Two Channels Worth $1B on Paper

The next time you find yourself on EDGAR, consider using my SECSift reader instead!

It turns SEC filings into clean, readable documents. You can search every word, export anything, merge tables, highlight what matters and gray out the rest, and compare against previous filings.

Not familiar with the reader? Here’s a 3-minute interactive demo (viewable on your computer) and another video I made walking through the entire tool.

🌾 Welcome to StableBread’s Newsletter!

I wrote up Ambarella (Nasdaq: AMBA) on May 29, 2026, when the stock traded ~$90/share.

Ambarella is the only listed Western company built around edge AI as a pure play, with 80% of revenue from chips that run vision and AI on the device itself, across security cameras, drones, and driver-assistance systems.

It has robot design wins too, but none of them are producing revenue yet.

AMBA stock has experienced quite a volatile ride since my write-up:

At ~$64/share, Ambarella is a $2.82B company (44.14M shares × $64) with ~$260M of net cash ($5.88/share). Strip the cash out and the operating business carries an enterprise value (EV) of $2.57B, or 5.8x EV/Sales on the FY2027 guide midpoint ($2.57B / $440M).

In May, at $90/share, that multiple was 8.4x. It’s compressed 31% since.

Below, I provide an updated analysis on Ambarella’s position as a pure-play edge AI chipmaker, and on whether the lower price reflects a weaker business or the same one at a discount.

With robotics and physical AI running as hot themes right now, that’s a distinction worth getting right.

The May 29 write-up rested on (1) a product cycle that lifts revenue per chip, (2) two long-term agreements (LTAs) that lock in multi-year volume, and (3) a robotics pipeline that hadn’t produced revenue yet.

Pricing

Ambarella’s second-generation CNN chips, the ones that detect and classify what a camera sees, sell for $10-75 per chip. Its third-generation chips, which also run transformers (the model architecture behind generative AI), sell for $20-400 per chip.

As the mix shifts toward those third-generation chips, revenue per chip rises without a single new customer.

Ambarella’s system-on-chip (SoC) average selling price (ASP) rose ~15% in FY2026 vs. FY2025, and management says every new edge AI chip in development is expected to price well above the current ASP.

That ASP was ~$15 in FY2026, and Wang said at Citi's 2026 Global TMT Conference on September 9 that the higher prices don't widen the margin. More AI performance takes a physically larger chip, and a larger chip costs more to build, so the higher price is meant to cover the higher cost rather than add to profit.

Just look at the company’s own ASP slide below:

Long-Term Agreements

These are multi-year customer commitments that cover more than one chip generation, so a customer signs up for the current product and its successors.

Ambarella had two in May, the 2-nanometer semi-custom CV8 program and the Hanwha Group agreement worth >$800M in potential revenue over 10+ years.

It still has those two. Neither the Q2 release nor the call disclosed a new one, and the semi-custom program books its first production revenue in FY2028.

Robotics

In May, Ambarella had 15 robot design wins, drones included, carrying ~$100M of cumulative pipeline revenue.

On the September call, Fermi Wang, Ambarella’s President and CEO, said more wins have been added and the revenue target is higher, without giving new numbers.

Q2 FY2027 revenue was $108.1M, up 13.2% y/y and just above the $108M guide midpoint. Edge AI and automotive revenue both set records on the 5-nanometer CV72 and CV75 ramp, and IoT and auto grew sequentially.

GAAP net loss narrowed to $6.7M from $20.0M a year ago. But $9.0M of that improvement didn’t come from selling chips.

An automotive autonomy customer had paid Ambarella a $13.5M deposit to fund a development project. Ambarella terminated the project on May 12, refunded $4.5M, and booked the remaining $9.0M as a reduction of R&D expense in Q2.

Strip it out and the operating loss is $17.1M, not the $8.1M reported ($8.1M + $9.0M). R&D was $50.6M with the credit and $59.6M without it, against $59.7M a year ago. So R&D was flat y/y, not down 15.3%.

Non-GAAP results exclude the credit, so the $0.18 of adjusted EPS (vs. $0.15 a year ago) doesn’t contain any of the $9.0M.

Non-GAAP gross margin was 59.3%, down from 60.5% a year ago and below the midpoint of the 59-60.5% guide. Q3 is guided to 59-60%, the floor of the 59-62% long-term model.

Margin came down for two reasons: (1) automotive carries a lower margin than IoT and set a record in the quarter, and (2) the 5-nanometer chips cost more to build than the generation they’re replacing.

Distribution concentration also shifted. WT Microelectronics, the Taiwanese fulfillment partner that ships Ambarella’s chips to multiple customers across Asia, fell to 60.2% of revenue from 71.3% a year ago. Hakuto, the Japanese distributor, came in at 11%.

A year ago WT was the only customer Ambarella disclosed above the 10% threshold.

Behind those distributors, Ambarella highlighted wins in four markets:

Three announcements landed alongside the quarter, two channel deals and one new chip.

Two Channel Deals

Wang said virtually all of Ambarella’s revenue is generated by its direct sales teams, and that the two new partners target customers who have largely been unserved so far.

