Nvidia is in early talks to acquire Reflection AI, the American open-weight model developer valued at roughly $25 billion, or to deepen its existing investment in the company, the Financial Times reported on Saturday, citing people with direct knowledge of the discussions. Reuters and Bloomberg quickly followed up on the FT's report, though all three outlets stressed that no agreement has been reached and the talks could still fall apart.
The discussions are notable less for the target than for what Nvidia is reportedly willing to consider: according to the FT, structures on the table include an outright acquisition, an expanded equity stake, a much larger chip-supply arrangement, or an acqui-hire in which Nvidia would license Reflection's technology and hire its staff without buying the company outright.
Why the Deal Structure Matters as Much as the Price
The acqui-hire option may sound like a technicality, but it could decide whether the deal happens at all. As AI Weekly's summary of the FT report notes, that structure would be the version least likely to draw a formal Federal Trade Commission review, since Nvidia would technically be hiring researchers and licensing code rather than completing a merger. Silicon Valley has been down this road before: the FTC scrutinized Microsoft's 2024 handling of Inflection AI's team, and the playbook for hiring a startup's people without buying its shares is now well understood.
Nvidia already knows Reflection well. The chipmaker holds a stake of roughly $800 million in the startup, which recently raised at a $25 billion pre-money valuation — a steep climb from the $2 billion round at an $8 billion valuation it announced in October 2025. The company was founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, is based in Brooklyn, and began life building autonomous coding agents before repositioning as a frontier lab for open-weight models.
Beam, the Model at the Center of the Talks
The timing of the report is no accident. Reflection shipped its first open-weight frontier model, Beam, earlier this week. The 501-billion-parameter model is the company's opening bid in the open-weight tier, with full weights and a technical report due under an Apache 2.0 license later this month.
The benchmarks Reflection published for Beam are honest about where the model stands: by the company's own numbers, it trails DeepSeek and Kimi K3 on several measures. What Beam offers instead is provenance — an American-built open-weight model that enterprises and governments uneasy about running Chinese-origin weights can adopt as a domestic alternative.
Training Beam was expensive. According to Startup Fortune, Reflection ran a four-week reinforcement learning session across 10,500 Nvidia GB300 GPUs, rented through a $6.3 billion multi-year compute deal at SpaceX's Colossus 2 facility that has the startup paying roughly $150 million a month through 2029. Every dollar of that compute spend flows back to Nvidia — which hints at the strategic logic of a deeper tie-up.
Washington Wants an Open-Weight Champion
The geopolitical layer is hard to miss. As our AI industry coverage has documented repeatedly over the past year, DeepSeek's rise rattled Washington because an open-weight Chinese model rivaled American closed models at a fraction of the training cost, exposing how thin the US moat in the open tier really was. The FT's report notes that President Donald Trump's administration hopes Reflection AI will rival cheap Chinese alternatives, and company officials have signaled exactly that expectation.
If Nvidia folds Reflection into its own roadmap, it would be buying more than a model team. It would be buying the argument that the open-weight future does not have to run through Beijing — and locking in a captive customer for its most expensive hardware.
Part of a Pattern of Nvidia Writing Checks
The Reflection talks are not an isolated move. Nvidia is separately in talks to invest in Perplexity at a valuation above $30 billion, up more than 50 percent from the roughly $20 billion mark the search startup set about a year ago, according to reports citing people familiar with the matter. The pattern is consistent: Nvidia keeps investing in companies that burn enormous amounts of its own compute, so each dollar it invests generates demand that eventually lands back on its own income statement.
Reflection is a different kind of bet, though. Perplexity sells a product; Reflection builds the foundation-model layer that OpenAI, Anthropic and Google already dominate — a layer where Nvidia has sold the shovels but never owned the mine.
What Happens Next
For now, caution is warranted. None of the three deal structures Bloomberg and the FT described is finalized, and early-stage talks can collapse or shrink into something far more modest than an acquisition — a bigger chip contract, say, or a plain equity top-up. Neither Nvidia nor Reflection AI has commented publicly on the report at the time of writing.
But the mere fact that full ownership is on the table, barely a year after Nvidia's first check, says something about how fast the ground has shifted under the open-weight category — and about how centrally Washington now sees it in the AI competition with China.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
