Nvidia in early talks to buy or deepen stake in Reflection AI, valued at $25 billion

For most of the past three years Nvidia has been the company that sells the picks and shovels of the artificial intelligence boom, not the company that digs with them. Its data center graphics processors power nearly every large model training run of consequence, and its own open-weight model family, Nemotron, has been a modest side project rather than a strategic centerpiece. That posture is now being tested by a widening race over open-weight models, the freely downloadable systems that Chinese labs such as DeepSeek have used to undercut the pricing power of American labs.

Reflection AI is one of a small number of American startups built specifically to answer that challenge. The company was founded in March 2024 by Misha Laskin and Ioannis Antonoglou, both veterans of Google DeepMind, where Antonoglou worked on the Gemini research effort. Reflection pitches itself as a builder of frontier open-weight models, in contrast to closed labs that guard their weights, and it has assembled an unusual mix of backers, including Nvidia, Sequoia Capital, Lightspeed Venture Partners, DST Global, Disruptive and 1789 Capital, where Donald Trump Jr. is a partner.

The chipmaker is now in early talks to acquire Reflection AI outright, to increase its existing stake, or to strike an acqui-hire arrangement, the Financial Times reported on Saturday, October 10, citing people with direct knowledge of the matter. Nvidia has already put $800 million into the startup, making it one of its largest financial backers, and Reflection was valued at $25 billion in its most recent round. SiliconANGLE reported on October 11 that an agreement could come within weeks, while cautioning that the talks could still collapse.

Reflection stepped into public view five days before the report, when it unveiled Beam, its first open-weight model. Beam is a mixture-of-experts system with 501 billion total parameters pitched against lower cost Chinese open models such as DeepSeek and Kimi on coding and agentic work. Beam's training ran on Nvidia GB300 systems that Reflection rents from SpaceX, under a June agreement worth up to $6.3 billion, with payments of $150 million a month for capacity at SpaceX's Colossus 2 data center in Memphis through 2029.

Key Facts

The Financial Times reported on Saturday, October 10 that Nvidia is exploring several structures for a deal with Reflection AI. One is a full acquisition, which would be the cleanest route but the most likely to trigger a lengthy regulatory review. Another is an acqui-hire, in which Nvidia hires Reflection's staff and licenses its technology without taking control of the company. A third is a larger equity check, possibly accompanied by a bigger supply of chips and computing power. TechStartups reported on October 12 that the potential deal could give Nvidia something it has yet to build on its own: a leading American developer of advanced open-weight AI models.

Nvidia's investment history with the startup is substantial. The chipmaker led Reflection's $2 billion round in October 2025, when the company was valued at $8 billion, and it has committed $800 million in total. CEO Misha Laskin told CNBC in April that the startup's latest round closed at a $25 billion pre-money valuation, which means the price tag roughly tripled in under a year. StartupFortune reported on October 11 that Bloomberg subsequently confirmed the talks.

Beam's specifications are unusually detailed for a launch. The model carries 501 billion parameters with 23 billion active parameters, was trained on 23.8 trillion tokens, and supports a context window of 1 million tokens. Training involved a four week reinforcement learning run across 10,500 Nvidia GB300 GPUs. Reflection says Beam matches GLM-5.2 from the Chinese developer Z.ai on reasoning benchmarks while using three to four times less inference compute, though those claims have not been independently verified. Full weights and a technical report are due under an Apache 2.0 license later in October.

Nvidia's acquisition playbook is already visible. In December the company licensed technology from the inference chip startup Groq in a deal reported at $20 billion, with founder and CEO Jonathan Ross moving to the chipmaker. On September 3, Nvidia agreed to pay $12.9 billion for Hugging Face. Yahoo Finance reported on October 10 that a $25 billion purchase would equal roughly 20 percent of the $127 billion in free cash flow Nvidia generated over the past four quarters, or about 10 weeks of cash generation. At July 26, Nvidia held $56.6 billion in cash, cash equivalents and marketable debt securities.

