reflects the preferred. The new DCF work raised the floor's honesty, not the ceiling's height.

The honest conservative floor is the diluted DCF, roughly $8 to $9.50 once the preferred resets and converts, which is above the current tape. On top of that floor sits the probability-weighted catalyst optionality the series has carried, which takes the expected value into the mid-teens, and the bull and moonshot rungs are intact because nothing this month lowered the ceiling. Anyone reading this as a low-single-digit story is reading the stack upside down. The price is near the bear rung while the base, bull, and moonshot rungs are unchanged.

And the discipline that keeps this honest is the kill list. From Part 10, seven things would end this thesis. A failed design review. A contract cancellation. A contradiction of the customer-program claims. An FAA enforcement suspension. A management exodus from the certification bench. A runway-breaking cash event. An adverse action on the preferred. As of today, after the deepest audit I have done on this name, not one has occurred. The CDR cleared. The contracts stand. The customer briefs the program. There is no FAA action. The certification bench, the safety-assurance and engineering leadership, is intact even though the founding technical team has rotated out. The runway reaches 2028. And the preferred is converting on its own terms, not being weaponized against the common. The thesis is intact because the things that would break it have not happened, and I will tell you the day one of them does.

The Read

Strip it to the core.

A skeptical Wall Street desk and my own conservative model independently landed on an $11 floor in early June, with the desk's undiscounted math near $27, on a business that has since cleared its biggest design gate. The stock then fell to $5. The fully diluted share count is higher than I previously carried, so the honest floor is now $8 to $9.50, not $11.52, and the stock is still below it. Demand for the core capability is being written into draft federal law, and the FAA is moving to upgrade the entire civil airspace into the traffic picture Merlin's autonomy is patented to fuse. The whole distance between a $5 tape and a model that runs from $8 to $27 to $58 is one variable, the execution-risk discount, and that discount comes down on milestones the company is already contracted to hit.

The only honest costs are timing and supply. The revenue is a 2027 event. The September lockup and the October preferred reset are real, dated, mechanical, and largely already in a $5 price. I would rather buy a known overhang on a known date in front of an unknown ramp than the reverse.

The asymmetry is the spread between what is publicly known to be working, a sole-source Phase III, a cleared CDR, a cited patent moat, a debt-free balance sheet, and what is publicly known to be coming, the autumn mechanics, the fiscal 2027 add, the first flights, and a commercial half the Street values at zero. That spread is the widest it has been in this series, and it widened while the company got better, not worse.

I am long. This remains my largest position. Part 13 will mark the September lockup, the October preferred reset, and the second-quarter print as they resolve.

Shawarma Capital

Disclosure. MRLN is my largest position. This is research synthesis, not investment advice. You should not buy or sell securities based on anything I write. I am not a registered investment advisor and I do not owe you a fiduciary duty. Financial projections, including my discounted-cash-flow and catalyst model, are illustrative model outputs based on publicly available data and my own assumptions. They are labeled conservative and are not guarantees or company guidance. Third-party analyst figures are attributed to their published source and are summarized, not reproduced. Microcaps are illiquid and can lose 50% or more in a single session. Do your own due diligence.

Read more

Open the $CIFR feed on X right now and count the rocket emojis. I did. Bernstein just slapped a Buy on it and called the company a "power landlord of AI." A fund manager is bragging about being up seventy-four million dollars on the position. Tier lists rank it in the neocloud cohort. Somebody is posting a Gann angle that unlocks $64. The stock is up something like 600 percent in a year and it is sitting at an all-time high.

This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber.

This is the single most-hyped name on finance Twitter. It is a ten-billion-dollar company. It breaks two of the three rules I run this book on: it is not sub-$1 billion, and it is the opposite of undiscovered.

So why am I writing 4,000 words on it.

Because underneath the rocket emojis there is exactly one thing I actually hunt for a living. A physical chokepoint. Cipher is not an "AI stock." It is a landlord that spent four years quietly assembling energized, grid-interconnected power in West Texas, the single scarcest input in the entire artificial-intelligence buildout, and then signed that power away to Amazon and a Google-backed tenant for $11.4 billion of contracted lease revenue before it had built most of the buildings.

That part is real. The question this post answers is which part is real, which part is a press release, and whether the gap between the two is priced into a $26 stock at an all-time high.

Let's go.

I. The Two Rules This Breaks

I am going to be honest about why this name does not look like anything else I cover, because the honesty is the entire reason to read me instead of the emoji accounts.

Rule one I break: size. My edge is sub-$1 billion names that no sell-side desk covers. Cipher's market cap is $10.7 billion against 409 million shares outstanding at $26.22. There are now multiple analysts on it. Bernstein, the firm that wrote the "power landlords of AI" note, is not a microcap shop. This is a mid-cap with institutional sponsorship. The information asymmetry I usually trade is mostly gone.

Rule two I break: the crowd. I like names the tape has not found yet. The tape found this one. It ran from roughly $5 in the spring to a $30 print last week. Every momentum account on X is long. When I wrote about AmpliTech at $5.04 a month ago, the whole point was that nobody trusted the print. Here, everybody trusts the print. That is a different, and more dangerous, setup.

What I do not break is rule three, the one that matters most. The asset has to be a physical chokepoint with a multi-year catalyst window and a moat a competitor cannot simply copy with capital. On that test, Cipher passes harder than almost anything I own. That tension, a name that fails two of my filters and aces the third, is exactly why it is worth a Part 1. I am not here to tell you to buy a ten-bagger at the top. I am here to tell you what the asset actually is, so that when it pulls back, and a name that has run this hard always pulls back, you already know whether you want it.

II. What Cipher Actually Is Now

Cipher Mining was born as the US bitcoin-mining subsidiary of Bitfury Group. It went public in August 2021 through a merger with Good Works Acquisition Corp, a SPAC, at a roughly $2 billion valuation. Tyler Page has been CEO the entire time. For its first three years it was what every bitcoin miner is: a company that converts cheap electricity into hashrate, sells the bitcoin, and dilutes shareholders to fund the next fleet of machines.

In February 2026 the company changed its name from Cipher Mining to Cipher Digital. The rename is not cosmetic. It marks the pivot that is the whole thesis. Cipher is converting itself from a bitcoin miner that happens to own power into a data-center landlord that happens to still mine bitcoin.

The mechanism is simple to state and brutally hard to execute. Cipher spent years acquiring land and, more importantly, securing large-scale grid interconnection capacity in the ERCOT market in West Texas. Interconnection is the bottleneck. Anyone can buy land. Getting a utility to commit hundreds of megawatts, or a gigawatt, of deliverable grid power to a specific site, on a specific energization date, is a multi-year regulatory and engineering slog that money alone does not shortcut. Cipher did that slog while it looked like a sleepy miner. Now it is renting the result to hyperscalers who need power yesterday and cannot wait three years for their own interconnect queue.

That is the company. A power-interconnection portfolio with a bitcoin mine still running on part of it and AI data centers being built on the rest.

III. The Power

Everything in this thesis reduces to one number that is hard to fake: megawatts of energized, grid-connected power.

Cipher's stated portfolio is approximately 4.2 gigawatts of grid power across operating, contracted, and pipeline sites, with management guiding total capacity toward that 4.2 GW figure by 2030 and beyond. That number is the headline. It is also the number to be most careful with, because "pipeline" is doing a lot of work inside it.

Here is the honest breakdown of where the power actually is today.

Read that table twice. Of the 4.2 gigawatts everyone quotes, roughly 207 megawatts is energized and producing revenue right now, and it is producing that revenue from bitcoin, not AI. The 600 megawatts under contract at Barber Lake and Black Pearl is real, financed, and being built, but it is not yet generating a dollar of rent. Everything past Stingray is pipeline: land and interconnection rights on a 2028-2029 clock, dependent on the same grid queue, supply chain, and capital markets that could slip a year in either direction.

The power is the moat. The energization date is the risk. Hold both of those in your head for the rest of this post.

IV. The Leases

This is where Cipher separated itself from the bitcoin-miner pack. It signed long-dated, large-dollar colocation leases with credible AI tenants and got paid to build.

The Fluidstack deal, announced September 2025 and expanded in November, fills the entire 300 MW Barber Lake site under a 10-year hosting agreement. Total contracted revenue across the Fluidstack relationship runs to roughly $3.8 billion over the initial term, with extension options that management frames as up to $7 billion to $9 billion if exercised. The critical feature: Google is a strategic backer. Google agreed to backstop a portion of Fluidstack's lease obligations, $1.4 billion initially plus an additional $333 million on the expansion, and in exchange received warrants for roughly 24 million Cipher shares, about 5.4 percent pro forma. A trillion-dollar balance sheet standing behind part of the rent, and taking equity to do it, is the strongest validation a former bitcoin miner could buy.

The AWS deal, announced November 2025, is the anchor. Amazon Web Services signed a 15-year lease for 300 MW of gross capacity at the Black Pearl site in Wink, Texas, valued at approximately $5.5 billion. It phases in beginning July 2026 with rent effective August 2026, and Cipher provides both air and liquid cooling. AWS is investment grade. AWS is the tenant that makes the whole portfolio bankable.

A third campus lease with an "investment-grade hyperscale tenant," disclosed on the Q1 2026 call, plus the Stingray financing, brings management's total contracted revenue figure to approximately $11.4 billion. Bernstein, counting extension options and the broader pipeline, frames the order book at roughly $24 billion.

So the lease stack, stated plainly.

That is a genuinely impressive book for a company that was a sub-$1 billion miner two years ago. It is also where the skepticism has to start.

V. The Gap

Here is the number the rocket emojis never put next to the $11.4 billion.

Bernstein, the bull, models Cipher's AI revenue at $19 million in 2026, rising to $1.2 billion by 2030. Nineteen million dollars. This year. Against $11.4 billion of contracted lease value and a $10.7 billion market cap.

