Pure-Play Below NVIDIA

🌾 Welcome to StableBread's Newsletter!

On May 20, 2026, NVIDIA (NVDA) reorganized its financial reporting into two segments: (1) Data Center and (2) Edge Computing.

That matters because of where AI money goes today. 95% of all AI spending flows into the data center. At least that's what Fermi Wang, CEO of Ambarella (NASDAQ: AMBA), claimed at the Morgan Stanley March 2026 Conference.

When the company selling more AI hardware than anyone else carves out a dedicated segment for everything outside the data center, it suggests the industry sees that ratio starting to shift.

A model can't brake a car, steer a drone, or pull the right box off a warehouse shelf from inside a server rack. Those jobs need AI running on the device itself, in real time, on a small power budget, without sending video to the cloud and waiting for an answer.

That's edge AI. The subset where a machine has to perceive its surroundings and then act on them is called physical AI.

NVIDIA sells the expensive brain for that world, and it's pushing further into edge AI too. But only at the top end, where chips cost thousands of dollars per unit.

There's a much larger layer beneath that top tier, where the job is cheap and low-power rather than maximum horsepower. Cameras, drones, mainstream cars, and smaller robots all live in that layer.

Ambarella has spent more than a decade and over $1.3B building chips and software for exactly that market. It's the only listed Western company built around it as a pure-play, and edge AI made up 80% of its revenue last fiscal year.

Revenue has grown from $226M to $391M over two fiscal years (+73%), and the company turned adjusted profitable in FY2026. Ambarella sizes its serviceable addressable market (SAM) at ~$12.9B by FY2031 (18% 5-year CAGR), against a current market cap of ~$4B.

The company just reported Q1 FY2027 earnings on May 28, 2026, with revenue of $100.4M (+16.9% y/y), tracking above its 10-15% full-year growth guide. It also announced a 10-year long-term agreement (LTA) with Hanwha Group worth >$800M in potential revenue, covering video security, robotics, industrial automation, and life sciences.

Below we'll discuss the business, the LTAs, the long-term robotics opportunity, and the valuation at $90/share.

Ambarella, founded in 2004, builds system-on-chip processors (SoCs). These are single pieces of silicon that pack video processing, AI computation, and system control into one chip.

The AI muscle inside each chip is Ambarella's own design, an accelerator it calls CVflow.

An accelerator is the dedicated circuitry that runs neural networks fast and at low power. CVflow is what lets a battery-powered camera or robot run models that would otherwise need a far bigger, hotter chip.

Software matters as much as the silicon. Ambarella's Cooper Development Platform lets a customer write its application once and move it across Ambarella's chip lineup without rebuilding it each time. That portability is a big reason customers stay.

The platform now supports more than 200 different AI model architectures in production.

The business splits into two end markets. In FY2026, IoT was 78% of revenue and automotive 22%.

IoT is mostly cameras. Security cameras for businesses and homes and portable video (action cameras, 360-degree cameras, wearables, and drones) are roughly equal at ~45% each. Robotics and access control are the smaller third bucket.

Automotive is mostly safety, not self-driving.

Around 95% of auto revenue comes from advanced driver assistance systems (ADAS), fleet telematics, dash cams, electronic mirrors, and driver-monitoring cameras. The other 5% is autonomy, and almost all of that is commercial trucks.

The commercial telematics market is a key growth driver. The installed base sits at 100M+ vehicles growing ~10% CAGR, with only 10-20% of it AI-enabled today.

Ambarella sells largely through a single distributor, WT Microelectronics, which accounted for ~70% of FY2026 revenue and 60.7% in Q1 FY2027. WT is a fulfillment partner, not an end customer, and it ships Ambarella's chips to many different buyers across Asia.

No single end customer is disclosed at more than 10% of revenue, so the underlying base is more diversified than the WT figure suggests. The real exposure is to a few end markets, security cameras, drones, and ADAS, rather than to any single buyer walking away.

Ambarella has shipped more than 46 million edge AI chips, and almost all of its revenue today comes from one model type, the convolutional neural network (CNN).

CNNs detect and classify objects in an image. A security camera spotting a person, a car reading a lane line, a drone avoiding a wall, those are CNN jobs.

Ambarella's second-generation chips, the CV2 family, run CNNs, and they're ~80% of revenue.

The next wave is the transformer, the model architecture behind generative AI and LLMs. Transformers let a device do more than spot objects. They let it reason about a scene, follow an instruction, and plan an action.

Ambarella's third-generation chips, the CV7, CV72, CV75, CV3, and N1 families, run transformers as well as CNNs. CV72 and CV75 started ramping into production late in FY2026. CV7, the company's first 4-nanometer chip, is set to ramp at the end of FY2027.

The transformer shift matters for pricing. Ambarella's CNN chips sell for $10-75 each. Its third-generation chips sell for $20-400, because running a multi-billion-parameter model needs a bigger, more capable chip. CV7 alone delivers 2.5x the AI performance of the chip it replaces.

