Nvidia and Reflection AI: talks to buy the open-model start-up it already backs

Nvidia is in early talks to buy Reflection AI, or to invest more in it, the Financial Times reported at the weekend. Reflection is an American start-up that publishes the weights of its AI models, so anyone can download, change and run them. A deal could come within weeks, but the talks could still fall apart, the newspaper’s sources said, according to Reuters. Neither company has commented.

Nvidia already has a large stake. It has invested $800m in Reflection, the FT said. The start-up was founded in 2024 by Misha Laskin and Ioannis Antonoglou, two former researchers at Google DeepMind. Laskin told CNBC in April that it was raising money at a pre-money valuation of $25bn. On 5 October it launched Beam, its first open-weight model, which is aimed at coding and agent tasks and competes with cheaper Chinese models such as DeepSeek and Kimi.

Nvidia and Reflection AI: a purchase that may not look like one

The talks cover several kinds of deal. One is an acqui-hire, in which Nvidia would hire Reflection’s staff and license its technology without buying the company itself. That route would probably avoid a long review by competition regulators, the FT reported. The other options are a larger stake, or more chips and computing power for Reflection, or both.

The choice matters because of Nvidia’s size. At Monday’s close of $226.32 on Nasdaq, its 24.1bn shares were worth about $5.5tn, by our calculation. A full purchase of a model maker by the company that sells the chips such models run on would probably face a long review, which is why the acqui-hire is on the list. It gives Nvidia the people and the technology while leaving the company, in law, independent.

Nvidia’s growing stake in open models

Reflection would not be Nvidia’s first deal of this kind. On 2 September Nvidia agreed to buy Hugging Face, which runs a platform for sharing and deploying open-source models and datasets. It will pay about $11.9bn to Hugging Face’s shareholders, plus a retention programme of up to $1.0bn for staff who join, according to its filing with the SEC. Nvidia promised to keep the platform open, including to chips from other vendors.

The same filing explains why open models matter to Nvidia. Demand for them “promotes the use of our products worldwide”, it says. It also names a risk. Many of the most popular open-source models “originated in China”, the filing says, and any rule that limits support for models from one region, China included, could have a material effect on Nvidia’s business.

The logic is simple. Every open model that becomes popular needs chips to train it and more chips to run it, wherever it is downloaded, and Nvidia sells those chips. A strong American open model also gives developers an alternative to the Chinese ones, which protects Nvidia if Washington ever restricts them. American labs have been making that case for some time: as we reported on 2 October, OpenAI and Anthropic both warned about cheap Chinese models, while saying they oppose a ban.

Washington wants American open models

The White House has backed the idea. “We need to ensure America has leading open models founded on American values,” says America’s AI Action Plan, published in July 2025. The plan argues that open models have “geostrategic value”, because they could become global standards in business and research, and says the federal government should create “a supportive environment” for them.

That gives Nvidia a political argument for a deal. Buying or backing the leading American open-weight lab can be presented as a contribution to a national goal, rather than as the market leader buying one of its customers.

A chipmaker that invests in its customers

The talks fit a wider pattern. At 26 July, Nvidia had equity investments of $99bn and commitments to invest a further $25bn, its latest quarterly report shows. The commitments include stakes in “AI model makers, infrastructure financiers, and other private companies”. Its holdings in unlisted companies alone rose from $22.3bn at the end of January to $51.2bn six months later.

The same report sets out the catch. Future demand for Nvidia’s chips “may be affected by the ability of customers and partners, including those developing open models, to generate revenue”. An open model is given away, so its maker must earn money in other ways, for example by selling services to companies and governments. Beam is eight days old, and Reflection has not published any revenue.

How far Nvidia now stands behind the companies that buy its chips is the subject of our lead story today.

What to watch

The FT’s sources expect a decision within weeks. If Nvidia chooses an acqui-hire, the deal may never come before a regulator. Nvidia’s next quarterly report will show how much more of its money has gone into the companies that depend on its chips, and whether the Reflection stake is in it.