That is not a contradiction. It is the entire structure of the trade, and you have to understand it or you will misprice the stock in either direction.

The leases are signed. The rent has not started. AWS rent turns on in August 2026. The Fluidstack site at Barber Lake energizes through 2026 and into 2027. Until the buildings are energized and the tenants move load in, the $11.4 billion is a contract, not a cash flow. 2026 total company revenue is consensus-modeled around $228 million, and almost all of that is still bitcoin mining, the business Cipher is actively winding down. The AI revenue line is a rounding error this year on purpose, because the sites are under construction.

The whole company is a bet on closing that gap on time. If the sites energize on schedule and the tenants take the power, the revenue ramp from $19 million to north of a billion over four years is one of the steepest in the public market, at lease-economics margins Bernstein pegs near 93 percent EBITDA because Cipher is the landlord, not the GPU operator. If the sites slip, every quarter of delay is a quarter of that contracted book pushed right while the stock trades on the assumption it arrives on time.

The bulls are pricing the $11.4 billion. The business is delivering the $19 million. Part 1's job is to tell you that both numbers are true.

VI. The Capital Structure

The most underappreciated thing Cipher did is how it financed the build without torching shareholders.

Bitcoin miners dilute. That is the genre. Cipher itself went from a low-nine-figure share count at the SPAC to 409 million shares today, much of it raised by selling stock to buy machines. The pivot changed the funding model. Instead of issuing equity to build the AI data centers, Cipher financed them with non-recourse project debt secured against the contracted leases themselves.

Black Pearl Compute LLC priced $2.0 billion of 6.125 percent senior secured notes due 2031 in February 2026 to fund the AWS site, plus a $200 million revolving credit facility from a syndicate of global banks. The Barber Lake project carries roughly $1.7 billion of its own project debt. A separate ~$810 million senior secured note financing supports the Stingray/AWS capacity. Total principal long-term borrowings sat around $5.2 billion at the end of Q1 2026, the large majority of it non-recourse and ring-fenced at the project level, against roughly $4.25 billion of cash, equivalents, and restricted cash.

Why this matters: non-recourse project debt means the lender's claim is on the specific data center and its lease, not on the parent. If a project underperforms, the damage is contained to that subsidiary. The parent's equity is not on the hook for the full $5.2 billion. And because the build is debt-funded against investment-grade rent, Cipher is not issuing hundreds of millions of new shares to construct these sites. The dilution that defines the bitcoin-mining sector is, for the AI buildout specifically, largely absent. The Google warrants for ~24 million shares are the notable equity issuance, and Cipher got a Google backstop in exchange.

That is the "masterstroke" the more careful bulls are pointing at. A miner that learned to use its contracted cash flows as collateral instead of using its stock price as an ATM. It is genuinely well-engineered. It also means the company now carries real , and is fine until an energization date slips and a debt-service reserve gets tested.

VII. The Counterparty Problem

The bull case leans on the phrase "investment-grade hyperscaler tenants." Half of that is precise. Half of it is marketing. The distinction is worth real money.

AWS is investment grade. The $5.5 billion Black Pearl lease is as close to a sovereign-quality rent stream as exists in this sector. No argument.

Fluidstack is not investment grade. Fluidstack is a neocloud startup, founded by Gary Wu and Cesar Maklary, that hit roughly $180 million of annualized recurring revenue in late 2024 and is reportedly raising around $700 million at a $7 billion valuation, with Situational Awareness, Leopold Aschenbrenner's firm, in the lead. It manages over 100,000 GPUs and serves AI labs like Meta, Mistral, and Character.AI. It is a real, fast-growing company. It is also a venture-backed business assembling one of the most complex financing structures in the sector, a Macquarie facility sized up to $10 billion, a few hundred million of equity, and Google backstops layered across multiple landlords, Cipher and Hut 8 and TeraWulf among them.

The Google backstop is what makes the Fluidstack rent bankable, and the backstop is partial, not total. Google guarantees a defined slice of the obligation to support the project debt. It does not guarantee the full 10-year, multi-billion-dollar lease. If the AI capex cycle cools, the entity standing behind most of Cipher's Barber Lake rent is a startup whose own survival depends on the same cycle, with a trillion-dollar partner backstopping part, but only part, of the bill.

This is not a reason to dismiss the name. It is a reason to weight the $11.4 billion correctly. The AWS portion is gold-plated. The Fluidstack portion is a high-growth startup wearing a Google-shaped life jacket that covers the torso but not the legs. Treat them differently when you price the book.

VIII. The Bitcoin Engine Being Dismantled

The part of Cipher that actually makes money today is the part it is taking apart.

Odessa runs 207 megawatts of bitcoin mining at roughly 11.6 exahash per second and 17.2 joules per terahash of fleet efficiency. In Q1 2026 it mined 346 bitcoin. The company also holds a treasury of about 1,808 bitcoin and 3,952 ether as of the end of Q1. Q1 revenue was $35 million, down from $60 million the prior quarter, precisely because Cipher is winding down mining at Black Pearl to hand that power to AWS. The net loss was $114.3 million and adjusted EBITDA was negative $48.2 million, both reflecting a company spending heavily to build the AI sites while its only live cash engine shrinks.

That is the awkward middle. For the next several quarters, Cipher is a company whose reported financials get worse, not better, as the high-margin future revenue is still under construction and the legacy revenue is being deliberately switched off. Anyone buying the headline numbers in isolation is buying a company with a falling top line and a widening loss. Anyone buying the thesis is buying the moment the AWS and Fluidstack rent switches on and the income statement inverts. The earnings prints between here and mid-2027 will look ugly to a screen and fine to a model. That divergence is where the volatility lives.

IX. The Comp Set

Cipher does not trade in isolation. It trades inside the neocloud cohort, and the cohort is not one thing.

There are two distinct business models wearing the same "AI infrastructure" label. There are GPU operators, companies like CoreWeave, that buy the chips, take technology and obsolescence risk, and sell compute. And there are power landlords, companies like Cipher, TeraWulf, IREN, Hut 8, and Core Scientific, that own the power and the building and rent it to someone else who takes the chip risk. Bernstein initiated Cipher and TeraWulf together precisely because they are the same model: capital-light at the chip layer, capital-intense at the power layer, paid in long-dated contracted rent.

The landlord model is the safer half of the trade. The landlord does not own a single GPU that depreciates to zero in four years. The landlord owns interconnected power, which only gets scarcer. When you hear "neocloud," separate the ones that own silicon from the ones that own substations. Cipher owns substations. Within the cohort, that is the side I would rather be on, and it is the cleanest argument for the name.

The comparison that should keep a buyer honest is the rest of the cohort has run just as hard. IREN, WULF, CORZ, and the others have all repriced violently off the AI-landlord narrative. This is a crowded, correlated trade. On a sector drawdown, they sell off together regardless of which one signed the best lease.

X. The Bear Case

I give the bear case more room than the bull case in every Part 1, and this name has earned a long one.

One. Execution and energization risk is the whole ballgame. The $11.4 billion converts to cash only if Barber Lake, Black Pearl, and Stingray energize and accept load on schedule through 2026 and 2027. Grid interconnection dates slip. Transformer and switchgear lead times are measured in quarters. Liquid-cooling buildouts at this scale are new. Every quarter of slippage pushes the revenue ramp right against a stock priced for it to arrive on time. This is the single load-bearing risk and it is not in management's full control.

Two. Counterparty concentration. Strip out AWS and a large share of the contracted book sits behind Fluidstack, a venture-stage neocloud whose rent is only partially Google-backstopped. If the AI capex cycle cools or Fluidstack's own financing stack wobbles, the Barber Lake revenue is the first thing markets will discount.

Three. The reported financials get worse before they get better. Falling revenue, widening losses, negative EBITDA, and heavy capex for several more quarters. A macro risk-off that hits before the AWS rent turns on in August 2026 would hit a company that, on a screen, looks like a melting bitcoin miner.

Four. . $5.2 billion of project debt is non-recourse and well-structured, but it is still $5.2 billion against a company earning negative EBITDA today. Debt-service reserves are funded on the assumption the sites energize and pay. A delayed energization tests those reserves. Non-recourse limits the blast radius. It does not make the free.

Five. Valuation and crowding. The stock is at an all-time high, up roughly 600 percent in a year, the most-posted ticker in its cohort, with a fund manager publicly bragging about the gain. Bernstein's own $32 target implies only about 22 percent upside from here, and that is the Street's bull. When the marginal buyer is a momentum account and the marginal analyst target is 22 percent away, the asymmetry is no longer in your favor at this entry. The thing trades like everyone already owns it because, on FinTwit, everyone does.

Six. Bitcoin. Cipher still carries a bitcoin and ether treasury and still runs Odessa. A sharp crypto drawdown dents the one cash engine that is actually live and reminds the market of the genre Cipher is trying to leave.

Seven. Dilution history. This management team raised equity aggressively as a miner. The non-recourse pivot is a real change in behavior, but the track record says watch the share count, watch the ATM, and watch any convertible issuance the moment the project-debt market tightens.

None of these are theses-killers on their own. Together they say the same thing: this is a well-built asset at a fully-priced, fully-crowded entry, with the one risk that matters, energization timing, sitting outside management's hands.

XI. Scenario Math

I do not hand-tune probability weights for a Part 1. I give you the frame and let you weight it.

The market cap is $10.7 billion. The contracted book is $11.4 billion of gross lease revenue over 10-to-15-year terms. At Bernstein's ~93 percent landlord EBITDA margins, the steady-state economics, once fully energized, are a multi-hundred-million-dollar annual EBITDA stream from the first three campuses alone, with the 2.5-plus gigawatts of pipeline as free options on top. On 2030 AI revenue of $1.2 billion at landlord margins, today's $10.7 billion cap is not obviously expensive. That is the bull's math and it is not crazy.