It also opens new uses. A transformer-capable chip can turn a store's existing security cameras into something that reads what customers buy and how they move, not just records footage. Ambarella's N1 family powers "AI boxes" that pull in feeds from several cameras at once and add that generative-AI layer, with no need to swap the cameras.

The first commercial proof point landed in March 2026, when i-PRO (one of Ambarella’s customers) launched the first edge endpoint camera to run GenAI locally, powered by CV72.

The early numbers back the shift. Edge AI revenue grew about 50% last year and reached that 80% share of the total. Over five years it has compounded at a 64% rate.

The customer mix tells the same story. All 11 representative customer engagements in Q1 FY2027 were on edge AI SoCs (3 from the CV2 family, 8 from the newer CV7x family). It was the first time the entire list was on edge AI.

Revenue hit $390.7M in FY2026, up 37% from the year before, after $284.9M in FY2025 and $226.5M the year before that.

The fourth quarter came in at $100.9M, up 20% from a year earlier and down 7% from the prior quarter on normal seasonality.

The quarterly path through FY2026 shows both the ramp and a cooling growth rate:

The deceleration looks worse than it is, because it's mostly a base effect.

FY2024 was a cyclical trough: Chinese surveillance and drone customers were hit by U.S. Entity List restrictions, the broader IoT customer base destocked, and quarterly revenue bottomed well below $60M.

FY2025 and the first half of FY2026 then lapped those depressed comps, which is what produced the 50-63% growth rates.

By Q3 FY2026 the comparison was against a normalized base, and growth settled into the high teens. That's closer to the underlying pace of the business, not a sudden slowdown.

Profitability improved on every line in FY2026. Gross margin held around 60%. The operating loss narrowed from $126.6M to $82.5M. The net loss narrowed from $117.1M to $75.9M, or $1.78/share.

The bigger marker is adjusted profit. Measured on a non-GAAP basis, which strips out SBC and acquisition costs, Ambarella earned $26.9M, or $0.62/share. A year earlier that figure was a $6.8M loss. Free cash flow (FCF) was $58M, about 15% of revenue.

Q1 FY2027 came in at $100.4M, up 16.9% y/y and tracking above the 10-15% full-year guide. Adjusted EPS was $0.11, the fifth straight quarter of non-GAAP profitability. Q2 is guided to $108M at the midpoint, or 13% y/y growth.

Automotive hit an all-time record on commercial-vehicle AI adoption, but auto typically carries a lower margin than IoT. That's why gross margin softened to 58.4% GAAP and 59.9% non-GAAP, down from 60.0% and 62.0% a year earlier.

Q2 non-GAAP gross margin is guided to 59.0-60.5%, and management reiterated the 59-62% long-term non-GAAP model.

FY2026 revenue grew 37% y/y. But this growth can’t continue, at least not by Ambarella's own guidance.

On the Q4 FY2026 call in February 2026, management guided FY2027 revenue growth at 10-15%, ~$440M at the midpoint. That's a significant drop from 37%, and there are two ways to read it.

The optimistic read is that Ambarella guides conservatively. The company entered FY2026 guiding to mid-to-high-teens growth and delivered 37%, raising its forecast as customer orders came in.

Here's what the VP of corporate development, Louis Gerhardy, had to say about Ambarella's guidance:

The cautious read is that FY2026 had a tailwind that won't repeat. Growth slowed every quarter last year, from 58% in the first quarter to 20% in the fourth, and Q1 FY2027 came in at 16.9%. The deceleration is real and continuing.

Q1 FY2027 supports the conservative-guide read at the margin. Revenue of $100.4M was above the 10-15% full-year pace, and the Q2 guide midpoint of $108M (+13%) is also slightly hotter than the full-year midpoint.

To hit the $440M guide midpoint, H2 would only need to grow ~10.5% y/y, well below where Q1 and Q2 are tracking. That's the pattern of a conservative full-year guide.

Fermi Wang explicitly held the FY2027 guide at 10-15% on the Q1 call, which is informative both ways. The beat without a raise tilts toward the conservative-guide thesis at the margin, but it's also a signal that management isn't yet confident enough in H2 to commit.

The real read on the year comes from Q2 and whether the full-year guide moves then.

The auto side has its own tailwind. Industry forecasts, mentioned by Fermi, call for the global auto semiconductor market to grow 10-15% this calendar year against an overall auto production decline of 1-2%. Ambarella expects to outpace that rate, driven by commercial fleet telematics and safety, where new product cycles are landing.

Both reads still hold some truth. The fair conclusion is that 10-15% is a floor management expects to lift through the year, not a hard ceiling.

Ambarella spent $238.5M on R&D in FY2026, up just 5.5% while revenue grew 37%. Total operating expenses rose about 5%. Revenue climbed ~7x faster than the cost base.

