Top Computer Vision Stocks 2026: How Robots See

Everyone is arguing about humanoid robots, robotaxis, and $14 billion "robot brain" startups. Meanwhile, the boring companies that sell the eyes for all of that are printing money.

Cognex just posted its highest revenue quarter ever. Keyence grew sales 33% last quarter at a 54% operating margin. Basler's order book jumped 59%. Ambarella crossed $100 million in quarterly revenue for the third straight time.

Computer vision is the part of AI that turns pixels into decisions. A camera sees a scratch on a phone screen and rejects it. A chip in a truck sees a pedestrian and brakes. A robot arm sees a box that's slightly rotated and grabs it anyway. Every one of those "sees" is a sale for someone on this list.^[1]

Two things flipped the demand curve this year:

1. AI moved to the edge. You don't want a warehouse robot pinging a data center to decide whether that's a shelf or a person. The vision has to run on-device, in milliseconds. That's a chip sale, a camera sale, and a software license, all at once.
2. Physical AI got real budgets. Robotics and physical AI startups raised a record $27.6 billion in 2025, and Nvidia opened CES 2026 by declaring the "ChatGPT moment for robotics." Whether or not you buy that, every robot those dollars build needs to see.

We're slicing this theme by where the seeing happens:

• Industrial Vision — the systems that inspect, count, and guide inside factories and warehouses. Proven, profitable, and accelerating.
• Vision Silicon — the chips that do the seeing on-device, in cars, cameras, and robots.
• Sensor Supplier — the picks-and-shovels layer that makes the actual eyeball.
• Private Bellwethers — the labs building "world models" and robot brains. Not tradable yet, but they set the pace.

This is where computer vision already pays rent. Factories and warehouses have used cameras for quality control for decades. What's new is AI: the same camera now handles messier jobs (odd lighting, weird packaging, robot guidance) that used to need a human. Three companies own most of this market, and all three are growing faster than they have in years.

Cognex NASDAQ: CGNX

If you've ever received a package that was scanned, sorted, and routed without a human touching it, there's a decent chance a Cognex camera did the looking.

Q2 2026 was the company's best quarter ever: $291 million in revenue, up 17%, with adjusted EBITDA margin at 32.2%, a jump of 1,150 basis points from a year ago. That's the eighth straight quarter of both revenue growth and margin expansion. Logistics has now grown double-digits for ten quarters running. Free cash flow rose 70%.

The strategic story is just as good. Cognex launched two AI vision systems this year, one built on Nvidia silicon and one on Qualcomm, that run inspection models on the camera itself instead of a bolted-on PC. New CEO Matt Moschner is also cutting dead weight: a Japanese trading business sold, roughly $22 million of low-margin revenue exited, and $35 to $40 million in annual cost savings targeted by year-end.

One caveat. Management flagged a lumpier second half: an electronics order pulled into Q2, a $30 million one-time partnership benefit from 2025 that won't repeat, and planned portfolio exits in Q4. The stock dipped on that. The underlying demand didn't.

Keyence TSE: 6861

Keyence is the company every other name on this list wishes it could be.

It sells sensors, measurement systems, and machine vision directly to factories through an army of engineers who show up, solve your problem, and leave with a purchase order. No distributors. No discounts. The result is an operating margin north of 50%, which is absurd for a company that ships physical hardware.

Fiscal 2026 (ended March) delivered ¥1.17 trillion in sales, up 10.4%, and the dividend jumped from ¥350 to ¥550 per share. Then the first quarter of fiscal 2027 came in with sales up 33% and operating margin at 54%. The stock popped 10% on the news.

The knock on Keyence is always the same: it's expensive, and it's a Tokyo listing that's fiddly to buy from a U.S. brokerage. But if you want one name in this theme you can hold for a decade and not check, this is it.

Basler XETRA: BSL

Basler is the small one. And small is where the leverage lives.