The bear's math is simpler. You are paying $10.7 billion today for $19 million of 2026 AI revenue and a promise that the rest arrives on a construction schedule. The Street's most bullish target, Bernstein's $32, is 22 percent above the print. The average analyst target sits around $32 to $33. You are buying near the analyst ceiling, not the floor.

The honest synthesis: the long-term value is plausibly much higher than $26 if execution lands, and the next 12 months of entry-price asymmetry are poor because the stock has already priced a clean energization. The way you make money on this name is not by chasing the all-time high. It is by owning the asset through a cohort drawdown that resets the entry while leaving the leases intact. The leases do not care what the stock did last week.

XII. The Catalyst Calendar

The dates that move this name from here.

• Q2 2026 earnings, August 6, 2026. The first print that should show AWS rent beginning to flow and the first read on whether Black Pearl phase one energized on time.

• AWS rent commencement, August 2026. The single most important operational milestone. Rent turning on at Black Pearl is the moment the income statement starts to invert.

• Stingray energization, Q2 2026. The 100 MW front-of-the-meter site is the near-term proof that Cipher can bring contracted power online on schedule.

• Barber Lake energization through 2026 into 2027. The Fluidstack revenue ramp.

• Colchis and the 2028-2029 pipeline. Direct-connect with American Electric Power on the 1 GW JV, the option value on the back half of the 4.2 GW.

The pattern to watch is simple. Every energization that lands on time validates the model and de-risks the gap. Every one that slips a quarter does the opposite. The thesis is mostly a schedule.

XIII. Kill Criteria

What would make this a closed thesis rather than a watch item, if I owned it.

One. A disclosed energization slip of two or more quarters on Black Pearl or Barber Lake. That converts a financing-and-construction story into a broken-schedule story.

Two. Any public sign of distress at Fluidstack, a down round, a financing failure, a lease renegotiation, or a withdrawal of any portion of the Google backstop.

Three. AWS pausing, descoping, or delaying the Black Pearl phase-in. The anchor tenant is the load-bearing wall.

Four. A new equity raise or convertible issuance above a few hundred million at a depressed price, signaling the non-recourse project-debt market has closed to them and the old dilution reflex is back.

Five. A debt-service reserve draw or covenant waiver on any project entity.

Six. A bitcoin drawdown deep enough to impair the treasury and force Odessa offline before the AI rent replaces the mining cash flow.

None of those are tripped today. The first testable one is the August 6 print and the AWS rent commencement in the same month.

XIV. Why I Have No Position

I will tell you plainly. I do not own this and I am not buying it here.

I missed it. It ran from $5 to $30 while I was deep in defense autonomy and compound semis, and I am not going to pretend a ten-bagger at an all-time high is a fresh idea. Chasing a 600 percent move into the analyst ceiling, in the most crowded name in its cohort, is the exact behavior I built my rules to prevent. It breaks my size mandate and it breaks my crowding mandate, and "but the asset is great" is what every top-tick buyer tells themselves.

What would change my mind is not a higher price. It is a lower one with the leases intact. If a cohort-wide drawdown resets CIFR back toward the high teens while AWS rent is turning on and Black Pearl is energizing on schedule, then the chokepoint, energized ERCOT power rented to investment-grade tenants on 10-to-15-year terms, is exactly the kind of asset I want to own for years. At that point the size rule still fails, but the asymmetry comes back, and I would size a position and say so in a Part 2.

Until then I am writing it, not buying it. That is the difference between research and a rocket emoji.

XV. Bottom Line

Cipher Digital is the best-engineered version of the bitcoin-miner-to-AI-landlord pivot in the public market. It spent four years securing the one input nobody can shortcut, energized grid power in West Texas, and it rented that power to Amazon and a Google-backed neocloud for $11.4 billion of contracted, mostly non-recourse-financed lease revenue without torching its share count to do it. The asset is real. The moat is power interconnection, and it is the hardest moat in this entire sector to copy.

The stock is a different question than the asset. It is a $10.7 billion company collecting $19 million of AI revenue in 2026, priced for a flawless energization that has not happened yet, trading at an all-time high inside the most crowded trade on finance Twitter, with its most bullish analyst target only 22 percent away. The contracted book is gold-plated where AWS sits and merely promising where Fluidstack sits. The financials get worse before they invert. The whole thesis is a construction schedule, and the schedule is the one thing management cannot fully control.

The bulls are pricing $11.4 billion. The business is delivering $19 million. Both numbers are true. The gap between them is closed by energization dates in the back half of 2026 and through 2027, and the first one prints on August 6.

I have no position. I am watching the schedule, not the chart. If the cohort breaks and the leases hold, I will have a lot more to say.

Shawarma Capital

I am long the names disclosed in my portfolio. I have no position in CIFR. This is research synthesis, not investment advice. You should not buy or sell securities based on anything I write. I am not a registered investment advisor. I do not owe you a fiduciary duty. My conclusions could be wrong in ways I have not anticipated. Financial projections are model outputs based on publicly available data. They are not guarantees. These names are volatile and can lose 50 percent or more in a single session. Do your own due diligence. Verify every number against primary sources.

Abstract. On June 18, 2026, the Federal Energy Regulatory Commission (FERC) acted on Docket RM26-4-000 to establish, for the first time, a standardized federal framework for interconnecting very large electricity consumers — data centers above all — to the interstate transmission system. The proceeding originated in an October 2025 directive from the Secretary of Energy under Section 403 of the Department of Energy Organization Act, and was paralleled by a December 2025 order (193 FERC ¶ 61,217) compelling the PJM Interconnection to create new tariff services for co-located load. Together these actions compress and standardize interconnection study procedures, formalize the co-location of load with on-site generation, and propose to assign the full cost of network upgrades to the connecting load. This paper situates the reform within the institutional architecture of U.S. transmission and tests a single proposition: that interconnection procedure is not the binding constraint on the rate at which firm, interconnected electrical load can be added, but rather a comparatively minor term in a system governed by physical manufacturing, transmission construction, generation adequacy, and capital-allocation constraints. Using PJM — the regional grid hosting the densest concentration of data-center load in North America — and drawing on interconnection-queue census data, three years of capacity-auction results, the North American Electric Reliability Corporation's reliability assessment, equipment lead-time surveys, and the FERC primary docket, we decompose the speed-to-power problem and quantify each constraint. We find that the procedural stage the reform targets occupies the shortest segment of a five-to-seven-year critical path; that the reform's most consequential effect is therefore distributional (through cost allocation) rather than temporal; and that the binding constraints — gas-turbine and transformer lead times measured in years, a single domestic source of grain-oriented electrical steel, and a capacity market that has now cleared below its reliability requirement for the first time in its history — lie almost entirely outside the reform's reach. The reform is necessary; it is not, by itself, sufficient.

Keywords: electricity markets; transmission interconnection; data centers; capacity markets; cost allocation; resource adequacy; PJM; FERC. JEL classification: L94, Q41, Q48, L51, K23.

1. Introduction

1.1 An electricity-demand shock without modern precedent

For roughly two decades, U.S. electricity consumption was essentially flat; utilities planned around demand growth of well under one percent per year. That regime has ended, and the proximate cause is computation. Lawrence Berkeley National Laboratory's congressionally-mandated assessment documents U.S. data-center electricity consumption rising from 58 terawatt-hours (TWh) in 2014 to 176 TWh in 2023 — 4.4% of national consumption — and projects 325 to 580 TWh by 2028, or 6.7% to 12% of all U.S. electricity (Shehabi et al., LBNL, 2024). The Electric Power Research Institute's 2026 scenarios place the 2030 figure higher still, at 9% to 17% (EPRI, 2026). On a global basis, the International Energy Agency estimates data-center consumption growing from about 415 TWh in 2024 to roughly 945 TWh by 2030 — a doubling — and notes that by 2035 U.S. data centers alone will consume more electricity than the production of aluminum, steel, cement, and chemicals combined (IEA, Energy and AI, 2025).

The defining feature of this load is not only its magnitude but its velocity. An artificial-intelligence training campus is conceived, financed, and brought toward commissioning on a horizon of eighteen to thirty months, while the transmission system that must serve it is planned, permitted, and built over five to ten years. The mismatch between the clock speed of compute and the clock speed of the grid is the organizing tension of contemporary U.S. electricity policy, and it is the problem to which Docket RM26-4-000 is addressed.

1.2 The policy response

On October 23, 2025, Secretary of Energy Chris Wright invoked Section 403 of the DOE Organization Act — a seldom-used authority permitting the Secretary to direct FERC to consider a rulemaking — to instruct the Commission to ensure "the timely and orderly interconnection of large loads to the transmission system," citing AI data centers as an "urgent issue." FERC transmitted the resulting Advance Notice of Proposed Rulemaking (ANOPR) on October 27, 2025 (Accession 20251027-3056), opening Docket RM26-4-000. More than 3,500 pages of comments followed. On April 16, 2026, FERC issued a procedural "Order Regarding Intent to Act," committing to a final action by the end of June 2026 in a manner "quick, efficient, and legally durable," and on June 18, 2026, the Commission took up the matter as Agenda Item E-1.

In parallel, on December 18, 2025, FERC issued a unanimous order — PJM Interconnection, L.L.C., 193 FERC ¶ 61,217 (2025) — finding PJM's tariff "unjust and unreasonable" for lacking clear terms for co-located load, and directing the creation of three new transmission-service constructs. The political framing of both actions is straightforward: regulatory friction is the obstacle to powering the AI build-out, and removing it will unlock supply.

1.3 Thesis and contributions

This paper subjects that framing to scrutiny. Our central proposition is that the procedural reforms, although sound and overdue, operate on the least binding constraint in the speed-to-power system. We make five contributions. First, we provide an integrated account of the reform and its institutional antecedents (Sections 2–3). Second, we develop a constraint-decomposition framework that locates interconnection procedure within a broader critical path (Section 4). Third, we assemble the empirical record — queue economics, demand projections, the PJM case study, and the equipment supply chain — and quantify each constraint (Sections 5–6, 9). Fourth, we analyze the distributional economics of participant funding, which we argue is the reform's most consequential feature (Section 8). Fifth, we develop scenarios and a research agenda (Sections 11–12). The analysis is positive rather than prescriptive, and deliberately avoids security-specific or investment recommendations.