Reflection's commercial footprint extends beyond its models. The company has established relationships with the Pentagon and the U.S. Department of Energy, and agreements to develop AI models for allies including South Korea. Nebius agreed in July to sell Reflection another $1 billion in AI capacity. Neither Nvidia nor Reflection has commented publicly on the talks.

Analysis

The most striking feature of this reported deal is how it would change Nvidia's position in the AI stack. Nvidia currently sells the hardware that every major lab depends on, a position that makes it agnostic to which model wins. Buying or effectively absorbing Reflection would push the chipmaker into direct competition with its own customers, including OpenAI, Anthropic and Google, at the model layer. That is a strategic shift with real commercial risk, and it explains why the acqui-hire structure is being floated so prominently.

What this really means is that Nvidia has concluded the open-weight layer is too important to leave to others. DeepSeek and other Chinese labs have shown that capable open models can be released cheaply and adopted widely, undercutting the pricing power of closed American systems. Washington wants a domestic answer, and Reflection has positioned itself as exactly that answer, with government relationships that a pure research lab would lack. Nvidia, as the most valuable company in the sector, is the only private actor with both the capital and the incentive to fund that answer at scale.

The financial mechanics are almost trivial for Nvidia. A $25 billion check represents about 20 percent of its trailing four quarter free cash flow, and the company held $56.6 billion in cash and marketable securities at July 26. Even a full purchase would leave roughly $31.6 billion on the balance sheet before other spending. The constraint is not money but regulation. An acqui-hire, the structure Nvidia used with Groq at a reported $20 billion, is the version of this deal least likely to trigger a Federal Trade Commission review, following the agency's scrutiny of Microsoft's handling of Inflection AI's team in 2024.

There is also a question about whether Reflection is the right asset at the right price. Its $25 billion valuation tripled from $8 billion in October 2025, a pace that reflects enthusiasm as much as revenue. Reflection's claimed parity with Z.ai's GLM-5.2 has not been independently established, and its compute costs are enormous: $150 million a month to SpaceX through 2029 under a deal worth up to $6.3 billion. Owning the company would give Nvidia both the model and the bills.

Why It Matters

If the talks produce a deal, Nvidia would become the first major American chip company to own a frontier open-weight model developer. That would put it directly against the Chinese labs Washington has spent two years trying to counter, and it would give the Trump administration a domestic champion in a market where cheap Chinese models have gained real traction. The White House has made clear it hopes American open-weight developers will rival those cheap alternatives, and Reflection's government work gives it a foothold that a purely commercial lab would struggle to match.

The structure of any agreement matters as much as the price. An acqui-hire would let Nvidia absorb Reflection's researchers and license its technology while avoiding the lengthy merger review that a full acquisition would invite. That template, first tested with Groq and now potentially with Reflection, is becoming a standard playbook for AI consolidation in the United States, and it raises questions about how much consolidation can occur outside the traditional antitrust process.

The deal also lands in a broader wave of AI investment. StartupFortune reported on October 11 that Nvidia is separately in talks to invest in Perplexity at a valuation above $30 billion, up more than 50 percent from roughly $20 billion about a year earlier, with Perplexity's annualized revenue reportedly climbing past $750 million. Nvidia is not merely selling chips into the boom. It is increasingly buying pieces of it.

Next Up

The immediate question is whether an agreement arrives within weeks, as people familiar with the matter told the Financial Times, or whether the talks collapse as the paper also warned they could. Either way, both companies have left the door open by declining to comment, and the reported interest from Nvidia has already reframed how the market values open-weight model developers.

Investors should also watch for the release of Beam's full weights and technical report later in October under an Apache 2.0 license, and for any sign of how Nvidia's existing Nemotron family would fit alongside a Reflection acquisition. If a deal closes, the next test will be whether the combined effort can hold its ground against DeepSeek and its Chinese peers on cost and capability, a race now measured in months rather than years.