Nvidia weighs buying the lab it already funds

Nvidia is in talks to deepen its investment in Reflection AI, the open-weight model developer it already backs, or to acquire it outright, the Financial Times reported on Saturday, citing people with direct knowledge of the matter. Reuters, which relayed the report, said it could not immediately verify it. Nvidia and Reflection did not immediately respond to requests for comment outside business hours, and neither had commented on the record when Reuters published.

According to Reuters' account of the FT report, the talks are at an early stage and a deal could take several forms. One is a so-called acqui-hire, in which Nvidia would hire Reflection's staff and license its technology rather than buy the company, a route the FT said could potentially avoid a lengthy regulatory review. The FT said an agreement could come in the coming weeks, and that the discussions could still fall apart. The Next Web, summarising the same report, said other options include supplying the startup with more chips or raising Nvidia's stake.

The FT reported that Nvidia has already invested $800m in Reflection. Nvidia was among the investors in the startup's $2bn round at an $8bn valuation in October 2025, and TechCrunch reported, citing PitchBook, that Reflection has raised roughly $4.7bn in total from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners. By our arithmetic, $800m is about 17 per cent of that total, though the two figures come from different sources and neither company has confirmed either.

The valuation is less settled. Reuters noted that Misha Laskin, Reflection's chief executive, told CNBC in April that the company was raising fresh capital at a pre-money valuation of $25bn. TechCrunch, writing on 5 October, described $25bn as the valuation of the company's last round. The reports we reviewed do not say whether that round has closed, or what price Nvidia might pay.

Reflection, founded in 2024 by the former Google DeepMind researchers Laskin and Ioannis Antonoglou, unveiled what it calls its first open-weight model, Beam, on Monday 5 October, and said it would publish the weights and technical details later in October. As Komposite reported the next day, the weights were not yet available. Per TechCrunch, Beam is a text-only mixture-of-experts model with 501bn total parameters and 23bn active. The company says it scores on par with Z.ai's GLM-5.2 on advanced reasoning benchmarks using "3-4x less inference compute". TechCrunch noted that Reflection's performance claims have not been independently verified.

Nvidia is also tied to Reflection through compute. TechCrunch reported that Reflection signed compute deals this summer with SpaceX and Nebius, collectively worth more than $7bn, to secure access to Nvidia's GB300 chips through 2029. Reflection is pitching Beam to enterprises and sovereign customers as the basis for local "AI factories", an idea TechCrunch noted Nvidia chief executive Jensen Huang has long championed and that would run on Nvidia GPUs.

The structure matters because of Nvidia's recent record. TNW reported that Nvidia agreed in September to buy Hugging Face for $12.93bn, a deal that has yet to close. Nvidia also has a licensing deal with the chip startup Groq, after which Groq's founder and chief executive, Jonathan Ross, and its chief operating officer, Sunny Madra, moved to Nvidia, as Komposite reported in September. TNW, citing the New York Times, reported that the Justice Department is examining whether the Groq deal was structured to avoid antitrust review. An Nvidia spokesperson, quoted by TNW, called the Groq story "a prime example of the American system working as designed". Nothing in the reports on Reflection says any regulator has looked at these talks.

The reports give no price, no stake size and no timetable beyond "the coming weeks". We have relied on Reuters' and TNW's summaries of the FT's reporting, not the FT's own text. Any hire-and-license arrangement would also depend on terms that have not been published.

For Reflection, the open question is what it would mean to be tied more closely to the company whose chips its compute deals secure. TechCrunch described its pitch as a Western answer to Chinese open models such as DeepSeek, Qwen and Z.ai, which rests partly on the company being an independent lab. Komposite has tracked the capital flowing to model developers, including TypeSafe AI's $870m raise at a $7.5bn valuation, and will report further if Nvidia or Reflection confirms the talks.