Report: The Chipmaker Eyes the Open-Model Lab It Already Backs

Nvidia, whose chips power most of the AI industry, is in early-stage talks to buy Reflection AI, a US startup that builds open models, the Financial Times reported on Saturday, citing people with direct knowledge of the matter. ^[1][6]

The story is unconfirmed. Neither Nvidia nor Reflection responded to the FT's requests for comment, and Reuters said it could not independently verify the report. ^[1][6] The talks could still collapse. An agreement could come in the coming weeks, according to the report. ^[1][6]

If it happens, it puts a model builder under the roof of the company that sells the hardware every other model builder depends on. That raises an awkward question about what "open" means when the chip vendor owns the shop.

Four possible shapes

The FT says the talks could produce a full acquisition, an additional equity investment, a deal to supply more chips or computing power, or an acqui-hire. ^[1][6]

An acqui-hire is a deal in which a company hires another firm's people and licenses its technology, rather than buying the firm outright. The structure could let the deal avoid a lengthy regulatory review, according to the report. ^[1][6]

Nvidia has used it before. In December 2025 it paid about $20 billion for a deal with Groq, a startup whose technology runs finished AI models to answer users' questions, a job the industry calls inference. Nvidia hired Groq's founder Jonathan Ross and other employees while Groq stayed an independent company. ^[1]

Nvidia is already a major backer of Reflection. It has put in $800 million, and the startup was most recently valued at $25 billion. ^[1][6] Reflection's other investors include Sequoia Capital, DST Global, Lightspeed Venture Partners and 1789 Capital. ^[2]

What Reflection is

Reflection was founded in 2024 by two former Google DeepMind researchers, Misha Laskin, its chief executive, and Ioannis Antonoglou. ^[1][2] It has about 320 employees in New York, San Francisco and London. ^[2]

It has cast itself as the American answer to Chinese open-weight labs such as DeepSeek, Z.ai and Alibaba's Qwen. Its first model, Beam, launched on October 5, five days before the FT report. ^[2][3]

Beam has 501 billion parameters, the adjustable numbers a model tunes during training. It is built as a mixture of experts, which means only 23 billion of those parameters are active at any moment, fewer than one in 20. That makes it cheaper to run than its size suggests. ^[3]

Reflection says Beam is competitive with Chinese open models such as GLM 5.2 and approaches Qwen 3.8-Max on coding and agentic benchmarks, meaning tests of tasks where the AI plans and carries out several steps on its own. Those results are the company's own and have not been independently checked. ^[3]

There is a plain irony in the training bill. Reflection trained Beam on 10,500 Nvidia GB300 chips over four weeks. ^[3] Its biggest supplier may be about to become its owner.

Reflection also has compute contracts with SpaceXAI, Nebius Group and Shinsegae, and is a member of the US Department of Energy's Genesis Mission, supplying models to the national laboratories. ^[2][4]

What "open" means here

Reflection describes its philosophy as "open intelligence": releasing model weights, publishing research including technical reports, and open-sourcing software tools. ^[2][4] Weights are the learned numbers inside a model. If you have them, you can run your own copy on your own machines.

That is not what the Open Source Initiative means by open. Its Open Source AI Definition, published in 2024, also asks for the training code and enough information about the training data. ^[2][4] By that test, most "open" releases, Reflection's included, are better called open-weight. ^[2][4]

The label has already been stretched thin. A buyer with the power to set the terms on which a model is shared stretches it further.

Reflection says it will release Beam's weights, a technical report and developer tools later this month. ^[3] Whether that survives a takeover is one of the questions the report leaves open.

Nvidia's record

Huang has spoken for open weights in public. He co-authored an open letter on their importance to the AI economy. ^[5][6]

The company has a decent record in practice. Its September deal for Hugging Face, the repository where developers share and find models, came with a promise from Huang that the site would remain an open platform. "Nvidia compute will not be required to build on or deploy through Hugging Face," he wrote. ^[5] The price was $12.93 billion. ^[5]

But that promise covers a neutral meeting place. Reflection is not a meeting place. It competes with OpenAI and Anthropic, both of them Nvidia customers. A chipmaker that owns the hardware, the biggest model repository and a model lab will be asked to explain how it stays neutral between them.

That is this paper's view, not a finding. The Hugging Face pledge was easy to keep because a repository serves everyone. A model lab has its own roadmap, and one day it will have to choose which of its models to release and which to keep.

Huang's letter says what he wants the ecosystem to look like. A purchase would show what he does when he owns a piece of it.

The acqui-hire route shows the tension. It might be the quickest path and the least scrutinised. It might also leave Reflection as a shell with a licence agreement and fewer of the researchers who made Beam. An open lab that has lost its people is open in name only.

What to watch

Three things will tell. The first is whether either company confirms or denies the talks. The second is whether Reflection publishes Beam's weights, the technical report and tools this month, as it has said it will. The third is the structure of any deal, because a full purchase invites antitrust questions that an acqui-hire is designed to dodge. ^[1][3][6]

The FT says a deal could be reached within weeks. It could also fall apart. ^[1][6]