On the seventeenth of February 2026, SharonAI Holdings traded on the OTC Pink market at $1.90 a share. Two days later, on the nineteenth, it priced a Nasdaq offering at $30.00. Seven weeks after that it touched $97.48. Today it sits at roughly $50.
Nothing about the business changed in those forty eight hours. What changed was the venue, the underwriter, and the story. That gap between the venue and the business is the entire subject of this piece, and once you understand how it opened, everything else about this company becomes legible.
On the thirtieth of April, Bleecker Street Research published a short report calling SharonAI a neocloud built on phantom contracts and questionable financing. The stock was around $63 at publication. It is $50 now. The short is working, and most people have concluded the report was simply right.
We spent three research rounds and sixteen agents testing that conclusion against primary sources. We pulled the 555 page Indian prospectus of the customer at the centre of the report. We pulled the bankruptcy docket the report cites. We pulled the Form D that names the underwriter's fee. We read the audited accounts of the CEO's private vehicle in Australia. We enumerated the company's certificate transparency logs, its autonomous system number, its PeeringDB facility registrations and its status page incident history.
What we found is that the report was right about the paperwork and wrong about the price. Its two load bearing claims have not survived. Several of its sharpest lines are materially imprecise. And the single most serious conflict at this company is one the report did not identify at all.
The thesis in two sentences
SharonAI is a real, very small, asset light GPU cloud with roughly two to four megawatts actually connected, wrapped in promotional numbers its own filings do not support, and the market has already discounted the promotional numbers to approximately zero. At $50 the enterprise value is $1.18 billion against $1.86 billion of cash, which means cash is 94 percent of the market capitalisation, only $29.60 a share is at risk on the operating business, and two independent valuation back solves both land on the capacity figure in the audited filings rather than the one in the press releases.
That is not a defence of the company's disclosure. It is an observation that you are being asked to pay for the conservative case and handed the rest as optionality.
1. What this company actually is
Strip away the language and SharonAI is a tenant.
It does not own data centres. It sold the one it had. The flagship Texas campus, a 250 megawatt joint venture in the Permian Basin called Texas Critical Data Centers, was a fifty fifty partnership with a small helium and natural gas company. SharonAI sold its entire fifty percent stake in January 2026, one month before the Nasdaq pricing, for $70 million. It was paid early and in full, taking $74.8 million, and booked a $65.9 million gain. That gain, plus $5.0 million on the shares it received, is the only thing that made money in the first half of 2026.
This matters for a reason that is not obvious. The company carries $26.3 million of property and equipment against $2.32 billion of total assets. That looks alarming until you know the data centre was sold and the GPUs are not owned either. The fleet sits in two leased halls belonging to NEXTDC, the Australian colocation operator, in West Footscray Melbourne and Artarmon Sydney. A thousand GPU Blackwell cluster is consumed under a Lenovo pay as you go agreement, meaning it never touches the balance sheet at all. A further $302.65 million sits in prepayments to named suppliers, principally $256.19 million to WWT Australia, for hardware that has not been delivered.
So the accounting is not hiding anything. The accounting is telling you precisely what this is: a company that has paid cash to vendors for equipment it has not received, to install in buildings it does not own, for customers whose contracts begin in 2027.
The cloud itself is real, and we had to correct ourselves on this point. An early pass concluded the company was a brochure site with no product. That was wrong, because it probed the wrong domain. The actual platform lives on a fourth domain registered in August 2025, with working single sign on, self serve registration, 352 pages of technical documentation covering managed Kubernetes, GPU virtual machines, serverless inference and per second metering, and production hostnames resolving inside the company's own announced address space. The company holds a genuine autonomous system number registered to its founders' pre existing Australian vehicle, announcing 1,536 addresses across six blocks.
It is a real cloud. It is also a small one. The company's own status page, live since October 2025, lists two regions, both Australian, and records three incidents in eleven months. Self reported peering traffic is twenty to fifty gigabits per second. A production cloud carrying meaningful load generates considerably more of all three.