Louis Gerhardy, VP of corporate development, described them on the call:

Wang sized both on the call:

Those are seven-year figures, not annual ones. Spread evenly, ~$1.0B across Macnica and Capgemini is ~$143M/year at full run rate, and Hanwha’s >$800M over 10+ years is ~$80M a year. That means ~$223M/year of potential revenue once all three are ramped, against the $440M guided for FY2027.

Management put meaningful revenue from the indirect channel 2-3 years out.

That lag is because the distribution side starts with small, fragmented design wins that have to aggregate into volume, while the Capgemini side brings large enterprise deals with long and complex sales cycles.

Macnica has already won designs in drones, retail, and manufacturing, so the first small revenue should show up within a year.

X7 AI Accelerator

The third announcement is a product, and it’s new. Wang introduced X7 on the call as Ambarella’s first standalone AI accelerator, sampling now, with initial design wins expected in edge infra.

It hadn’t appeared in a prior earnings call or investor deck, and Wang was careful to say it still hasn’t been formally announced.

Ambarella’s camera chips are SoCs that pack camera perception, the AI accelerator, the CPU, and encoding onto one piece of silicon, which means the customer builds the whole system around that chip. X7 isn’t camera-specific, so it runs as a co-processor next to an ARM or x86 host instead.

Therefore, a customer already using an Ambarella chip can add X7 for more AI performance without redesigning the board, and a customer running someone else’s CPU can add it anyway.

X7 is built for on-premise inference rather than the data center. Wang said it runs LLMs on 4MB of memory inside a 4-5W power envelope, against edge GPUs that need far more of both.

Wang was fairly direct about the competition. He said NVIDIA (NVDA) and Qualcomm (QCOM) already have similar products in this space, and counted ~50 startups building the same kind of chip.

That’s a more crowded field than the camera SoC business, where Ambarella is the only listed Western pure play.

So how does Ambarella compete with the top players? Wang’s argument is performance per watt plus mature hardware and software, and he put Ambarella among “the very few” that have both.

All of this feeds the number management led with. Ambarella’s rolling five-year serviceable market (SAM) now runs from $8.5B in FY2027 to $22.9B in FY2032, a CAGR management puts at ~20%. The prior forecast, announced in May 2025, ran to $12.9B by FY2031 at ~18%.

IoT is now ~70% of the terminal-year SAM, and Wang called edge infra the single most important reason for the raise.

A SAM number is what management thinks it could sell into, not what it expects to sell. Ambarella guides FY2027 revenue to $440M against an $8.5B SAM for the same year, or ~5% of that market.

But the direction is a change worth noting. In May, Ambarella sold chips that go inside other people’s cameras and cars. It now also sells chips for the on-premise boxes and servers those cameras send their video to.

Ambarella maps both halves on one slide:

The right-hand column of that slide, edge infra, had one product in the market in May, the N1-655, and it now carries most of the SAM raise.

Developer AI workstations, AI vision boxes, and on-premise servers all run on that N1 line, which is separate silicon from the camera SoCs. So this is new chips sold into new sockets, not a repackaging of what Ambarella already ships.

Memory prices are the reason the full-year guide now carries a condition.

Memory vendors are prioritizing AI data center demand, which has pushed DRAM and flash prices up and supply down across the industry. Ambarella’s customers build the cameras, dash cams, and drones that need that memory, and Wang said the company is helping them design workarounds.

The effect on Ambarella’s own margin is smaller than it looks. Ambarella doesn’t buy or resell memory, so the price doesn’t touch its gross margin. Wang said so on the call:

The pressure lands one step downstream instead.

Ambarella’s chip is one component inside a security camera or a dash cam, and memory is another component in the same device. When memory costs more, the finished camera costs more, and Ambarella’s customer either absorbs the hit to its own margin or raises the retail price.

In other words, fewer cameras sell, fewer cameras get built, and fewer Ambarella chips get ordered.

The price effect is the smaller risk. The bigger one is customers not being able to get memory at all. Asked whether the 10-15% full-year guide still holds, Wang’s answer was conditional:

Wang was more specific six days later, at Citi’s 2026 Global TMT Conference on September 9. Q3 is covered, Q4 isn’t. His customers told him that none of their memory suppliers will commit to a November allocation, and that they don’t know what price they’d pay if they got one.

Wang also said it doesn’t end this calendar year. It runs into 2027, and if it runs long enough it reaches Ambarella.

Worth remembering that this company guides low. As I covered in the May write-up, Ambarella entered FY2026 guiding to mid-to-high-teens growth and delivered 37%.

So 10-15% is most likely a floor management expects to lift through the year rather than a ceiling, unless memory takes that decision out of their hands.

Ambarella is also facing higher supply chain costs of its own, and plans to pass them through. Wang said the company will “pass this cost to our customer to maintain our long-term gross margin target of a 59-62%.”

That’s a price increase on Ambarella’s chips landing on a buyer who is already paying more for memory. Wang gave no size, no timing, and no indication of which products it covers, and Q3 gross margin is still guided to 59-60%, so none of it shows up in the guide yet.