1.4 Roadmap

Section 2 establishes the institutional background. Section 3 details the reform and its primary record. Section 4 develops the conceptual framework. Section 5 presents the queue evidence. Section 6 is the PJM case study. Section 7 addresses co-location and the Susquehanna precedent. Section 8 analyzes cost allocation. Section 9 examines generation adequacy and the supply chain. Section 10 synthesizes; Section 11 offers scenarios and sensitivity; Sections 12 and 13 address limitations and conclude. An appendix provides a glossary and methodology note.

2. Institutional Background: The Architecture of Interconnection

2.1 Generators, loads, and the pro forma agreement

Across roughly two-thirds of the United States, the bulk power system is administered by independent system operators and regional transmission organizations (ISOs/RTOs) operating under FERC-jurisdictional tariffs. Historically, interconnection reform has concentrated on generators: a power plant seeking to inject energy executes an interconnection service agreement (ISA) whose terms derive from a federal pro forma standard established in Order No. 2003 (2003). Agreements that deviate from that pro forma — "non-conforming" agreements — bear a high justificatory burden, requiring demonstration of reliability concerns, novel legal issues, or unique operational factors. This doctrinal point, seemingly technical, becomes decisive in the co-location analysis of Section 7.

2.2 Order No. 2023 and the generator queue

By the early 2020s, generator interconnection queues had become severely congested, dominated by speculative renewable and storage projects. FERC's Order No. 2023 (2023) responded by replacing the serial, "first-come, first-served" study process with a "first-ready, first-served" cluster-study approach, imposing commercial-readiness deposits and study deadlines to deter non-viable applications. Crucially for the present analysis, Order No. 2023 reformed the generator queue. It did not establish a comparable national framework for large loads, which had long been treated under the retail jurisdiction of states and the bespoke practices of individual utilities.

2.3 The procedural gap and the jurisdictional question

It is precisely this gap that RM26-4-000 targets. A hyperscale data center is, electrically, a large load rather than a generator; yet a 500 MW load can stress a transmission network as severely as a 500 MW plant. In the absence of a standardized federal large-load interconnection process, treatment varied by jurisdiction, producing uncertainty, litigation risk, and delay. Filling the gap, however, raises a contested legal question that runs through the entire proceeding: whether FERC possesses jurisdiction over load interconnection at all. The Department of Energy argues that the Federal Power Act, Section 201(b) — governing wholesale transmission access and practices "directly affecting" wholesale rates — supplies the authority; states and many utilities counter that load interconnection is a retail matter reserved to them (American Bar Association, 2026). As we discuss in Section 3, FERC's parallel PJM order suggests a strategy designed to sidestep, rather than resolve, this collision.

3. The Reform in Detail

3.1 Provenance and the six principles

RM26-4-000 originated not in an industry petition but in the Secretary of Energy's October 2025 directive. The ANOPR set out six reform principles (Mayer Brown, 2025; CSIS, 2025):

1. Standardized study deposits, readiness requirements, and withdrawal penalties, modeled on the generator-interconnection discipline of Order No. 2023, together with an "option to build" network upgrades;

2. A joint study process for hybrid facilities that co-locate load with generation;

3. System Support Resource / Reliability-Must-Run (SSR/RMR) studies when a large load co-locates with an existing generator, to capture the reliability consequences of diverting that generator's output;

4. An expedited interconnection study — potentially 60 days — for curtailable or dispatchable loads, an explicit reward for flexibility;

5. One hundred percent participant funding, under which the large load bears all assigned network-upgrade costs; and

6. Transition planning for projects already in study.

3.2 The 20 MW threshold and the federalism problem

The ANOPR provisionally defines a "large load" as one exceeding 20 MW, while expressly soliciting comment on whether that threshold is appropriate or should be eliminated. The figure is contested precisely because it is low: hyperscale facilities routinely exceed several hundred megawatts, so a 20 MW floor could "sweep in mid-sized manufacturers and facilities traditionally served under state tariffs, effectively federalizing retail matters" (CSIS, 2025). The comment record fractured along this line (Table 6): the Edison Electric Institute urged a floor of at least 100 MW; PJM's transmission owners argued for 200 MW; the R Street Institute opposed a size threshold entirely, proposing instead a test based on wholesale-market participation; and the Data Center Coalition supported the 20 MW figure and the expedited timelines.

3.3 Participant funding and cost causation

The reform's economic core is the 100% participant-funding model: the connecting large load pays the full cost of the network upgrades its interconnection triggers, on a cost-causation rationale. We analyze the welfare properties of this choice in Section 8; here we note only that it is a significant departure from the prevailing treatment of transmission as shared infrastructure recovered through regional rates — a structure under which, by one leading estimate, more than 90% of U.S. transmission investment is presently socialized rather than assigned to specific beneficiaries (Pfeifenberger / Brattle Group, 2024).

3.4 The PJM co-location order (193 FERC ¶ 61,217)

The December 2025 order is the reform's operational vanguard within PJM. Finding the PJM tariff "unjust and unreasonable" for lacking "clear rates, terms, and conditions of service" for co-located loads — a deficiency that produced "uncertainty" and "disparate treatment" — FERC directed PJM to create three new transmission-service constructs (Table 7) and to reform behind-the-meter generation (BTMG) rules it likewise found no longer just and reasonable, imposing a three-year transition, a materiality threshold, and grandfathering of existing contracts. Co-located loads must pay for Regulation and Black Start service on a gross demand basis, addressing the cost-shifting concern directly. Compliance filings were ordered for January 17 and February 16, 2026.

Chairman Laura Swett framed the order's animating purpose in consumer terms — "I don't want any American family to feel that anxiety because of anything that we do" — while Commissioner Judy Chang observed that "the broader challenge of reliably, efficiently, and fairly interconnecting large loads … remains before the Commission, the States, and industry." Commissioner Rosner's concurrence is analytically telling: it described the order as achieving "the ANOPR's goals without jurisdictional confrontation." This suggests that FERC's preferred strategy is to accomplish large-load reform through its unambiguous authority over wholesale transmission-service tariffs, rather than by asserting the legally novel and contested jurisdiction over retail load interconnection discussed in Section 2.3.

3.5 The June 18, 2026 action

At its June 18, 2026 open meeting, FERC acted on RM26-4-000 (Agenda Item E-1), advancing the large-load framework consistent with the ANOPR record — standardized study procedures, co-location pathways, an expedited-study incentive for curtailable load, and the participant-funding principle. Consistent with the Rosner concurrence, the instrument is best read as grounded in FERC's wholesale-service authority, preserving legal durability at the cost of comprehensive coverage of loads that never participate in wholesale markets. (The precise instrument — a notice of proposed rulemaking advancing toward a final rule, versus a direct final rule — and its operative provisions should be confirmed against the issued order; pre-meeting materials did not definitively resolve the instrument type.)

4. Conceptual Framework: Decomposing Speed-to-Power

4.1 The interconnection timeline as a critical path

We model the interval from a developer's decision to build to energization at scale as a critical path of partially overlapping stages. For a greenfield large load — or the generation that must accompany it — this interval typically spans five to seven years. Table 2 decomposes it and identifies which stages the reform reaches.

4.2 A taxonomy of constraints

It is useful to classify the constraints on speed-to-power into five categories. Procedural constraints are study queues and agreement execution. Physical constraints are the manufacturing and construction of transformers, turbines, and transmission lines. Supply-chain constraints are the upstream inputs — grain-oriented electrical steel, high-voltage bushings, skilled labor. Capital constraints are the financing and cost-allocation terms that determine whether a project proceeds. Reliability constraints are the engineering limits planners impose to keep the system stable. RM26-4-000 operates almost entirely on the procedural constraint, with a secondary effect on the capital constraint through cost allocation. It does not touch the physical or supply-chain constraints, and it bears on reliability only insofar as the expedited-study incentive presumes a flexibility that, as Section 7 shows, large loads have been reluctant to provide.

4.3 The marginal value of relaxing the procedural constraint

The policy-relevant quantity is the marginal reduction in total elapsed time from relaxing the procedural constraint while the others bind. If the procedural stage is the shortest element of the critical path — and Table 2 indicates it is — then compressing it yields a correspondingly small reduction in total time. The intuition is captured in the practitioner observation that the regulatory delay amounts to "a three-to-six-month edit on a five-to-seven-year wait": grid connections take years not because approvals take years, but because the manufacturing and construction of physical assets does. Figure 4 visualizes this asymmetry.

This is not an argument against the reform. Removing a discretionary, variable, and litigation-prone source of delay has genuine value: it improves cost discovery, reduces uncertainty, and materially accelerates the minority of projects for which the procedural stage is on the critical path — most importantly co-located projects that avoid network upgrades entirely (Section 7). The point is narrower and more precise: a reform confined to the procedural stage cannot, by construction, relax the physical and supply-chain constraints that dominate the remaining years of the path.

5. Queue Economics: The Evidence on "Shovel-Ready" Capacity

5.1 The scale of the queues

A recurring claim in public discourse — visible in the response to the very Bloomberg report that frames this paper — is that vast capacity sits "ready to connect." The empirical anchor is Lawrence Berkeley National Laboratory's Queued Up: 2025 Edition, which reports approximately 2,290 GW of generation and storage actively seeking interconnection at the end of 2024 — roughly 1,400 GW of generation plus 890 GW of storage across some 10,300 projects, nearly twice the entire installed U.S. generating fleet (Rand et al., LBNL, 2025). The peer-reviewed treatment in Joule reports a comparable figure of about 2,600 GW against an installed base of roughly 1,280 GW (Gorman et al., 2024). On its face, this supports the abundance narrative.