2. The short report, adjudicated
We took twenty six discrete claims from the Bleecker report and tested each against primary documents. Thirteen confirmed. Six overtaken by events. Five overstated. Two false. Two unresolved.
The pattern is consistent and worth stating plainly, because it is to the author's credit. Where Bleecker quoted a document, it quoted correctly. The failures are inferential rather than documentary. This was not a sloppy report. It was a report written in April about a situation that moved in May, June, July, August and September.
What died
The centrepiece was ESDS Software Solution, an Indian company that signed a $1.25 billion five year contract with SharonAI on the thirty first of March. The report's argument was arithmetic and looked devastating: ESDS earned $39.9 million and held $69.5 million of assets, yet had committed to pay roughly $250 million a year and post $140 million in letters of credit, a figure exceeding its entire balance sheet.
We pulled ESDS's own Red Herring Prospectus, filed with the Indian securities regulator on the twenty fourth of August. It is 555 pages, audited, and legally binding on its signatories. It did not exist when Bleecker wrote.
It confirms the contract term for term. Entered the thirty first of March 2026. Five years plus a two year option. Total contract value of $1.25 billion, stated in rupees as ₹118,312.50 million. Approximately 8,208 NVIDIA B300 GPUs, a more precise figure than SharonAI's own "8K". An Australian data centre. Monthly service fees.
The phantom contract claim is false. The contract exists, and the customer's own auditors have examined the disclosure.
More importantly, the prospectus reveals the funding mechanism, which nobody outside the two companies could previously see. ESDS's subsidiary has already received an advance of ₹11,766.35 million, approximately $124.3 million, from an enterprise customer incorporated outside India, for this project. That advance is what tripled ESDS's balance sheet, taking cash from ₹607 million to ₹12,534 million.
This is a prefunded back to back resale. ESDS is not a small company promising to pay a quarter of a billion dollars a year out of its own earnings. It is an intermediary that has already collected a large advance from the ultimate end user. And the number ties to something unexplained on SharonAI's own balance sheet: a customer deposits line of $143,879,911, carrying no note, no policy and no counterparty disclosure. The two sides of the trade reconcile.
The second load bearing claim was financing. The report attacked a $500 million facility from USD.AI, a decentralised finance protocol, noting correctly that it had roughly $284 million of lending capacity against $1.2 billion of approvals.
Every word of that arithmetic was right, and it remains right. We pulled USD.AI's live proof of reserves data. Its entire loan book across twenty seven deals sits in Sweden, Canada, France and the United States. Australia is zero. New Zealand is zero. Its first accountant attestation, published on the first of September, states the book is deployed across facilities in the United States, Canada and Sweden.
SharonAI never drew a dollar. Every filing for eight months has carried the identical sentence, that USD.AI had approved a facility subject to execution of definitive documentation. It was never upgraded to "we entered into" or "we drew". USD.AI appears zero times in the second quarter report.
The report named a lender that was never used. And in the interim the company raised approximately $2.15 billion from institutions that are emphatically not decentralised finance protocols: Oaktree, which holds $375 million across both convertible tranches, plus Millennium, Soros, Point72, Blackstone, Diameter, Sona and Lone Pine. Cash went from $71 million to $1,861,347,822. Annual interest is $54.6 million against 34 times coverage, and nothing matures before May 2031.
Three further claims decayed. The assertion that ESDS cannot legally serve Indian customers from Australia rests on Indian data localisation rules that cover payment system data only, and the end customer here is offshore, so the rule never bites. The Gazprombank exposure was confirmed to the decimal for fiscal 2025 at 20.15 percent, and has since fallen to 2.80 percent. The supervoting structure described as sixty five percent of votes on one percent of economics was a mismatched pairing then, and today stands at 38.2 percent of votes against 45.7 percent of economics, which is not a voting majority at all.
What survived, and what got stronger
The governance critique stands, and in one respect we strengthened it against the company.
The DSS acquisition, in which the founders sold a private Australian entity into the public shell, drew a comment from the Securities and Exchange Commission itself during the merger review, questioning whether it was a common control transaction that should carry no goodwill at all. Bleecker did not have that. It is corroboration from the regulator.
