Nvidia is in talks to deepen its investment in Reflection AI or buy the startup outright. The Financial Times reported this on Saturday, citing people with direct knowledge. Reuters picked it up and said plainly that it could not verify the report; neither Nvidia nor Reflection responded over the weekend.
The talks are early. They could still fall apart.
Three shapes, and one of them skips the antitrust review
Per the FT, several structures are in play. One is an acqui-hire: Nvidia hires the staff and licenses the technology rather than acquiring the company. That would avoid a lengthy regulatory review.
A deal could land in the coming weeks. Nvidia is already inside Reflection anyway, to the tune of $800 million.
Reflection is two years old
The company was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both previously at DeepMind. It builds tools that automate software development. Laskin told CNBC in April that it was raising at a $25 billion pre-money valuation.
On Monday, Reflection shipped its first open-weight model — Beam, 501 billion parameters. The pitch was to compete on coding and agentic work against the cheap Chinese models, against DeepSeek and Kimi.
Five days later its biggest backer is discussing a takeover. That’s a conspicuous pace.
The compute partner is the same one
Reflection is also the company that signed a $6.3 billion compute deal with SpaceX in June. So whoever buys Reflection buys a model, an engineering team, and an infrastructure position.
Nvidia would own the open answer
Nvidia earns on every model that runs on its chips, so in theory it shouldn’t care which lab wins. Open weights are different: if the cheap Chinese models define the open end of the market, demand drifts somewhere Nvidia has less say.
Beam is built against exactly that. Which is why the talks read less like buying technology and more like refusing to let the American answer to DeepSeek slip away. An acqui-hire would be the fastest version of that — and the one that raises the fewest questions in Washington.