That gap is why the operating loss is shrinking quickly, and why the company reached adjusted profitability for the first time.

SBC alone explains why the GAAP loss persists. Ambarella expensed $98M of it in FY2026, the full gap between the headline loss and adjusted profit.

SBC is also falling. It was $111M in FY2024 and $108M in FY2025. As a share of revenue, it has dropped from 49% to 25% over three years. Q1 FY2027 came in at $21.9M, down 16% y/y, putting the run-rate near $88M for the full year.

That trend decides when Ambarella turns GAAP-profitable. If research spending stays roughly flat (Q1 R&D was actually down slightly y/y) while revenue continues to grow, the company plausibly reaches a GAAP operating breakeven around FY2028-2029, somewhere near $550-580M of revenue.

Adjusted profit and FCF are already here, and Ambarella has now run positive FCF for 17 straight years on an annual basis (Q1 FY2027 was a -$29.6M outflow driven by the inventory build, which doesn't break the annual streak).

The CNN-to-transformer upgrade is the base case. Around it, Ambarella has started signing a new kind of contract that didn't exist in the model six months ago.

Management calls them Long-Term Agreements, or LTAs. They are multi-year customer commitments that cover more than one chip generation, so a customer signs up for the current product and its successors.

Each is structured around volume and pricing over five or more years. Two have now been disclosed.

The first LTA is for CV8, Ambarella's first 2-nanometer semi-custom chip. The design was finalized and sent to the foundry in January 2026, with the (undisclosed) customer funding part of the engineering.

CV8 is targeted at IoT endpoint applications with complex AI workloads. CV8 starts production in the first half of FY2028. Ambarella retains the right to sell the same design as a standard product into other markets.

The second LTA, announced on May 28, 2026, is with Hanwha Group, the Korean industrial conglomerate with more than $60B in annual revenue across aerospace, defense, robotics, physical security, life sciences, industrial automation, and other categories.

The agreement runs 10+ years and is valued at more than $800M in potential revenue. It covers Hanwha's video security, robotics, industrial automation, and life sciences product lines, structured as co-development across multiple chip generations.

Management calls it "one of the first agreements of its kind in the edge AI semiconductor market."

The Hanwha LTA implies $80M per year, against an FY2027 guide midpoint of $440M. This lays a multi-year revenue floor under the industrial, robotics, and life sciences applications Ambarella has been describing as future markets.

Management also said it's in discussion on additional LTAs beyond these two, without specifying. LTAs typically come with non-recurring engineering (NRE) funding from the customer, so the model also reduces some of Ambarella's silicon and software development burden alongside locking in volume.

Beyond the LTAs, Ambarella's other bets are still earlier:

Ambarella has also locked in 2-nanometer manufacturing capacity at Samsung. On May 28, Samsung publicly named Ambarella and NVIDIA as its 2-nanometer process customers, a competitive signal in a market where leading-edge capacity is tight.

Ambarella's management keeps pointing at the same long-term prize, and it isn't cameras or cars. It's robots.

But the company is still early in this transition. Ambarella shipped its first warehouse-robot perception chip and ran its first full quarter of drone production revenue in FY2026.

A robot, whether it's a warehouse picker, a delivery drone, or a humanoid, needs exactly the chip Ambarella builds. It has to fuse data from several sensors, understand its surroundings, plan a move, and drive motors, all on the device, all on a battery.

Ambarella has multiple chip families pointed at this market:

Humanoid robots have moved from research demos to funded commercial programs in the past two years. But every one of them runs into the same problem the moment it tries to ship in volume: the prototype chip is too expensive and too power-hungry.

Most humanoid prototypes today run on NVIDIA's Jetson modules, including the new Jetson Thor platform that NVIDIA built specifically for physical AI. Jetson Thor is the most capable option on the market, but it's a high-power, premium-priced part.

Once a manufacturer is building thousands of units a year, cost per robot and battery life turn into the binding constraints.

That transition is exactly what Ambarella's design is built for. Its whole pitch—the most AI performance per watt—only starts paying off when a buyer is shipping real volume and the per-robot cost line decides the business. Ambarella also brings radar perception, which matters because robots, like cars, need to judge distance where cameras alone struggle.

On the Q1 FY2027 call, management disclosed 15+ robotic design wins, including aero drones, with lifetime revenue exceeding $100M and more than 30 customers in the robotics pipeline.

Moreover, the Hanwha LTA signed in May 2026 adds a contracted multi-year ramp across industrial robotics, industrial automation, and life sciences at one of Asia's largest industrial customers.

Fermi Wang also confirmed that some of the design wins are humanoid applications, though Ambarella isn't yet disclosing which.

Realized humanoid revenue is still a multi-year wait, but the design-win pipeline behind it is now visible.

Ambarella's biggest risk isn't in its products. It's in its geography.