5.2 Attrition and the "ready to connect" fallacy

The interpretation collapses under the completion data. Of the capacity that requested interconnection between 2000 and 2019, only 13% had reached commercial operation by the end of 2024; approximately 77% had been withdrawn; and 10% remained active (LBNL, 2025). The Joule analysis, spanning more than 38,000 project-level observations, reports a withdrawal rate of roughly 78–80% (Gorman et al., 2024). Far from clearing, the funnel is lengthening: the median time from request to commercial operation for projects that energized in 2024 was 55 months (4.6 years), up from 22 months for the 2008 cohort and 36 months for the 2015 cohort — a near-tripling. And the attrition is not merely early-stage speculation: in 2024 a record 700 GW withdrew — exceeding the roughly 500 GW of new requests — with one-third of withdrawals occurring at the late-stage facility-study or interconnection-agreement phase, after the system had already incurred planning and study costs (LBNL, 2025).

A queue, properly understood, is not an inventory but a funnel with heavy attrition; a project with secured equipment, financing, transmission, and offtake is the exception. The headline gigawatt figure overstates deliverable near-term supply by nearly an order of magnitude once historical completion rates are applied.

5.3 The load–generation distinction

A further subtlety is decisive. The queue census describes generation. Data centers are load, and the large-load interconnection pathway is both procedurally distinct from and far less mature than the generator queue that Order No. 2023 reformed. Extrapolating generator-queue abundance to load-interconnection readiness is therefore a category error. In PJM specifically, the load side is already congested in its own right: the Independent Market Monitor reports more than 3,300 projects awaiting interconnection — the largest backlog of any RTO (Monitoring Analytics, 2025), and PJM's recent cycle drew applications representing roughly 220 GW, of which only about 21 GW had reached engineering/procurement and 8.2 GW were under construction as of January 2026.

6. The PJM Interconnection: A Case Study

6.1 Why PJM

PJM coordinates the bulk power system across thirteen states and the District of Columbia and administers the largest competitive wholesale electricity market in the world. It contains "Data Center Alley" in Northern Virginia, the single densest cluster of data-center load on the continent. PJM therefore offers the sharpest available test of how a grid absorbs concentrated, rapid large-load growth — and, as the data below show, the first system in which the strain has become acute enough to break a previously reliable market mechanism.

6.2 Demand: a forecast dominated by data centers

PJM's load forecast has been transformed in three years. The aggregate of utility ten-year demand forecasts for the region rose roughly six-fold, from 24 GW of projected summer growth in 2022 to 166 GW in 2025 (LBNL Large Load Literature Review, 2025). PJM's own 2025 Long-Term Load Forecast projects peak load growth of 32 GW between 2024 and 2030, of which data centers account for approximately 94%; over 2025–2040, data centers represent 422 TWh of a projected 605 TWh of annual load growth — about 70% (PJM, 2025). NERC's reliability assessment, applying a more conservative lens, still projects PJM summer peak rising by 56 GW and winter peak by 62 GW by the mid-2030s (NERC, 2026). The composition is the salient fact: large loads account for essentially all net peak growth, as conventional demand is flat.

It is worth noting that even these figures embed downward corrections. PJM tightened its data-center counting methodology to avoid double-counting speculative or duplicative requests — an admission that "announced" load materially overstates probable load, a theme we return to in Section 11.

6.3 The capacity market and the price signal

PJM secures resource adequacy through a forward capacity market, the Reliability Pricing Model (RPM), whose Base Residual Auction (BRA) procures committed capacity three years ahead. As surging demand projections met a tightening, retiring supply stack, the clearing price escalated across three consecutive auctions, as shown in Figure 1.

Three features of the 2027/28 auction (held December 17, 2025) deserve emphasis. First, the entire RTO — every locational deliverability area, and even the Fixed Resource Requirement areas — cleared at the price cap of $333.44/MW-day. Second, the auction procured 134,479 MW through the BRA plus 11,299 MW of FRR resources, for 145,777 MW combined, and still fell 6,623 MW short of the one-event-in-ten-years reliability requirement — the first such shortfall in PJM's history, leaving a reserve margin of 14.8% against a 20% target. Third, approximately 5,100 MW of the 5,250 MW year-over-year increase in the demand forecast — about 97% — was attributable to data-center load (PJM, 2025).

6.4 The price cap and the structural trap

The 2027/28 result reflects a price-collar settlement negotiated after Pennsylvania challenged the prior auction at FERC; the floor was set at $179.55/MW-day. A counterfactual simulation indicates the auction would otherwise have cleared at roughly $529.80/MW-day RTO-wide, and $542.83 in the Dominion zone — implying the cap suppressed the price by about 59% (Avalon Energy Services, 2026). This produces a structural trap. The cap was designed as a consumer protection for a market with surplus capacity; it now suppresses the very investment signal needed to close a 6,623 MW shortfall. The Independent Market Monitor has accordingly declared the capacity auctions for the 2025/26, 2026/27, and 2027/28 delivery years "not competitive," "in significant part as a result of forecast demand for data centers" (Monitoring Analytics, 2026). The market is simultaneously pricing scarcity at its ceiling and prevented from pricing it high enough to resolve it.

6.5 Consumer incidence and the wider market

The escalation reaches consumers through both capacity and energy channels. The IMM's 2025 State of the Market report records the full-year real-time locational marginal price rising 50.4%, from $33.74 to $50.73/MWh, and the total cost of wholesale power rising 48.9%, from $55.52 to $82.67/MWh, with the capacity-cost component alone up 262.3% year-over-year (Monitoring Analytics, 2026). Real-time load rose 3.7% in 2025, and PJM set new winter and summer peak records in the same year. These are not abstractions: as Section 8 details, the increases have already produced double-digit monthly bill increases across PJM states and have forced administrative intervention in a market mechanism.

6.6 Reliability and the supply side

NERC's 2026 Long-Term Reliability Assessment classifies PJM at "elevated risk," with the anticipated reserve margin falling below the reference margin level beginning in 2029; the 2026 summer anticipated reserve margin deteriorated from 35.7% in the prior assessment to 29.7% in a single year (NERC, 2026). NERC also revised its ten-year North American summer peak-demand growth forecast upward by 69%, from +132 GW to +224 GW. Notably, the greatest risk of unserved energy in PJM falls in winter, under below-normal temperatures with natural-gas infrastructure constraints — a reminder that resource adequacy is a seasonal and fuel-security problem, not merely a nameplate-capacity problem. An adversarial review argues NERC overstates the risk by excluding likely-to-connect queue resources, and that counting them would resolve most identified shortfalls (Grid Strategies, 2026) — a useful caution, but one that itself depends on the queue-completion assumptions Section 5 shows to be historically generous.

7. Co-location: The Escape Valve and Its Precedent

7.1 The Susquehanna proceeding

The cleanest way to avoid a congested transmission network is not to use it — the logic of co-location, in which a large load is sited at a generator and served directly. Its defining test was the arrangement between Talen Energy and Amazon at the 2.5 GW Susquehanna nuclear plant in Pennsylvania. In November 2024, FERC, by a 2–1 vote, rejected an amended interconnection service agreement that would have increased behind-the-meter supply to a co-located Amazon Web Services facility from 300 MW to 480 MW. The majority held that PJM had not met the Order No. 2003 burden of justifying a non-conforming agreement; the chairman dissented, characterizing the configuration as "first-of-a-kind" and warranting approval (Utility Dive, 2024). Talen sought review in the Third Circuit.

7.2 Net versus gross: the analytical crux

Co-location's central technical dispute concerns the basis on which a hybrid load-plus-generation facility is studied. Developers argue for a net basis — measuring grid impact by net injections and withdrawals — on the ground that on-site generation reduces system burden. Utilities and planners insist on a gross basis, arguing the grid must withstand contingencies when on-site generation fails, and that net accounting shifts reliability risk and cost onto other ratepayers (CSIS, 2025). The December 2025 order's requirement that co-located loads pay for Regulation and Black Start on a gross basis is a partial resolution in favor of the conservative view. The stakes are not trivial: an analysis of PJM territory finds that traditional firm interconnection of new large load imposes roughly $764 million per gigawatt in system supply costs, whereas a flexible "bring-your-own-capacity" approach — accepting curtailment of only 40–70 hours per year while maintaining grid availability above 99% — can reach full operation in about two years, three to five years faster than firm service (Camus Energy, via Institute for Progress, 2026).

7.3 The restructured template

Market behavior is instructive. In June 2025, Talen and Amazon restructured their arrangement into a long-dated power-purchase agreement — reported at roughly $18 billion over 17 years for up to 1,920 MW — delivered front-of-the-meter, through the grid rather than behind it. The restructuring preserved the economic substance while conforming to FERC's reliability and cost-allocation concerns, and the December 2025 order generalized the lesson by building standardized service options. RM26-4-000 carries that template to the national level. The episode illustrates both the appeal of co-location and the regulatory friction it must still navigate — friction that the reform is, genuinely, designed to reduce.

8. Cost Allocation: The Distributional Heart of the Reform

8.1 Cost causation versus beneficiary pays

Beneath the engineering lies a distributional question with a substantial economics literature. Two principles compete. Cost causation assigns upgrade costs to the party whose connection triggers them — the basis for participant funding, and the position defended on both legal and efficiency grounds by, among others, the R Street Institute (2024) and formalized in recent scholarship (Ribeiro et al., 2025). Beneficiary pays assigns costs in proportion to benefits received, recognizing that transmission, once built, lowers congestion and improves reliability for all users; it is the principle embodied in FERC's Order No. 1920 (2024), which requires that costs be allocated to beneficiaries "at least roughly commensurate with estimated benefits," subject to a benefit-to-cost threshold no greater than 1.25 to 1.

