$AMBA: The Eyes of Physical AI, an $800 Million Handshake, and a Tape That Gave the Entry Back
Ambarella designs the low-power vision chips inside every machine that has to see and decide on the spot. Rosenblatt called it the physical AI pure play at $120. The stock trades at $69. The full build, holes included.
Position disclosure: I have no position in AMBA as of this writing. This is the funnel, not a victory lap. If the name survives its own risk list, and this piece walks every item on that list, I may open one. You will hear it here first when I do.
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.
WT Microelectronics exceeds 75 percent of revenue for two consecutive quarters. Concentration ratchets are how single-distributor stories end.
Insta360 is confirmed dual-sourcing or replaced in a flagship generation, or an appellate ruling reinstates an import ban on its US products.
The Hanwha agreement produces no disclosed design win, NRE milestone, or revenue attribution by the fiscal 2028 first half.
Non-GAAP gross margin prints below 55 percent in any quarter. The moat claim dies there.
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
Ambarella Q1 FY2027 financial results, May 28, 2026. investor.ambarella.com and SEC Form 8-K.
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.
Hanwha and Ambarella long-term edge AI agreement, May 28, 2026. Business Wire.
Ambarella FY2026 results, February 26, 2026. Business Wire.
Rosenblatt Securities, second-half 2026 top picks, June 30, 2026, and reiterated buy, $120 target, Kevin Cassidy. CNBC, Yahoo Finance, Investing.com coverage.
Form 4 filings, Ambarella insiders, June 17 through July 9, 2026. SEC EDGAR, summarized via Finviz insider table.
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.
GoPro founder loan disclosure, July 9, 2026. Digital Camera World.
Q1 FY2027 earnings call coverage, May 28-30, 2026. Motley Fool transcript, Globe and Mail, GuruFocus highlights.
MarketBeat, July 15, 2026, AMBA down 8.9 percent, volume analysis.
Price and market data as of July 15, 2026 close, Yahoo Finance and StockAnalysis.com.