The German camera maker got crushed in the 2023–2024 industrial slump. It cut costs, held its organization flat, and waited. In 2026 the wait paid off: first-half orders up 59% to €180 million, revenue up 36% to €152 million, and EBIT up from €8.6 million to €31.1 million. Guidance was raised twice in one year. The German vision industry as a whole grew orders 25%; Basler more than doubled that.

The shares fell after the H1 report anyway, because management warned gross margins would ease from 51.6% to 47–49% in the second half, and admitted some orders may have been pulled forward by customers worried about lead times. Watch that. If second-half orders keep climbing, this becomes a very different stock.

The camera is dumb. The chip behind it does the seeing. This segment sells the processors that run vision AI inside cars, security cameras, and robots, at low enough power that the device doesn't melt. It's a nastier business than industrial vision (design wins take years, and you live and die by a handful of customers), but the winners get locked in for a decade.

Ambarella NASDAQ: AMBA

Ambarella started life making the video chips inside GoPros. Today it makes low-power system-on-chips that run vision AI at the edge, and it's mid-transformation into something bigger.

Revenue in the quarter ended April was $100.4 million, up 16.9%, the third straight quarter above $100 million. Automotive hit an all-time record on AI adoption in commercial vehicles. The company counts over 15 robotics design wins and guided next quarter to $105–111 million.

The number that matters: Ambarella signed a long-term agreement with Hanwha that management says could exceed $800 million over ten-plus years. That's a company with $400 million in annual revenue locking in one customer for two years' worth of sales. Expect more of these. The stock ran 40% after the report, so you're not early. But Ambarella is still losing money on a GAAP basis, and the first long-term deal is rarely the biggest one. One more wrinkle: NXP is reportedly in talks to buy the whole company (Bloomberg, July 31), and nothing is signed.

Next earnings: September 3.

Mobileye NASDAQ: MBLY

Mobileye shipped 10 million EyeQ chips last quarter. That's ten million cars that see the road because of a computer vision company in Jerusalem.

The stock has been a punching bag for two years, and the reasons were real: China share losses, a base-ADAS business that's mature, and the perpetual "when does the advanced stuff ramp?" question. Here's what changed in Q2 2026. Israel passed an R&D incentive law worth $180–200 million a year to Mobileye, which lifted guidance to roughly $2 billion in revenue. The company won a high-volume Stellantis program with more than double the per-unit profit of base ADAS. First-half revenue grew 13% while its customers' production fell 3%. And a U.S. robotaxi launch is penciled in for 2027, alongside a fresh acquisition (Mentee Robotics) that puts it in humanoids.

Two things to watch: Surround ADAS, the higher-priced product, isn't expected to move average selling prices until 2028, and founder Amnon Shashua said he plans to step down once a successor is found. Mobileye is cheap for a reason. It's also the only pure-play way to own automotive vision at scale.

Nvidia NASDAQ: NVDA

Yes, you already own Nvidia through every index fund on Earth. It's here because you can't write about robot eyes without it.

Nvidia is the layer everyone else's vision runs on. Jetson Thor delivers 7.5x the AI compute of its predecessor for on-robot inference, and the July launch of the cheaper T3000 and T2000 modules pushes that into mainstream robotics and vision AI. Cosmos, its family of "world models," generates the synthetic video that trains robots before they ever touch the real world. Cognex builds its top camera on Nvidia. World Labs took Nvidia money. Skild trained on Nvidia platforms.

At $81.6 billion in quarterly revenue, robotics is a rounding error today. But Nvidia now reports "Edge Computing" as its own market platform, covering robotics and automotive. When that line starts growing faster than data center, this list gets a lot more crowded.

Somebody has to make the actual eyeball. Image sensors are a picks-and-shovels business with real physics moats: pixel size, dynamic range, and the stacking tricks that let a sensor see in the dark.

