Nvidia is circling Reflection AI

The Financial Times reported on Saturday, October 10, that Nvidia is talking with Reflection AI about a deeper tie-up, citing people with direct knowledge of the matter. The discussions are early and could take several forms: a full acquisition, a larger equity investment, or a so-called acqui-hire, where Nvidia would hire Reflection’s staff and license its technology rather than buying the company — a structure that could avoid a lengthy regulatory review.

The FT’s sources said an agreement could come in the coming weeks, and were equally clear the talks could collapse. Reflection declined to comment; Nvidia did not immediately respond. Separately, Bloomberg reported the discussions could also end with Nvidia supplying Reflection with more computing power instead of fresh capital.

Reflection was founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou and builds open-weight models aimed at coding and AI-agent work. On October 5 it launched its first one, Beam, a 501-billion-parameter model it claims matches GLM-5.2-level reasoning at 3–4x less inference compute, with weights promised under Apache 2.0 later this month.

The company is not cheap anymore. It raised $2 billion last October at an $8 billion valuation — Nvidia put in $800 million of that — and CEO Laskin told CNBC in April it was raising again at a $25 billion pre-money valuation. It has also locked up more than $1 billion of compute capacity from Nebius through 2029, plus capacity at SpaceX’s Colossus 2 data center.

Nvidia has been leaning hard into open weights. The FT notes the White House has been counting on U.S. open-weight models to rival China’s, and Nvidia’s incentives here are obvious: open models that anyone can download and run sell GPUs. Beam is positioned explicitly as a Western alternative to DeepSeek and Kimi in coding and agentic tasks.

Owning a credible open-weight lab — rather than waiting for one to show up — also settles a strategic question for Nvidia. If agents and coding models commoditize into downloadable weights, the company that makes the chips benefits from every lab that ships them. Buying the lab is just the faster route.

For now this is a report about talks, not a deal. The FT described early-stage discussions with several live shapes — buyout, acqui-hire, bigger investment, or more compute — and named no price. If it closes in any form, it would be one of the largest open-weight transactions yet, and a signal that the chip seller wants a seat on both sides of the model market.

Watch two things: whether the Apache 2.0 weights for Beam actually ship this month, and whether any deal keeps them free. A buyer whose business is selling GPUs has every reason to keep open weights open — but the people who built Beam haven’t commented on any of this yet.