8.2 The efficiency case and its limits

Participant funding has a clean efficiency rationale: it confronts the connecting load with the social cost of its location decision, sharpens siting incentives toward unconstrained areas, and protects existing ratepayers from subsidizing private expansion. Yet even its proponents acknowledge the logic is incomplete. Transmission exhibits positive externalities and economies of scale; charging the first mover the full, lumpy cost of an upgrade that subsequently benefits many can deter socially valuable investment (CSIS, 2025). And the equity incidence is uneven: capital-intensive industrial loads — an electric-arc-furnace steel mill, for instance — "often cannot" absorb full network-upgrade costs in the way a hyperscaler can (CSIS, 2025), a concern the industrial commenters pressed directly (IECA/ELCON, 2025). FERC's Order No. 1920, with its beneficiary-pays orientation, sits in tension with a strict participant-pays model and frames the unresolved question.

8.3 Incidence: the cross-subsidy is already running

The most important empirical point is that the distributional stakes are not hypothetical. The cross-subsidy is already operating at multi-billion-dollar scale. PJM's Independent Market Monitor attributes 63% of the 2025/26 capacity-price increase — about $9.3 billion — to data-center load, and finds that across the last three base auctions data-center forecast load above existing levels accounted for $21.3 billion, or 45%, of the $47.2 billion in cleared capacity costs (Monitoring Analytics, 2025–2026). Separately, the Union of Concerned Scientists documents $4.4 billion in 2024-approved local transmission projects serving data centers — more than 150 projects across seven states, with nearly $2.2 billion in Virginia alone — socialized across all ratepayers (UCS, 2025). Because more than 90% of U.S. transmission investment is currently categorized as "reliability-driven" and rolled into broad zonal rates (Brattle Group, 2024), the existing structure provides almost no mechanism to trace these costs back to the load that triggered them.

8.4 The household bill

The incidence reaches households directly. Synapse Energy Economics, in studies commissioned by state consumer advocates, documents capacity-driven monthly bill increases of approximately $21 in Washington, D.C., $18 in western Maryland, and $16 in Ohio, beginning in June 2025 (Synapse, 2025), and projects that PJM residential bills could rise nearly 60% by 2036–2040 absent interconnection reform (Synapse / Evergreen Action, 2025). The NRDC estimates cumulative data-center-driven capacity costs of $100–163 billion through 2033 — about $70 per month for the average PJM household — with Dominion-zone ratepayers facing capacity fees 65% above the PJM average (NRDC, 2025). A White House "ratepayer protection pledge" signed by hyperscalers in March 2026 carries no enforcement mechanism; absent a binding FERC rule, it is, as one analysis observes, legally unenforceable (Latitude Media, 2026). Cost allocation is therefore not a peripheral feature of the reform but its most consequential lever.

8.5 Compromise designs

A middle path is emerging in the comment record and at state commissions: large loads fund upgrades upfront but receive partial refunds or credits if their facilities later deliver system-wide benefits. Such designs attempt to preserve the incentive properties of participant funding while internalizing the positive externalities that pure cost causation ignores — a pragmatic reconciliation of the two principles that bracket the debate.

9. Generation Adequacy and the Equipment Supply Chain

If the binding constraint is physical, the supply chain is where the analysis must conclude. The evidence here is the strongest in the paper, and it is the reason the procedural reform cannot, on its own, move aggregate speed-to-power.

9.1 Gas turbines

The fast-ramping gas capacity utilities would normally deploy to serve new load is effectively unavailable on the relevant horizon. GE Vernova's gas-power backlog grew from 83 GW to 100 GW in the first quarter of 2026 alone, with management guiding to at least 110 GW by year-end and stating that reservation slots will be "sold out through 2030" by the end of 2026 (GE Vernova Q1 2026 Form 8-K). New orders placed today are not deliverable until late 2028 at the earliest; Mitsubishi Power quotes 2028–2030; and Siemens Energy carries a record backlog of roughly €131 billion (about $148 billion), with three manufacturers serving more than 75% of projects under construction (RMI, 2025). S&P Global reports U.S. turbine wait times of as much as seven years, and procurement costs rising sharply — Duke Indiana's Cayuga project cleared at $2,340/kW, 36% above the prior year, a roughly $900 million overrun on a 1,476 MW facility (S&P Global, 2025).

9.2 Transformers and grain-oriented electrical steel

Transformers are the second binding constraint. Large power transformer lead times average roughly 120–160 weeks (about 2.5–3 years), with specialized units reaching 210 weeks; the U.S. Department of Energy's 2024 Report to Congress cites commonly quoted lead times of 36 months and maxima of 60 (DOE, 2024). Generator step-up transformers average about 144 weeks. The IEA reports that procurement now takes up to four years — double the 2021 baseline — with transformer prices up roughly 75% since 2019 and DC-cable lead times beyond five years (IEA, 2025). Demand for high-voltage transformers has risen 116% since 2019, and for generator step-up units 274%. The upstream chokepoint is acute: transformer cores require grain-oriented electrical steel, of which the United States has a single domestic producer (Cleveland-Cliffs), so that more than 80% of new large power transformers are imported (DOE, 2024; NIAC, 2024). Within the transformer itself, high-voltage bushings and on-load tap changers — specialized, application-certified components served by few qualified suppliers — are the most consistent build-schedule bottlenecks (IEA, 2025).

9.3 The systemic consequence and the price feedback

These constraints aggregate into a build-out that lags its own demand signal. Industry analyses estimate that more than half of planned U.S. data centers may be delayed or cancelled for lack of electrical equipment — of roughly 12 GW expected to come online in the U.S. in 2026, only about one-third was under active construction, the remainder awaiting equipment with three-to-five-year delivery (trade synthesis, 2025–2026). Wood Mackenzie projects the U.S. data-center electrical-equipment market expanding from about $20 billion to $65 billion by 2030, a tripling that manufacturing capacity cannot match in time. The feedback into prices is direct: the Brattle Group's 2025 Cost of New Entry study for PJM found that the scarcity of turbines, transformers, and switchgear drove the cost of a new gas-fired plant 43–46% higher than in the prior study (Brattle, 2025) — which is precisely why the capacity market of Section 6 cleared where it did. The binding constraint, in short, is not capital: collective AI-infrastructure capital expenditure for 2026 exceeds $650 billion. It is energized megawatts.

9.4 New firm supply: nuclear, SMRs, and on-site gas

The fastest path to firm power has become bringing one's own. This takes three forms. Nuclear restarts and uprates — the Three Mile Island Unit 1 / Crane restart contracted to Microsoft; Palisades; Duane Arnold — add carbon-free firm capacity but on multi-year timelines. Small modular reactors remain years from first commercial megawatts. On-site gas, frequently aeroderivative units used as "bridge" generation, is presently the quickest deployable firm option and the principal beneficiary of the co-location reforms — though it, too, is gated by the turbine constraint of Section 9.1, and the IEA cautions that reliable on-site gas service requires overbuilding generation by 30–70% relative to demand, with an estimated 15–27 GW of on-site gas potentially serving U.S. data centers by 2030 (IEA, 2025).

10. Synthesis

10.1 For policy

The reform correctly identifies, and fills, a genuine institutional gap: the absence of a standardized large-load interconnection framework. But policymakers should hold realistic expectations about its effect on aggregate speed-to-power. Standardizing procedure will not, by itself, materially accelerate the rate at which firm load can be added while equipment and transmission constraints bind. The higher-leverage levers lie upstream — in transformer, turbine, and electrical-steel manufacturing capacity; in workforce; and in transmission permitting and construction — none of which is within the reform's reach. The reform is best understood as removing a real but secondary friction, and as setting the terms (through cost allocation) on which the binding constraints will be financed.

10.2 For reliability

By easing co-location and behind-the-meter pathways, the reform shifts part of the resource-adequacy question off the centrally-planned grid and onto private arrangements. This can accelerate individual projects, but it raises the net-versus-gross question of Section 7 at system scale, and it presumes a load flexibility that hyperscalers have been reluctant to provide in practice. The conservative, gross-basis treatment in the December 2025 order mitigates, but does not eliminate, the risk that planning models understate firm-capacity needs — a risk made concrete by the first RTO-wide capacity shortfall in PJM's history.

10.3 Economic incidence

The reform's most consequential effect is distributional. Participant funding would reallocate the cost of network upgrades from the general ratepayer base toward the connecting load — a shift with efficiency merits and real equity costs, and one whose final calibration (including any refund-for-shared-benefit mechanism) will determine the incidence of a sum already running into the tens of billions of dollars. Given that the cross-subsidy is already documented at $21.3 billion across three auctions and $4.4 billion in socialized transmission, the cost-allocation question — not the study-timeline question — is where the reform's stakes are highest.

11. Scenarios and Sensitivity

Base case. RM26-4 standardizes large-load interconnection and modestly compresses study timelines; co-location accelerates a meaningful minority of projects; equipment and transmission lead times keep aggregate speed-to-power gated through the late 2020s. PJM capacity prices remain at or near the cap; behind-the-meter generation proliferates; the cross-subsidy debate intensifies as bills rise.

Upside. Aggressive participant funding plus standardized co-location catalyzes faster private build-out of generation alongside load, easing the network-upgrade burden; manufacturing capacity for turbines and transformers expands on the strength of multi-year backlogs; the supply curve bends by approximately 2028–2029, and capacity prices ease as committed supply arrives.

Downside — and it is well-supported. A demand air-pocket. Goldman Sachs estimates that only about 60% of announced data-center capacity will materialize on schedule in the near term, falling toward 50% over two years, and frames the resulting stranded-asset risk explicitly (Goldman Sachs, 2025); the IEA independently flags that around 20% of planned projects are at risk of delay (IEA, 2025); and PJM's own methodology tightening concedes that announced load overstates probable load. If AI capital expenditure moderates, or if efficiency gains compress per-query energy intensity faster than base cases assume, the committed generation, equipment orders, and participant-funded upgrades would face a softer demand signal, concentrating stranded-cost risk. Rhodium's high-growth scenario, conversely, places data centers at 14% of U.S. power by 2030 and raises system costs 13–15% by 2035 (Rhodium, 2026) — the band of outcomes is wide in both directions.