Sony Group NYSE: SONY

Sony makes the image sensors in iPhones, and roughly half the world's smartphone cameras. That business plateaued, margins slid from 25% to just over 10%, and Samsung has been chipping away at the Apple account.

So Sony is repositioning. In August it firmed up a joint venture with TSMC, roughly $6.3 billion, to build next-generation sensors at its Kumamoto fab, with Sony holding 60%. The stated target beyond phones is "physical AI": sensors that help robots and vehicles recognize objects. Mass production is slated for 2029. Sony's CEO called it "the first step toward a fab-light model," and the company has reportedly explored spinning the semiconductor unit out entirely.

This is the slow, safe way to play the theme. You get games, music, and movies for free, plus a call option on machine eyes.

Here's the part of the market you can't buy, and it's arguably the most important. These labs are building the "brains" that decide what to do with what the cameras see. Their progress sets demand for everyone above.

World Labs (Private)

Fei-Fei Li built ImageNet, the dataset that kicked off the deep learning era. Her bet now is that language models have no real sense of geometry, gravity, or what's behind a wall, and that "spatial intelligence" is the next frontier. World Labs raised $1 billion in February from Nvidia, AMD, Autodesk ($200 million alone), and Fidelity, after Bloomberg reported talks at a $5 billion valuation. Its only shipped product is Marble, which turns images or text into editable 3D worlds. No IPO chatter yet. Watch for a robotics customer announcement; that's when this stops being a research project.

Skild AI (Private)

Two Carnegie Mellon professors, one "omni-bodied" foundation model that controls any robot, humanoid or arm or quadruped, without custom engineering. Skild AI raised $1.4 billion in January at a $14 billion valuation, the largest robotics AI round on record, co-led by SoftBank and Nvidia. It was worth $4.5 billion seven months earlier. That's the market telling you where it thinks robot value accrues: software. No IPO talk yet, but at this size the clock is ticking.

Physical Intelligence (Private)

The academic all-star team. Physical Intelligence builds the π-series vision-language-action models: one model that watches, reads instructions, and moves a robot. Confirmed valuation is $5.6 billion (November 2025); Bloomberg reported a round in talks at over $11 billion in March that hadn't closed by June. It's pre-revenue. If that round prints, the whole robot-brain category re-rates. If it quietly dies, that's the first crack in the physical AI narrative.

Failure mode #1: The brain eats the eye. If general-purpose vision models from Nvidia, Google, or the private labs get good enough, the "smarts" in a Cognex or Keyence camera become a commodity, and margins go with it. Watch whether Cognex's new AI systems hold pricing, and whether Skild-style platforms start shipping with their own perception stack bundled in.

Failure mode #2: The order book was a mirage. Basler admitted some 2026 orders came early from customers nervous about lead times. Cognex has a lumpy second half. If Q3 bookings roll over across the group, the "recovery" was a restock, not a cycle. Watch order intake, not revenue, this fall.

Honorable mention: China. Basler's mid-term guidance assumes the Chinese market stays open, and Chinese vision makers compete hard on price. A tariff or export shock lands here first.

Over the next 12–18 months, the theme goes from "AI is coming to the edge" to "show me the design wins." Catalysts, in rough order:

• September 3: Ambarella Q2 FY27. Look for a second long-term customer agreement, not just the numbers.
• October–November: Cognex Q3 and Basler Q3 order intake. This is the pull-forward test.
• Late 2026: Does Physical Intelligence's $11 billion round close? Does World Labs sign a robotics customer? These are the sentiment gauges for the whole space.
• January 2027: CES. Expect Nvidia to expand the Thor/Cosmos stack again and a wave of humanoid announcements, all of which spec cameras and vision chips from the companies above.
• 2027: Mobileye's planned U.S. robotaxi launch, and Stellantis Cloud-Enhanced ADAS beginning to roll across the fleet.
• By March 2027: Sony and TSMC finalize the sensor JV, with Japanese government support in discussion.

The humanoid gets the headlines. The companies above get the purchase orders.