Nvidia is in early talks that could end with it putting more money into Reflection AI, supplying the startup with more computing power, or buying it outright, according to reports published on October 10 and 11. The Financial Times broke the story on Saturday, citing people with direct knowledge of the discussions, and Bloomberg separately reported the talks the following day through its own source. Neither company has confirmed anything. Nvidia did not respond to requests for comment, and Reflection declined to comment to Bloomberg.
Both outlets stress how preliminary the discussions are. Bloomberg's source said no terms had been reached. According to the FT, a deal could take several forms: a larger equity stake, an agreement to provide more chips, or an "acqui-hire" in which Nvidia would hire Reflection's staff and license its technology without acquiring the company itself. That last structure, the FT noted, could spare the parties a lengthy regulatory review. The paper said an agreement could come within weeks, while cautioning that the talks could still collapse. Reuters, which relayed the FT report, said it could not immediately verify it.
Nvidia is not a newcomer to Reflection's cap table. The FT reported that the chipmaker has already invested $800 million in the company, making it one of Reflection's main financial and strategic backers. Reflection was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. Its chief executive, Laskin, told CNBC in April that it was raising money at a $25 billion pre-money valuation, and the Wall Street Journal has reported talks to raise $2.5 billion at that level.
The timing follows Reflection's first big product moment. On October 5 the company introduced Beam, a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, aimed at coding, reasoning and agentic work. Reflection says it will publish Beam's weights under an Apache 2.0 license later this month, together with a technical report and model card, and that Beam matches China's GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. Those performance figures are the company's own and had not been independently verified when Beam was announced, and the weights are not yet public.
The strategic backdrop is the push for American open-weight models. The FT described Reflection as a developer the Trump administration hopes can rival cheap Chinese alternatives, and Bloomberg pointed to the administration's 2025 AI action plan, which argued that open models founded on American values could become global standards. Nvidia chief executive Jensen Huang has been a vocal supporter of open-weight models, and the company publishes its own Nemotron family.
A tie-up would also extend a run of large Nvidia deals. Bloomberg cited a $13 billion agreement to buy Hugging Face in September and a licensing pact with chip startup Groq worth roughly $20 billion. Taking a larger role at Reflection would put Nvidia more firmly on both sides of the open-model market: selling the hardware used to train and serve models while backing, or owning, one of the teams building them.
For now, the story rests on anonymous sourcing from two news organizations, with no statement from either company and no disclosed price or structure. The range of outcomes on the table, from a computing deal to a full acquisition, means the eventual transaction, if any, could look very different from the headlines.