The NVIDIA misstatement is confirmed and remains live. The annual report filed on the thirty first of March, signed by the chief executive, stated that NVIDIA is a strategic shareholder in SharonAI. An 8-K filed on the thirteenth of April stated NVIDIA holds no equity securities. The annual report has never been amended, so the false sentence is still in the filing today, and the correction was filed under the "other events" item rather than the non reliance item. In fairness, the sentence sits in a single cell of a partner table whose adjacent row makes the identical claim about a different partner and is true, which reads as a copy into the wrong row rather than an invention. It is sloppy. It is not fabricated. It is also uncorrected six months later.
And the letters of credit do not exist. ESDS's certified indebtedness table, examined by an independent accountant as at the thirtieth of June, three months after signing, records letters of credit sanctioned nil and outstanding nil. Its entire non fund based limit is approximately $5.5 million against a required $140 million. The customer's own legally binding disclosure document does not record the security that SharonAI's lenders are being asked to underwrite.
What was overstated
Three lines in the report do not survive contact with the underlying record, and anyone repeating them should stop.
The report says the chief executive's prior public company is suing him for routing $11.5 million through a shipping company he controlled. There is no such case. A full sweep of all thirty Mawson Infrastructure dockets, plus searches on his name and his vehicles, returns zero federal actions naming James Manning as a party. Across 120 docket entries in the live adversary proceeding, his name and his company's name appear zero times. The allegation lives in Mawson's answer to an involuntary bankruptcy petition, a defensive pleading offering motive evidence against the creditors who filed it, and it has never appeared in any Mawson annual or quarterly report. The word identifying that shipping company returns zero hits in both.
The payments that did flow to that company from SharonAI were $167,638 in 2024 and $92,722 in 2025, and it appears nowhere in the 2026 related party note.
The report describes a $25 million self dealing transaction in DSS. The audited accounts of the CEO's own vehicle show it sold DSS for A$1,018,000 of stock, booking a A$900,074 gain on a carrying value of approximately A$118,000, which ties precisely to the $677,000 SharonAI disclosed. The $25 million was the whole company valuation across mostly unrelated holders. The insider take was about one million Australian dollars.
And the contractor arrangement the report attacked, under which the chief executive, chief financial officer and a co founder invoiced through their own companies, was terminated in April 2026 with no notice period and no payment.
On the regulatory record the negative is clean and deserves to be stated as plainly as the allegations. No action by the Securities and Exchange Commission. No banning or disqualification by the Australian regulator. No Nasdaq proceeding. No industry record. No criminal charge. No restatement, no non reliance filing, no auditor resignation over any of it. After three years of allegations, no regulator anywhere has acted.
These are serious, specific and entirely untested allegations. The fair weight is a standing governance discount, because the related party density is undisputed disclosure rather than allegation, plus a guidance credibility discount. It is not a fraud thesis.
3. What nobody has reported
Here is the part of this piece that is ours, and it is the most serious conflict at this company.
The underwriter that took SharonAI public bought its stock at $6.00 first.
The sequence is documented in two filings. A prospectus supplement filed on the thirteenth of February shows Lucid Capital Markets as one of three investors in $2,250,000 of December 2025 bridge notes that converted at $6.00 a share. Six days later Lucid acted as representative of the underwriters on the public offering priced at $30.00. On the seventeenth of April it initiated research coverage with a buy rating and a $50 price target. Nine days after that, a Form D filed on the fourth of May names Lucid Capital Markets as the sole recipient of sales compensation on the $350 million convertible offering, paid $14,875,000.
Bought at six. Sold to the public at thirty. Published a fifty target. Collected fourteen and a half million.
Lucid is placement agent on every financing this company has done, roughly $2.075 billion in total. And its Head of Capital Markets was the sole officer and director of the SharonAI registrant before the merger, and is himself registered to resell 465,343 shares. That same individual joined the board of the shell in 2023.