What would change the conclusion. The thesis — that procedure is not the binding constraint — would weaken if (i) equipment lead times compressed sharply through manufacturing expansion; (ii) co-location and behind-the-meter generation scaled fast enough to bypass the transmission constraint for a large share of load; or (iii) the procedural stage proved, on better project-level data, to occupy a larger share of the critical path than Table 2 assumes. Each is a testable empirical question, and each is the right place to look for evidence against the argument advanced here.

12. Limitations and Future Research

The stage durations of Table 2 are representative ranges, not point estimates; they vary by region, voltage class, and equipment type, and a more rigorous treatment would estimate them econometrically from project-level data. The capacity-price counterfactuals depend on the price-collar settlement and on PJM's evolving load-counting methodology. The precise legal instrument issued on June 18, 2026 governs the reform's ultimate effect, and compliance litigation — together with the unresolved Section 201(b) jurisdictional question — will shape implementation for years. Promising avenues for further research include: project-level survival analysis of large-load interconnection requests; estimation of the cross-elasticity between capacity prices and behind-the-meter adoption; welfare modeling of alternative cost-allocation designs, including refund-for-benefit mechanisms; and empirical measurement of realized large-load flexibility against the assumptions embedded in expedited-study eligibility.

13. Conclusion

FERC's large-load interconnection reform is sound policy directed at a real institutional gap, and it arrives at a moment of genuine urgency: PJM's capacity market has cleared at its cap three years running and, for the first time in its history, has cleared below the reliability requirement. But the reform should be understood for what it is — the relaxation of the shortest constraint in the speed-to-power system. The binding constraints are physical and financial: gas-turbine slots sold out to 2030, transformer lead times of two-to-four years resting on a single domestic source of electrical steel, a capacity market whose price cap suppresses the very signal needed to close a 6,623 MW shortfall, and tens of billions of dollars in cross-subsidy whose allocation the reform must still decide. The procedural fast-track is a necessary permission slip. It is not, by itself, the power.

References

• American Bar Association, Infrastructure & Regulated Industries. (2026). The Jurisdictional Collision over Large Loads and Data Center Interconnection. Spring 2026.

• Avalon Energy Services. (2026). PJM 2027/2028 Base Residual Auction Clears at Price Cap.

• Baker Botts LLP. (2025). FERC Issues Order Providing Guidance for "Co-locating" Power Plants with Data Centers within PJM.

• Brattle Group / Pfeifenberger, J. (2024). Transmission Cost Allocation for Order 1920 Compliance. NARUC–NASEO–DOE Webinar, Dec. 6, 2024.

• Brattle Group. (2025). 2025 Cost of New Entry (CONE) Report for PJM. April 2025.

• Center for Strategic and International Studies (Yang, A., Cai, R., Majkut, J., Zacarias, M.). (2025). What's at Stake in FERC's Large Load Proposal? Dec. 9, 2025.

• Federal Energy Regulatory Commission. Interconnection of Large Loads to the Interstate Transmission System, Docket No. RM26-4-000 (ANOPR transmitted Oct. 27, 2025; Order Regarding Intent to Act, Apr. 16, 2026; action June 18, 2026). Order No. 2003 (2003); Order No. 2023 (2023); Order No. 1920 / 1920-A (2024).

• Federal Energy Regulatory Commission. PJM Interconnection, L.L.C., 193 FERC ¶ 61,217 (Dec. 18, 2025); Fact Sheet and Commissioner concurrences (Chang, Rosner).

• Goldman Sachs Research / GS SUSTAIN. (2025). Data Center Power Demand: The 6 Ps Driving Growth and Constraints; and US Data Center Power Demand Projected to Double by 2027.

• Gorman, W., Mulvaney Kemp, J., Rand, J., Seel, J., Wiser, R., et al. (2024). Grid connection barriers to renewable energy deployment in the United States. Joule 9(2). DOI: 10.1016/j.joule.2024.11.008.

• Grid Strategies LLC. (2026). Review of NERC's 2025 Long-Term Reliability Assessment.

• Institute for Progress. (2026). Fast and Secure Grid Interconnection for American AI Leadership (citing Camus Energy PJM study).

• International Energy Agency. (2025). Energy and AI (World Energy Outlook Special Report); Building the Future Transmission Grid. (2026). Electricity 2026.

• Lawrence Berkeley National Laboratory (Rand, J., et al.). (2025). Queued Up: 2025 Edition. OSTI:3008763.

• Lawrence Berkeley National Laboratory (Shehabi, A., et al.). (2024). 2024 United States Data Center Energy Usage Report.

• Lawrence Berkeley National Laboratory. (2025). Large Load Literature Review, December 2025 Update; and Electricity Rate Designs for Large Loads (Jan. 2025).

• Mayer Brown LLP. (2025). FERC Large-Load Interconnection Preliminary Rulemaking: Key Takeaways. McGuireWoods LLP. (2025). FERC Establishes Proceeding to Consider DOE Directive.

• Monitoring Analytics, LLC (PJM Independent Market Monitor). (2025–2026). State of the Market Report for PJM — 2025 (Mar. 12, 2026); Q3 2025 State of the Market; IMM Comment, Docket RM26-4-000 (Nov. 25, 2025); IMM Complaint re Data Center Loads.

• Morgan Stanley Research. (2025–2026). Powering AI: Markets Race to Invest in AI Energy Solutions.

• North American Electric Reliability Corporation. (2026). 2025 Long-Term Reliability Assessment.

• NRDC (Rutigliano, T.). (2025). PJM Board to Decide on Higher Bills for 67 Million Americans Due to Data Centers.

• NIAC / CISA. (2024). Addressing the Critical Shortage of Power Transformers. (Draft, June 2024.)

• PJM Interconnection. (2024–2025). Base Residual Auction Reports (2025/26, 2026/27, 2027/28); 2025 Long-Term Load Forecast; Regional Transmission Expansion Plan.

• R Street Institute (Chandler, K.). (2024). Cost-Causation Is No Cause for Concern; and Initial Comments, Docket RM26-4-000 (2025).

• Rhodium Group (Carter, T., King, B., Larsen, J., van Brummen, A., Kolus, H.). (2026). The Impacts of Rising Electricity Demand from Data Centers on US Energy and Emissions. Apr. 7, 2026.

• Ribeiro, L., Street, A., Arroyo, J. M., Moreno, R. (2025). A Causation-Based Framework for Pricing and Cost Allocation of Energy, Reserves, and Transmission. arXiv:2505.24159.

• RMI (Cohen, J., Fitch, T., Shwisberg, L.). (2025). Gas Turbine Supply Constraints Threaten Grid Reliability. June 18, 2025.

• S&P Global Commodity Insights. (2025). US Gas-Fired Turbine Wait Times as Much as Seven Years; Costs Up Sharply.

• Synapse Energy Economics. (2025). For DC Office of People's Counsel, Drivers of PJM's Capacity Market Price Surge; for MD Office of People's Counsel (2026); for Evergreen Action, Tackling the PJM Electricity Cost Crisis.

• Union of Concerned Scientists (Jacobs, M.). (2025). Connection Costs Loophole Costs Customers Over $4 Billion.

• U.S. Department of Energy. (2024). Large Power Transformer Resilience Report to Congress.

• Utility Dive; American Nuclear Society. (2024–2025). Talen–Amazon Susquehanna interconnection coverage. GE Vernova Inc., Form 8-K / Q1 2026 (SEC EDGAR, CIK 0001996810).

Appendix A. Glossary

ANOPR — Advance Notice of Proposed Rulemaking. BRA — Base Residual Auction (PJM's forward capacity auction). BTMG — behind-the-meter generation. CONE — Cost of New Entry. Co-location — siting a large load at a generator to be served directly. FRR — Fixed Resource Requirement (an alternative to RPM participation). GOES — grain-oriented electrical steel. ISA — Interconnection Service Agreement. LDA — Locational Deliverability Area. LMP — Locational Marginal Price. RML/ARM — Reference Margin Level / Anticipated Reserve Margin. RPM — Reliability Pricing Model (PJM's capacity market). RTO/ISO — Regional Transmission Organization / Independent System Operator. SSR/RMR — System Support Resource / Reliability-Must-Run. UCAP — Unforced Capacity.

Appendix B. Methodology note

Quantitative claims are drawn from primary sources (FERC dockets and orders, PJM auction and load reports, the PJM Independent Market Monitor, NERC, LBNL, DOE, and equipment-manufacturer disclosures), peer-reviewed literature (Joule; arXiv), and reputable institutional research (Brattle, IEA, EPRI, Goldman Sachs, Morgan Stanley, Rhodium, RMI, Synapse). Where figures vary across sources (for example, data-center demand projections), ranges are reported rather than point estimates. Critical-path stage durations (Table 2) are representative syntheses intended to convey relative magnitude, and should be read as ordinal rather than cardinal.

This paper is for informational and educational purposes only and does not constitute investment, legal, or engineering advice.

That is the trade in one paragraph. The market is still pricing a failed Bitcoin miner. The insiders are pricing the power. We are siding with the insiders, and we have started buying.

This is Part 1. It lays out what the company is, how it got here, what the asset is worth, and the one event that closes the gap. Parts to follow will track the catalysts in real time.

How we got here

You cannot understand the opportunity without the wreckage that created it.

The company used to be Mawson Infrastructure, ticker MIGI, and it was one of the worst-run names in the public Bitcoin-mining sector. While direct peers compounded more than 600 percent on the shift to AI infrastructure, Mawson destroyed roughly 97 percent of its value. Management churned. Creditors filed an involuntary bankruptcy petition. The stock was staring at a Nasdaq delisting.

