Nvidia in Talks to Buy or Back Reflection AI, FT Says

Nvidia is in early-stage talks either to put more money into Reflection AI or to buy the New York open-weight model lab outright, the Financial Times reported on Saturday, 10 October, citing people with direct knowledge of the matter. According to Reuters’ account of the report, an agreement could come within weeks, but the discussions could still fall apart. Reuters said it could not independently verify the story, and neither company commented on it.

The FT described several possible shapes for a deal. One is an acqui-hire, in which Nvidia would hire Reflection’s staff and license its technology rather than buy the company, a structure that could spare both sides a lengthy regulatory review. The others are less dramatic: supplying Reflection with more chips, or simply raising Nvidia’s stake. Bloomberg, reporting separately, described talks about a bigger investment or computing power and said no terms had been agreed, The Next Web noted.

A company Nvidia already funds

Nvidia is not a newcomer here. It has already invested $800 million in Reflection, according to the FT, making it one of the startup’s biggest backers. Former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou founded the company in 2024, and Laskin told CNBC in April that it was raising fresh capital at a $25 billion pre-money valuation, the latest price on record. No figure for an acquisition has been reported. Reflection has also lined up compute from Elon Musk’s SpaceX and the cloud provider Nebius, and the FT reported that the Trump administration hopes the lab can become an American rival to cheap Chinese open models such as DeepSeek.

The timing gives the talks an obvious anchor. On 5 October Reflection previewed Beam, its first open-weight model, which BitsMinds covered in detail: a 501-billion-parameter mixture-of-experts with 23 billion active parameters per token, aimed at coding and agents, with weights promised later in October under an Apache 2.0 licence. Beam was built on Nvidia silicon from end to end, with pretraining on 6,144 GB300 GPUs and a reinforcement-learning run on 10,500 more. Reflection’s own comparison table placed it behind the newest Chinese open models on raw scores and argued for it on efficiency instead.

Part of a pattern

A Reflection deal would extend a run of Nvidia moves into the model layer. In September it agreed to buy Hugging Face for $12.93 billion, and in August it agreed to pay Poolside $6 billion to license its model factory and hire the 109 engineers behind it, in a deal Poolside insisted was not an acquisition. It already publishes its own open models under the Nemotron name, as TIKR pointed out, and its earlier licensing deal with the chip startup Groq brought Groq’s chief executive and chief operating officer over to Nvidia.

That Groq deal is also the reason the acqui-hire option will draw attention. The New York Times reported in September that the Justice Department is examining whether the arrangement was structured to avoid antitrust review; an Nvidia spokesman said it showed “the American system working as designed”. Money is not the constraint: TIKR puts Nvidia’s free cash flow over the past four quarters at about $127 billion, so even a check at Reflection’s last reported valuation would be roughly ten weeks of cash generation.

For the moment there is no signed agreement, no price and no comment from either company, only a report that the two are talking. If a deal does land, its form will say as much as its size: a straight purchase would invite the review an acqui-hire is designed to avoid, while a bigger stake would leave the most prominent American open-weight lab independent on paper and even more dependent on its largest supplier.