Nvidia has already written one large check to Reflection AI — roughly $800 million to back a two-year-old lab that builds open models for automated programming. Now the chipmaker is weighing whether to buy the rest of it.
The Financial Times reported on October 10 that Nvidia is in talks to deepen its stake, with options that include adding equity, supplying computing power, or an outright acquisition. Bloomberg confirmed the contact the same day, citing people familiar with the matter. The discussions are at an early stage, and no terms have been reached.
A deal, if one happens, could take the shape of an acqui-hire. Nvidia would absorb the core team and license the technology rather than merge two companies outright. That structure would let it skip the long antitrust review that a full purchase of a $25 billion lab would likely draw.
Reflection has moved fast even by the standards of the AI boom. Misha Laskin and Ioannis Antonoglou, two former Google DeepMind researchers, founded the Brooklyn-based company in 2024 to automate software development. It emerged from stealth in March 2025 with $130 million in financing, then raised $2 billion in October 2025 in a round that valued it at $8 billion. By June 2026 that figure had climbed to $25 billion, a tripling in less than a year.
The lab’s first product, Asimov, was a code-comprehension agent that read source code, email, Slack messages and documentation to answer questions about how a software system actually works. The broader bet has always been the same: that models trained with heavy reinforcement learning can take over the tedious parts of building software, from writing code to shipping it.
Last week the company released Beam, its first open-weight model. The 501-billion-parameter system, with 23 billion active, is aimed at coding, reasoning and agentic tasks at low token cost. The pitch is that open weights can do frontier work more cheaply than the closed labs that now dominate the market, and hold their own against the Chinese open models that have undercut Western prices.
Reflection’s rise tracks a wider shift in the market. Open-weight models from Chinese labs have pushed prices down and forced Western companies to answer with their own free releases, a dynamic that has made open weights central to the AI story rather than a niche. Reflection has positioned itself as the Western answer to that trend, a lab that gives away the weights and charges for the enterprise access around them.
The deepening tie is a bet for both sides. Nvidia gets closer to a lab whose open models generate demand for its chips — Reflection signed a $6.3 billion compute deal with SpaceX in June for Nvidia GB300 chips at the Colossus 2 data center — while Reflection gains the money and silicon it needs to keep training at scale.
For Nvidia, the appeal runs through the economics of open weights. Every company that downloads an open model still has to run it somewhere, and that somewhere is usually a cluster of Nvidia chips. A thriving open-model ecosystem is, in effect, a demand engine for the company’s hardware, one it does not have to build itself. Nvidia has backed the closed labs too, anchoring OpenAI’s record financing round, so its interest in Reflection reads less as picking a side than as being present wherever the models run.
Owning the lab outright would give Nvidia a stake in the open-source side of an industry argument it has so far profited from without taking sides. The chipmaker sells to everyone — closed and open, American and Chinese — and has been careful to stay in that middle lane. Buying Reflection would be a statement about which direction it expects the market to move.
The acquisitive instinct cuts against Nvidia’s usual posture. The company has spent the past two years insisting it is a neutral supplier of the AI build-out, an argument that has kept it out of the crosshairs that have found cloud providers and model makers. Taking control of a frontier lab would complicate that story, which is part of why the acqui-hire structure, with its lighter regulatory footprint, is on the table.
The talks could still end with nothing more than a larger equity position. Nvidia’s $800 million already sits on the company’s books, and a follow-on investment would deepen the tie without the burden of ownership. What is less ambiguous is the direction of travel: the supplier of the AI boom is moving from customer to owner, one open-model lab at a time.
The open question is what Nvidia actually wants to become. Running a model maker means taking on the safety debates, the scrutiny and the capital demands that come with owning a frontier lab — a very different job from selling chips to every lab in the field. Whether Nvidia is ready to step into that role, and whether regulators would let it, is the part of the deal no term sheet answers. For now, the talks are one more sign of how fast the middle ground in AI is shrinking, and of who is willing to pay to stay on it.