Which brings us to the shell. SharonAI did not arrive on Nasdaq through an initial public offering in any meaningful sense. The vehicle began as TKB Critical Technologies 1, which raised money in October 2021 to buy a connected car data company. That deal collapsed. In June 2023 the sponsor and directors sold 4,312,500 shares and 8,062,500 private warrants to affiliates of Roth Capital Partners and Craig-Hallum for one dollar. The vehicle was renamed, then failed two separate Nasdaq listing rules, one for having fewer than four hundred holders and one for not holding an annual meeting. It voluntarily delisted in April 2024, liquidated its trust, and saw ninety percent of its public shares redeemed.
That is the entity SharonAI merged into in December 2025, landing on the OTC Pink market. Two months later it uplisted to Nasdaq at $30.00, two days after trading at $1.90.
We also found a selling shareholder that bought at $6.00 whose trust belongs to the chairman of both the CEO's private investment vehicle and the shipping company named in the Mawson pleading. None of that relationship appears in the selling shareholder footnotes.
Two more items nobody has published.
The company bought land in Australia and did not tell its shareholders. On the ninth of September a regional newspaper reported that Visy sold its former paper mill near Albury, New South Wales, to SharonAI, with a company spokesperson confirming it on the record, corroborated by a second outlet and by a statement from the local member of parliament. This is the first evidence of any owned real property. It appears in no filing and in no press release. There is no development application lodged. It is bare dirt, disclosed to a regional paper and an MP but not to the market.
And the headline backlog includes the supplier. The second quarter release presents $8.8 billion of total contract value, and books $4.9 billion of that under a heading called customer momentum, attributed to NVIDIA. But the financial statements say the company committed to procure hardware for approximately $4 billion, and the chief executive describes the NVIDIA pricing as guaranteed as a minimum only, a floor rather than a ceiling, focused on the resale of that capacity. That is a purchase obligation and a price backstop presented as customer demand. Signed customer contract value is $3.893 billion, not $8.8 billion.
4. The capacity question, which is the real one
The company says 212 megawatts. Its own quarterly filing says 132 megawatts secured and 116 contracted. The number 212 appears nowhere in the filing's body.
We traced its origin. It first appears on the sixth of August, in an exhibit to the second quarter results. And contrary to our own earlier working note, it is not a contradiction of the release two days before. On the fourth of August the company said total capacity remains at 132 megawatts with 120 contracted. On the sixth it added a previously unannounced additional 80 megawatts, bringing the total to 212.
So a sixty one percent increase in claimed capacity was disclosed as a bullet in an earnings release and a remark on a call. No standalone filing. No facility named. No host named. No contract. Approximately 80 of the 212 megawatts has no disclosed home anywhere in the public record.
Against that, what is actually connected today is two to four megawatts, in two NEXTDC halls, verifiable through the company's own PeeringDB registration, which lists exactly two facilities and zero in the United States. The fleet was 432 GPUs as of September 2025, including sixteen H200s the company disclosed were acquired primarily for customer latency testing and proof of concepts.
The honest way to describe the gap is not fiction. It is tense. This is a 2027 business being described in the present tense. The largest tranche, 72 megawatts at NEXTDC worth roughly $623 million, does not deploy until April to August 2027. Three of the four headline contracts begin in 2027. The quarterly filing concedes the point in its own language, recording no liability because services have not commenced.
There is real corroboration that the contracts are binding. NEXTDC's own exchange filing on the twenty first of July reports contracted utilisation rising 73 megawatts in the same month SharonAI signed a 72 megawatt order. That is a landlord confirming the lease from the other side.
5. The tape, and why it fell
We began this work with a hypothesis that the decline from $97 to $50 was mechanical, driven by convertible arbitrage funds shorting the stock to hedge their bonds. We were wrong, and the data killed it cleanly.
Peak short interest was 2,744,284 shares, 7.8 percent of the class, against a theoretical fully hedged book of about 7.5 million. Short interest fell throughout the entire decline, from 2.74 million in July to 2.36 million at the end of August. Days to cover ran between one and one point four. Borrow is easy and cheap, with 800,000 shares available at 2.45 percent. The stock has never appeared on the threshold list. The options book is 3.3 to one long calls, which is retail buying upside, not desks hedging bonds.