If customers take the increase, the 59-62% model holds. If they push back, the margin gets defended out of volume instead.

Memory isn’t the only tight link. Wang said packaging and testing are tightening too, and Ambarella’s hedge is exclusivity.

Samsung Foundry manufactures chips for fabless companies, the ones that design silicon but don’t own a factory. Ambarella is the only fabless company that uses Samsung exclusively, and it does the same with the outside firms that package and test its chips.

Ambarella’s first 2-nanometer chip goes into production at Samsung, and Wang said staying with TSMC would have put Ambarella behind much larger customers in the queue. It also means one foundry on one process for the chip that carries the semi-custom program.

The May write-up called geography Ambarella’s biggest risk, and Q2 didn’t change that.

On a bill-to basis, Taiwan was $65.1M of the quarter and the rest of Asia Pacific $25.1M, so Asia was 83.4% of revenue ($90.2M / $108.1M). The United States was $2.1M, or 1.9%.

But concentration has started to ease at the edges:

Here’s the full geographic split from the 10-Q:

None of that changes the exposure, which has already cost Ambarella twice, once in revenue and once in the stock price:

The 10-Q flags two live constraints. Commerce applies to Hong Kong the same export controls it applies to China, including licensing on certain SoCs, and Ambarella ships the substantial majority of its finished chips through a Hong Kong warehouse.

Separate U.S. rules on transactions with Chinese entities could limit the R&D it runs in China, where a large part of its employee base resides.

Robot Rules Favor Ambarella

Trade policy isn’t only a risk here. On robotics it currently points Ambarella’s way, for the same reasons I covered in my Agility Robotics write-up on September 6.

Washington spent 2026 restricting Chinese robots in the U.S. market:

Every one of those rules pushes Western robot makers away from Chinese silicon, and Ambarella is one of the few non-Chinese options at the low-power end of that market.

That’s a tailwind on the design-win side of robotics at the same time that Asia concentration is a risk on the revenue side, which is why I still treat the robotics pipeline as optionality rather than a forecast.

Wang made the same point himself at Citi. Ambarella has a couple of humanoid design wins, and he said he wouldn’t bet the volume changes anything material about Ambarella’s financials. He also thinks humanoids are a harder technical problem than a level 5 self-driving car.

Ambarella pays no dividend and carries no debt, so capital allocation comes down to (1) funding chip development, (2) buying back stock, and (3) issuing stock to employees.

The board authorized a new $50M repurchase program on May 27, running July 1, 2026 through June 30, 2027. As of July 31, the full $50M was still available, and Ambarella bought nothing in Q2.

The last time Ambarella bought stock, it took 47,798 shares at an average of $51.04 in Q1, near the March lows during the Insta360 case. Whether August changed anything shows up in the Q3 10-Q.

Q2’s cash flow gives a reason to hold off:

Management framed the inventory build in May as positioning for new product ramps and tightening supply. Memory scarcity has since been added to that explanation.

H1 FY2027 is worse. Operating cash flow was a $25.9M outflow, against a $20.3M inflow a year ago, and FCF was a $36.7M outflow (-$29.6M in Q1, -$7.1M in Q2). Ambarella generated $58M of FCF in FY2026.

The bigger number here is the stock Ambarella issues to employees. It expensed $22.7M of SBC in Q2 alone, 21% of revenue (I’d prefer lower), and ~$98M across FY2026, against a repurchase authorization of $50M for the entire year.

Diluted shares went from 42.5M to 44.0M over the past year. At ~$64, the full $50M authorization would retire ~780K shares ($50M / $64), a bit over half of that increase. So the buyback would only slow the share count’s growth.

As a quick reminder, the Financial Times reported on July 31 that NXP Semiconductors was in talks to acquire Ambarella, which sent the stock up 16.1% to $86.00 that day.

Nothing has been announced since. No merger agreement, no 8-K, no confirmation from either company.

Across the prepared remarks and a full Q&A with nine analysts, NXP never came up. Management didn’t raise it and nobody asked.

That silence isn’t a signal. Live talks would sit under a confidentiality agreement, and dead talks would have nothing to announce, so both look the same from the outside.

At $86.00 you were paying for the deal. At ~$64 you’re ~$10/share below the pre-report close of $74.09, so the premium is gone and then some.

The strategic logic hasn’t changed. Automotive was 55.7% of NXP’s revenue in H1 2026 ($3.72B of $6.68B), and Ambarella’s edge AI and computer vision chips slot into that portfolio. This is also the second time in just over a year that Ambarella has drawn acquisition interest, after reports of a sale exploration in mid-2025.

A confirmed bid is upside you aren’t paying for at this price. A report that the talks have ended would knock the stock, though there isn’t much premium left to give back.

At ~$64/share, Ambarella has a market cap of $2.82B (44.14M shares × $64). Net cash is $260M ($272.3M of cash and securities less $12.6M of operating lease obligations), or $5.88/share, which leaves an EV of $2.57B…