OpenAI Reportedly Buys Glass Imaging for the Camera in Its First Hardware Device – AI Image Processing From an Ex-Apple Team

Industry NewsAI

by Nino LeitnerToday

Subscribe to News - Industry

OpenAI has reportedly bought Glass Imaging, the Californian startup behind the GlassAI neural image signal processor, in a deal said to value the company at more than $300 million. Neither side has confirmed the purchase, and OpenAI has not said what it wants with a team whose entire business is getting more real detail out of very small cameras. For anyone shooting on phones, drones, or pocket cameras, the question is what happens when an AI lab owns the step between the sensor and the file.

The report comes from The Wall Street Journal, which cites people familiar with the matter and says the purchase was completed quietly in recent months. TechCrunch reports that OpenAI did not immediately respond to a request for comment. Glass Imaging’s own journal carries no mention of a sale. As of this writing, more than a week after the report, neither company has published a statement.

The closest thing to an acknowledgment comes from the company’s early backers. David Stark of Ground Up Ventures, its first investor, wrote on LinkedIn that he could not confirm the details in the Journal’s report but looked forward to what the founders “will accomplish inside of OpenAI,” and partners at GV and Insight Partners posted similar farewells. Attar reshared them with thanks. The exit itself is not being disputed, then, while the price, the structure, and the fate of the team and its customers remain unconfirmed, and we treat the deal accordingly throughout.

Who is Glass Imaging?

Glass Imaging was founded in 2019 by Ziv Attar (CEO) and Tom Bishop (CTO), two former Apple engineers who, according to TechCrunch, led the team that developed the iPhone’s Portrait Mode. Attar had been through an imaging acquisition before: he co-founded LinX Imaging, the multi-aperture camera company Apple bought in 2015.

The company positioned itself as a supplier rather than a device maker. In its most recent funding announcement in 2025, Glass Imaging described its business as developing “licensable IP” for smartphones, drones, and wearables. That wording matters for what comes later in this article.

One trained network instead of a processing chain

A conventional image signal processor (ISP) turns sensor data into a picture through a sequence of separately tuned stages: demosaicing, noise reduction, sharpening, lens corrections, and multi-frame fusion. GlassAI replaces that sequence with a neural network trained for one specific camera module. According to the company, the network learns how that particular lens and sensor degrade an image, then reverses lens aberrations and sensor imperfections from bursts of RAW data, running on the device’s own neural processor.

Glass Imaging claims this can “boost camera performance 10x” and that results “remain true to life with no hallucinations or optical distortions.” Both are manufacturer claims. The first has no unit attached to it, and we are not aware of an independent measurement of the second. The practical consequence of the approach is easier to pin down: because each network is tailored to a module, every new sensor and lens combination requires new training work from whoever owns the tools.

It is not limited to stills

This is where the story becomes relevant beyond photography. At Qualcomm’s Snapdragon Summit last year, Glass Imaging demonstrated GlassAI Video on a Snapdragon 8 Elite Gen 5 reference device, processing 4K video in real time on the Hexagon NPU at 20x zoom or higher. In January, the company declared its RAW video neural ISP pipeline production-ready on the same platform, with a demo clip labeled 4K at 30fps.

The terminology needs care. “RAW video” here describes the input, not the recording format. GlassAI takes Quad-CFA RAW sensor data, bypasses the standard ISP and the usual remosaicing step, and outputs a finished RGB video stream with noise reduction, sharpening, and image finishing already applied, frame by frame and without a cloud connection. CTO Tom Bishop said the output comes “without generative artifacts.” The release gives no figures for latency or power consumption beyond describing both as acceptable.

The first shipping product followed in spring, when HONOR launched its 600 series with GlassAI handling zoom imaging. The phones are built around a 200-megapixel main camera, and the base model has no telephoto module at all, so its zoom range depends on cropping the main sensor and on how much detail the network can recover from that crop.

Why would an AI lab want a camera pipeline?

OpenAI has not answered that, so what follows is our reading. The company bought Jony Ive’s hardware startup io Products in 2025 in a deal valued at roughly $6.5 billion, and the two have been working on consumer devices since. A court filing in the iyO trademark dispute earlier this year stated that the first device will not ship to customers before the end of February 2027. In July, Bloomberg reported that it will be a portable, screenless speaker built as an AI companion for the home, equipped with a camera and sensors to read its surroundings, with an unveiling planned for this year and a launch in 2027. OpenAI itself has not described the device.

If that report holds, the camera will be small, and a vision model can only interpret what the optics and the processing deliver. A team that has spent seven years recovering detail from sub-par lenses and tiny pixels is an obvious fit for that problem, and The Wall Street Journal draws the same line to the hardware effort.

The Apple connection also has a legal backdrop. In July, Apple sued OpenAI, io Products, Chief Hardware Officer Tang Tan, and a former Apple engineer in the US District Court for the Northern District of California, alleging trade secret theft and breach of contract in connection with OpenAI’s hardware effort. The complaint states that more than 400 former Apple employees now work at OpenAI. OpenAI responded that it has “no interest in other companies’ trade secrets” and, according to Bloomberg, asked the court in August to dismiss the case as meritless. Nothing in that suit concerns Glass Imaging, whose founders left Apple before starting the company in 2019. It does mean that another group of former Apple camera engineers is arriving at a company that is already in court with Apple over how it built its hardware team.

We would be cautious about reading this as a generative video move. OpenAI shut down Sora in March, app and API included, after the copyright fight with Hollywood and the collapse of the Disney investment. GlassAI sits at the other end of the chain. It works on photons that actually reached a sensor, and its makers sell it on the promise that nothing is invented.

The pipeline is becoming the camera

Readers of our lab tests have already seen where this leads. When we put the iPhone 17 Pro through our standard procedure, we had to abort the dynamic range and latitude tests because the phone kept applying scene-dependent processing in full manual mode, in ProRes RAW and Apple Log 2. Our conclusion at the time was that the iPhone is “less a camera and more an image-generation pipeline that starts with a sensor.” By contrast, the DJI Osmo Pocket 4P behaved deterministically in manual mode and delivered 16 stops at SNR=2 from a 1-inch sensor, the best figure we have measured.

A neural ISP trained per module pushes further down the first of those two roads. When a single network handles demosaicing, denoising, and optical correction together, there is no individual stage left to switch off, and resolution or noise figures describe the training as much as the hardware. The results may well be accurate, but they are harder to verify from the outside, which is relevant for anyone who wants to cut small-camera footage against a cinema camera in the grade.

Provenance is the second open point. C2PA Content Credentials, which Sony extended to video last year, record which device wrote a file and what was edited afterwards. As far as we can see, they do not describe how much reconstruction happened before the file existed. Glass Imaging’s position is that its network restores captured information rather than generating it, and that distinction will carry a lot of weight as learned pipelines spread.

Then there is the licensing business. HONOR is a customer, and other manufacturers were the stated target. Whether OpenAI keeps licensing GlassAI to phone and drone makers or reserves the team for its own hardware is unknown, and no party has commented. Companies that planned future camera modules around the technology will want that answered first.

For more information, please visit the Glass Imaging website.

Would you trust a camera whose entire RAW-to-RGB conversion is a trained neural network, as long as the detail is real? Don’t hesitate to let us know in the comments below!