Then an investor group called Endeavor started buying. Through late 2025 it accumulated stock and issued shareholder letters arguing the obvious: the infrastructure was worth far more than the equity, and the problem was leadership, not assets. By April 2026 the campaign was over. Endeavor had won board control, the prior CEO was gone, five new directors were seated, the company was renamed Big Digital Energy, and the company set itself on a path back into Nasdaq compliance. The board has since retired the takeover poison pill, now that the new owners are the ones in control. The new team pointed the company at the highest-value use of its core asset: low-cost, energized power for AI and high-performance compute.

The stock has already moved off the lows, from a roughly 16 million dollar market cap to roughly 40 million. We think that is the first leg of a much larger re-rate.

What you are actually buying

Strip away the ticker change and the narrative, and BGDE is a power-first infrastructure operator with a real, verifiable footprint. From the company's own SEC filings, not a pitch deck:

The flagship is Midland, one of the largest sites in PJM operated by a public miner, with roughly 38,810 machine positions and carbon-free power supplied under a fixed-price agreement with Energy Harbor, owner of the adjacent Beaver Valley nuclear station. Bellefonte is a smaller complementary site, and Corning is an additional development parcel that carries optionality the market ignores. A second 250-megawatt substation already sits on the site, currently offline, next to an option on 200 adjacent acres. That is development upside, not operating capacity, and none of it is counted in the 129-megawatt figure.

This is not a shell. The infrastructure is real enough that the company runs production data-center management tooling and live energy metering across its sites. You are buying a genuine operator, not a story.

The power economics

Here is the part the market is not modeling. BGDE buys power under a fixed-price contract at a blended cost in the low 40s of dollars per megawatt-hour, roughly 4 cents per kilowatt-hour, while the 2025 PJM all-in wholesale price ran closer to 79 dollars per megawatt-hour. That spread is the entire business model: control cheap, contracted power in a market where power is repricing violently upward.

It monetizes that position two ways today. First, hosting and colocation, renting capacity and racks to customers. Second, and more interesting, energy management: getting paid by the grid to curtail and provide demand response when PJM is short. That second line is small relative to what it becomes if the company converts capacity to AI, but it is real, it is growing, and it is the proof that scarce power is a cash-generating asset, not just a cost center.

The business already generates real revenue

The bear reflex on any ex-miner is "no revenue." That is wrong here. In 2025 the company produced roughly 38 to 40 million dollars of revenue across two real lines: digital colocation at about 26 million, and energy management at 11.8 million, up 56 percent year over year.

That energy-management line is the important one. It is how a power-rich operator turns grid scarcity into cash. As power gets scarcer, that line gets more valuable, and it is already the second-largest revenue source in the company. And the platform has shown it can monetize the assets themselves: it previously sold its Sandersville, Georgia site to CleanSpark for up to 42.5 million dollars, more than the company's entire market cap going into this takeover.

The macro: power is the bottleneck, and BGDE sits on it

The AI build-out is not constrained by chips or capital. It is constrained by energized megawatts with grid interconnection already in place, the thing that takes years to permit and build. Nowhere is that tighter than PJM, and the price proves it.

PJM capacity cleared at 28.92 dollars per megawatt-day for 2024 to 2025. It then cleared 269.92, then 329.17, and the most recent auction hit 333.44, the regulatory cap, the first time the entire grid fell short of its reliability requirement. Data centers drove the majority of that move, and PJM is warning of a shortfall of up to 60 gigawatts over the decade. BGDE controls a scarce, grid-connected position in the exact market where everyone now needs power and nobody can get it fast.

The tell: the people who know the most are buying the most

Cheap is not a catalyst. Aligned owners buying their own stock is. And the insider record here is as one-directional as we have ever underwritten.

From late November 2025 through June 2026, the controlling group bought stock almost continuously. They did not buy the bottom and stop. They paid 4.07, then 4.84, then 5.37, then 6.01, then 8.38, and on one print 12.00. They paid up into their own thesis. Then, in June, the entire newly seated board followed with personal cash in the high 6 to mid 7 range. Total open-market insider buying runs past 6 million dollars. Real insider selling is zero, into a float of under 6 million shares. That is accumulation by the most-informed buyers in the building.

The prize: AI conversion

This is where the asymmetry lives. The same megawatt earns very different money depending on what runs on it. Bitcoin and legacy hosting generate roughly 20 to 45 million dollars of revenue per 100 megawatts at mining-grade margins. The publicly disclosed AI and HPC colocation contracts that former miners have signed imply 140 to 225 million dollars per 100 megawatts at 40 to 65 percent site-level margins.

This is not theoretical. Former miners have signed enormous, long-dated AI contracts that re-rated their entire equities. These are the precedents the market will use to price BGDE the day it signs its first deal.

Be clear-eyed: BGDE has not signed an AI offtake yet. That is the un-hit catalyst, and it is exactly what the new team was installed to deliver. But that is also why you are getting 129 megawatts at a price that assumes it never happens.

What is $BGDE worth?

We value this on the asset, because that is what a re-rate prices. Today BGDE trades around 0.3 to 0.5 million dollars per megawatt. The miners that have signed AI contracts trade at 11 to 30 million per contracted megawatt. That gap is the opportunity, and the risk. On the company's own platform math the existing assets carry a stated net asset value near 75 million dollars, so the stock changes hands at roughly half of that book value before any AI conversion is priced in.

The variable is dollars per megawatt, and the swing factor is whether the company converts capacity into contracted AI compute.

• Base case, roughly 18 dollars. The miner-plus-demand-response business stabilizes, the market gives partial credit for 129 megawatts of PJM power, and the stock re-rates toward 1.5 to 2 million per megawatt. Roughly 2.5 times from here, and that is the case where the AI thesis never even lands.

• Bull case, roughly 55 dollars. The team signs its first real AI or HPC offtake, even 30 to 50 megawatts, and the contracted capacity reprices toward 5 to 7 million per megawatt, still a steep discount to peers for a first contract. Roughly 7 to 8 times, after the shares issued to fund the conversion.

• Moonshot, 150 dollars and up. Multiple sites convert to executed contracts and the platform re-rates toward the peer band of 11 to 30 million per megawatt. Twenty times and up. This is the home run, and it requires the team to execute against the clock, but it is the deck's own platform math, not a fantasy.

Against a price around 7 dollars, the base case is a winner and the bull case is a multi-bagger. The targets are aggressive on purpose. They are also anchored to the asset and to peer multiples, not to hope.

The team

The thesis rests on whether this group can do the one thing the old regime never did: convert power into contracted compute. The control group, operating as Endeavor and SixThirty.AI, are the chairman, chief executive, and operating chief, and they collectively own about 29 percent of the company. Their backgrounds run across energy, AI and HPC infrastructure, real estate, and operations, and one built and sold a cybersecurity company. The most important fact about them is not a resume line. It is that they have bought more than 6 million dollars of stock with their own money and put it under multi-year performance milestones. Their outcome is the public shareholder's outcome.

What are the catalysts?

This is an event-driven setup. The things that re-rate it, in order of importance:

1. The first executed AI or HPC offtake with a named compute partner. This is the single re-rating event. It converts a few-hundred-thousand-per-megawatt valuation into a contracted multiple overnight.

2. A new long-dated power contract replacing the December 2026 agreement at a still-competitive rate, defusing the cost-reset clock.

3. Refinancing or terming-out the near-term loans, removing the going-concern overhang.

4. A clean, independently priced combination of the controlled development pipeline into the public company, turning the platform story into public-company megawatts.

5. Litigation resolution at de minimis cost, clearing the last legacy overhangs.

What are the risks?

We are bullish, not blind. At any real position size you hold these in view.

• The cheap-power agreement reprices at the end of 2026. Midland's fixed-price contract runs through December 2026; beyond it, power costs move toward the PJM market. A new long-dated contract is the catalyst we are watching, and its absence is the clock on the cost advantage.

• Funding and dilution. The balance sheet carries a going-concern note and roughly 26.5 million dollars of near-term loans, and the funding answer is mostly debt rather than fresh stock. The executive chairman's own entity has extended the company a secured revolving credit line of up to 40 million dollars, while the public at-the-market equity program is capped at just 9.6 million, which holds near-term dilution down. Two items inside that stack are worth watching: a defaulted legacy promissory note with about 11.1 million dollars outstanding that now sits in active litigation, and a Nasdaq condition that requires at least 5 million dollars of stockholders' equity in every quarter through the middle of 2027. The value per share still has to outrun the share count, and so far the insiders are buying faster than the company is issuing.

• No AI contract is signed yet. The thesis lives or dies on the first real offtake. A prior 2024 attempt did not convert. This is a bet that this team gets it done.

• Related-party structure. Part of the growth pipeline and the flagship colocation deal involve entities affiliated with the same principals who control the company, reviewed under a stated independent-committee framework with the conflicted principals recused. The alignment here is structural, not cosmetic. The affiliate's upside on the colocation deal is paid in warrants struck at 20 dollars, nearly three times the current price, so those insiders only monetize that piece by driving the stock to 20 and above. It still deserves ongoing scrutiny, and the offset is that the same principals hold common stock alongside outside shareholders.

None of these is disqualifying. All of them are why a 129-megawatt PJM asset is still available at a fraction of replacement cost, and why we treat this as a high-conviction, defined-risk position rather than a sure thing.

Is $BGDE a buy? The bottom line

Big Digital Energy is a real 129-megawatt power operator in the most power-starved grid in America, with a demand-response revenue line growing 56 percent, nuclear-adjacent power, and an activist team that put more than 6 million dollars of its own money in from 4 to 8 dollars with zero selling. It trades at a few hundred thousand dollars per megawatt while contracted peers trade in the tens of millions. The one thing it has not done, sign an AI offtake, is the one thing that closes that gap, and it is the catalyst we are positioned for.

We are buyers. This is the kind of asymmetry the portfolio exists to find.

Disclosure

Shawarma Capital holds a position in BGDE and intends to add to it. This is research and opinion, not investment advice. Microcap stocks are volatile and can go to zero. Do your own work and size accordingly.