Computer vision.
Seeing gets cheap.

Meta gives away a model that can find and track almost anything in an image or video when you describe it in plain words. Anyone can download it. Meanwhile Cognex, which sells cameras that spot defects on production lines, made a 71.5% gross margin last quarter.

Either Cognex's customers are paying for something Meta now gives away, or they're paying for something else. Working out which is the whole game for computer vision stocks in 2027.

It helps to remember what they've been paying for. Computer vision (machine vision, when it's on a factory floor) is software that turns pixels into decisions. A camera on a packaging line spots a crooked label and kicks the box off the belt. A chip behind a windshield sees a pedestrian and hits the brakes. For most of the field's history, every one of those jobs needed its own custom-trained model and an engineer to babysit it. The industrial vision companies sold exactly that, and they've been very good at charging for it. Now general-purpose models are starting to do a lot of the job out of the box.

You wouldn't know it from the incumbent computer vision stocks. Keyence ran a 54% operating margin last quarter, and Cognex stock is up about 80% this year. At the same time, private investors are paying up for the general-purpose "robot brains" that make seeing cheap.

Cheap seeing doesn't make computer vision a worse business. It moves the money. Here's where it's going, and how these computer vision stocks are grouped: companies that sell a result someone stands behind, the chips the free models have to run on, the hardware eyes no one can download, and the private labs building the brains.

When seeing is free, customers still pay for a result that works every time and someone to blame when it doesn't.

Keyence TSE: 6861

Keyence doesn't really sell cameras. It sells solved problems. Its own engineers walk onto the factory floor, figure out the inspection or measurement job, and leave with a purchase order, with no distributors in between.

That model is why a hardware company runs margins most software companies would envy: a 54% operating margin and an 85% gross margin last quarter. A free model doesn't threaten that much, because the customer was never paying for the model. They were paying for the line to stop throwing out good parts.

Cognex NASDAQ: CGNX

Cognex has decided not to fight the commodity. This year it launched one AI vision system built on Nvidia silicon and another on Qualcomm, running the models on the camera itself.

That's the right instinct. Cognex's value was never the pattern-matching code. It's the decades of knowing what a good barcode read or a clean solder joint looks like on a line running 24 hours a day, and the support team that shows up when it fails. Riding everyone else's chips lets it put that know-how on cheaper, better hardware every year.

The results so far back it up. Revenue hit a record $291 million last quarter with an adjusted EBITDA margin of 32.2%, up more than 11 points in a year.

Mobileye NASDAQ: MBLY

Mobileye shipped 10 million driver-assistance systems last quarter at about $48.50 each. A carmaker isn't buying a model there. It's buying a vision system that has been in production cars for nearly two decades, with a supplier standing behind it when a regulator comes asking.

The market hates the stock anyway. It's down nearly 30% this year, the company is taking a $3.8 billion goodwill impairment, and founder Amnon Shashua plans to step aside as CEO once a successor is found. Mobileye is a turnaround in sentiment more than in the business. The business still guides to about $2 billion of revenue this year, and its chips go into roughly 40 million new cars a year.

Free models still have to run somewhere. Inside a robot, a camera or a car, the chip decides what's possible.

Nvidia NASDAQ: NVDA

Nvidia has made cheap seeing its business model. It gives away Cosmos, its family of open models for training robots on synthetic video, and sells the Jetson Thor chips those robots run on.

Its edge computing segment, which covers robotics and cars, did $7.2 billion last quarter, up 27% from a year earlier. That's small next to the data center. But every free model that makes a robot easier to build is a Jetson sale waiting to happen, and Nvidia's new cheaper modules, aimed at "visual AI agents" and warehouse robots, arrive early next year.

Ambarella NASDAQ: AMBA

Ambarella makes low-power chips that run vision AI inside security cameras, trucks and robots. When the models are free, what matters is how many watts it takes to run them on a device with no fan and no data center behind it. That's the problem Ambarella has spent a decade solving.

The market is starting to pay for it. Hanwha signed a deal in May worth more than $800 million over ten-plus years, and revenue grew 13% last quarter to $108 million. In September it announced a partnership to run Ultralytics' popular open-source YOLO vision models on its chips, which is the thesis in one press release. The Financial Times reported in July that NXP was in takeover talks. Nothing has been announced.

Software can be copied. Sensors can't. These companies make the physical part of seeing, and they're the computer vision stocks a free download can't replace.

Sony Group NYSE: SONY

Sony Group makes the image sensors inside most of the world's premium phones, and that market has stopped growing. The next customer is the machine.

Sony and TSMC signed a preliminary agreement in May to work together on image sensors. Nikkei reported in August that the venture would cost about $6.3 billion, be based at Sony's Kumamoto site in Japan, and aim its sensors at robots and vehicles, with mass production as early as 2029. Sony hasn't confirmed those terms. For a company that's also a video game, music and movie business, it's a slow way to own the theme. It's also the company the rest of this list buys from.

Teledyne Technologies NYSE: TDY

Teledyne Technologies owns FLIR, the thermal-camera maker, and DALSA, a machine-vision camera business. It's here for the thermal side. A free model can label anything in a picture, but it can't see heat, and it can't see in the dark.

Thermal and infrared imaging are what militaries, border agencies and industrial inspectors need when ordinary cameras go blind. Teledyne's digital imaging sales grew 12.7% last quarter to $869 million, led by infrared and defense.

Basler XETRA: BSL

Basler is a German maker of industrial cameras, and it's on the list because nothing else shows camera demand so directly. Orders jumped 59% in the first half, and it has raised its 2026 guidance twice.

That signal works both ways. Management estimates 10–15% of second-quarter orders were pulled forward by customers worried about lead times, and the full-year guidance implies a much softer second half. This is a cyclical small cap. Buying it is a bet that demand for cameras outlasts the restocking.

These are the labs making general-purpose seeing cheap, and private investors are betting most of the value ends up with them.

Skild AI (Private)

Skild AI is building one foundation model that can control any robot body, whether it's an arm, a humanoid or a four-legged machine. It raised $1.4 billion in January in a round led by SoftBank, at a valuation above $14 billion. If one brain can drive every body, the robot maker's job shrinks to building cheap hardware.

Physical Intelligence (Private)

Physical Intelligence builds vision-language-action models. One model watches the scene, reads an instruction and moves the robot, with no separate perception system bolted on. It was valued at $5.6 billion in late 2025, and Bloomberg reported in March that it was raising at $11 billion.

That's the most direct threat on this list to anyone selling perception as a separate product. If seeing becomes one layer inside a model that also acts, nobody buys it on its own.

World Labs (Private)

World Labs is Fei-Fei Li's company. She built ImageNet, the dataset that kicked off modern computer vision. Her bet now is that machines need to understand space, including depth, geometry and what's behind the box, not just label what's in front of them. World Labs raised $1 billion in February from backers including Nvidia, AMD and Autodesk. Its first product, Marble, turns images and text into 3D worlds you can edit.

Free models won't mean fewer cameras. They'll mean far more of them.

When seeing costs a few watts and an open-source download, every robot, forklift, shelf and doorbell gets eyes, and a humanoid needs several pairs. The software scales at no extra cost. The sensors don't. Image sensors come out of a small number of specialized fabs, and the biggest supplier by far is concentrated in southern Japan.

This summer an earthquake hit Kumamoto, and Sony had to pause production at a sensor plant there. Within days, Basler was warning its own camera output would be disrupted into the fall.

That's a preview of the next few years. Intelligence is getting abundant, and the physical parts that catch light are not. The companies that look most like commodity suppliers today could end up holding the tightest supply chain in computer vision.