The actual explanation is supply, and it is dated to the day.
Class A shares went from 11,832,164 at the end of December to 35,667,164 by August, roughly tripling in six months. On the fourteenth of August, 13,094,719 shares priced at $68.73 became freely saleable. On the twenty first, a further 8,056,699 priced at $12.53 joined them. That is 21,151,418 shares, 59.3 percent of the entire class, unlocked in eight days, on registration statements the Securities and Exchange Commission expressly declined to review. The stock fell 23.2 percent in the four sessions between the two effective dates, and the high of $76.90 came three sessions after the first.
Seventy six percent of that freed equity is unhedged.
The largest holder is worth understanding precisely, because it is simultaneously the biggest risk and a meaningful signal. Situational Awareness, the fund run by the former OpenAI researcher Leopold Aschenbrenner, holds 8,070,950 shares and prefunded warrants, 21.1 percent on the basis its own filing uses. Its cost is $523,882,863, or $64.91 a share, leaving it roughly $105 million underwater. Its fund shrank from $45 billion to $10 billion. It converted from a passive to an active filing on the twenty eighth of August, and the nineteen point nine nine percent ownership blocker ceased to apply on the twenty seventh.
It has sold nothing. Zero shares. Through a forty six percent drawdown.
Neither have insiders, in either direction. Five ownership forms since listing, one of them a zero dollar grant, no sales, no plans, no buying.
And the retail story is a non story. The company did run a paid investor relations campaign, disclosed properly under the relevant rule, costing $200,000 in total, $50,000 in monthly fees and $150,000 for a four week television campaign that aired in March. It lapsed in the second quarter. The promotional window and the price window do not overlap. A census of 1,356 messages on the main retail forum returns no coordinated phrasing, and the twelve and a half percent decline on the fourteenth of September drew four messages. Retail has left. It did not make this price, and it is not defending it.
6. What you are actually paying for
This is where the argument turns, and it turns on arithmetic rather than opinion.
At $50.17 the basic share count is 35,803,505 and the treasury method count is 39,430,551. Enterprise value is $1,180.9 million, and across six different definitions the range is $0.83 billion to $1.39 billion, narrow enough that the choice does not matter.
Cash is 94.1 percent of the market capitalisation. Only $29.60 a share is exposed to the operating business at all.
Now run the two back solves.
At the peer median of roughly $9.0 million of enterprise value per megawatt, $1,180.9 million implies 131 megawatts. The quarterly filing says 132.
At the peer median of 3.48 times forward revenue, it implies $339 million of 2027 revenue. That is 44 percent of the signed book's run rate and 22 percent of the four analyst consensus.
Two independent methods, approached from different directions, both land on the number in the audited filing rather than the number in the press release.
The 212 megawatts is not in the price. The $8.8 billion is not in the price. The 68,000 GPUs are not in the price. Every promotional figure this piece has spent six sections dismantling has already been discounted to approximately zero by the market. The promotional record bear case is paid for.
Meanwhile the company trades at 0.77 times forward revenue against a peer median of 3.48, the cheapest in its sector by a factor of 3.2, while simultaneously trading at $328 million per live megawatt against a peer band of $45 to $64 million. Both are true at once, and the reconciliation is the cash.
The unit economics are also better than the narrative implies. The contracts price at $18.6 to $19.0 million per megawatt per year, which is below one listed peer's disclosed figure and inside another's stated band. The pricing is market. The delivery is the anomaly. Gross margin is driven far more by depreciation life than by utilisation: on a four year life the range across seventy to one hundred percent utilisation is negative 21.5 to positive 12.9 percent, while on a six year life it is positive 13.2 to positive 37.1 percent.
And the funding shortfall, correctly framed, is a growth gap rather than a solvency gap. Total uses through 2027 are $2,141.6 million against $1,861.3 million of cash. That is short $280.3 million with no customer prepayments, short $111.3 million at ten percent, and a $57.7 million surplus at twenty percent. We know prepayment is real, because ESDS has already advanced $124.3 million. Interest coverage is thirty four times and nothing is due before May 2031.
7. The dated tests, and what they are saying right now
This thesis is unusually falsifiable, which is the best thing about it. Four tests carry dates.
The first was the sixteenth of September, the contractual delivery date for the ESDS cluster. As of writing there is no filing, no announcement, and no evidence anywhere that the $140 million security has been posted. More pointedly, on the tenth of September the customer's own chairman told Indian television that the Sydney facility would go live in the next forty to fifty days, by the end of October. Six days before the contractual date, the customer was publicly guiding six weeks past it. He did not name SharonAI, and the connection runs through the only Australian arrangement ESDS discloses, so treat it as strong inference rather than proof.
The second is the twenty sixth of September, when a 15 megawatt Sydney facility was said to be going live. The landlord's last public statement was in May. Its newsroom has not posted since June and has never named SharonAI. The company's PeeringDB registration still shows two facilities, last updated on the twenty ninth of July. No third building has appeared.
The third is secured financing. On the twenty first of August the convertible holders consented to strip the covenants limiting debt and liens, which is unambiguous preparation for asset backed borrowing, and which Oaktree as a required holder with a veto necessarily approved. Twenty five days later nothing has been filed.
The fourth is the cleanest and cannot be spun. When the next quarterly report is filed, either it recognises an operating lease liability for the new capacity or it does not. Accounting standards require recognition when a lease commences. If those megawatts are live, the liability appears. If it does not appear, they are not live.
Two of four have failed so far. One is pending with the customer guiding past it. That is the honest scoreboard, and it is why this is a 2027 story.
8. What would make us wrong
If the September and October deliveries land and the next quarterly report shows a lease liability stepping up, the live footprint roughly quadruples and the 2027 book starts converting to revenue. That is the bull path and it is measurable.
Against it, here is how the long thesis breaks.
The ESDS contract is renegotiated or shrinks, which is the single most likely adverse outcome given no letter of credit exists, the customer is a reseller with a twenty percent gross margin, and its own analyst models an earnings margin collapsing from 50.8 to 18.1 percent.
The secured financing does not arrive, or arrives at punitive terms, leaving the 2027 tranche unfunded. Watch for any filing that follows the covenant strip.
Situational Awareness sells. Eight million shares, twenty one percent of the class, a hundred and five million dollars underwater, in a fund that lost most of its assets, with the blocker now removed. Nothing obliges it to be patient.
The May convertible strikes at $48.24, which is at the money today. A recovery arms that dilution rather than relieving it.
The auditor position deteriorates. Three firms in twenty months, no internal control audit ever performed, and a material weakness in complex financial instruments that remains unremediated.
Or delivery simply keeps slipping, and 2027 becomes 2028, at which point the cash cushion that makes this defensible starts being consumed by interest and overhead rather than by building.
The position
The short report was right about the paperwork. The capacity number is promotional, the backlog counts the supplier, the console has been removed from public DNS, the land purchase went to a regional newspaper instead of the market, and the underwriter bought at six and sold you thirty.
It was wrong about the price. The contract is real and prefunded. The financing it attacked was never drawn, and two billion dollars arrived from Oaktree and Millennium and Soros instead. The lawsuit it described does not name the man it describes. The twenty five million was one million. And the market has already marked this company at the number in its audited filing rather than the number in its press release.
At ninety four percent cash backing, you are paying roughly thirty dollars a share for a business the market is valuing at the 132 megawatts on file, with the 2027 book as free optionality and four dated tests in the next ninety days to tell you whether it converts.
That is not a story stock at this price. It is a financing and delivery question with a large cash cushion behind it, and the honest answer is that we will know a great deal more by the end of October than we know today.
Disclosure and disclaimer. This is our own research and opinion. It is not investment advice and not a solicitation to buy or sell any security. Figures are drawn from company filings, exchange filings, foreign regulatory prospectuses, court records and public infrastructure registries, and we make no warranty as to their accuracy or completeness. Bleecker Street Research has disclosed a short position in SHAZ. Do your own work.