Nvidia Reflection AI Talks Put $25B Open-Model Lab in Play

What happened

Nvidia, which has spent the past year buying talent and technology across the AI industry, is now in talks about Reflection AI, a startup it already funds. According to the Financial Times, which Reuters and Bloomberg later relayed, Nvidia is in early talks to buy Reflection AI or to put more money into the open-weight model developer.1714 The reports describe three possible outcomes. Nvidia could acquire the whole company. It could take a bigger stake, possibly with more chips and computing capacity attached. Or it could do an acqui-hire, hiring Reflection's staff and licensing its technology without taking ownership.11 People familiar with the talks said an agreement could arrive within weeks but that the discussions could still collapse.17 Reflection declined to comment, and Nvidia did not respond right away.17

Nvidia already has a large position in the company. The FT puts its total investment at $800 million, which makes it one of Reflection's biggest shareholders.1713 Accounts of that investment say Nvidia led Reflection's $2 billion round in October 2025, which valued the company at $8 billion. Sequoia, Lightspeed, DST and GIC also took part.12 Reflection has publicly confirmed the $2 billion raise and its list of investors. It has not confirmed the $800 million figure or the $8 billion valuation, so both should be read as press reports.7

The valuation math behind the talks

The deal may hinge on how fast Reflection's price has risen. In April, chief executive Misha Laskin told CNBC that the company was raising money at a $25 billion pre-money valuation.17 One market analysis citing StockAnalysis data puts the post-money figure at $27.5 billion after a $2.5 billion Series C. The same analysis says thin secondary trading on Hiive implied about $54.6 billion as of late July.1 If those numbers hold, Reflection's value roughly tripled in about six months, and secondary buyers have since priced it at about twice that again.131

Reports differ on some details. Several outlets date the $25 billion valuation to March.1364 Others tie it to the April round that Laskin confirmed on air.37 The likely explanation is that the Wall Street Journal reported the fundraising talks in March and the round closed in April.1 Estimates of total funding also vary a lot. One European outlet says Reflection has raised close to $2.6 billion.16 A profile cited in another analysis puts the total at about $4.63 billion.1 The gap probably comes down to whether the Series C money is counted. Either way, readers should not treat any headline total as audited.

The point several outlets stress, and the one that matters most, is that $25 billion is a fundraising valuation, not a price Nvidia has agreed to pay.43 No price has been reported.1 For a sense of scale, a purchase at the $27.5 billion post-money value would equal about half a percent of Nvidia's roughly $5.54 trillion market capitalization.1 Nvidia has about $99.9 billion in cash and short-term investments, so it can afford the deal.13 Affordability is not the real question.

Why an acqui-hire is the most likely structure

My reading is that the acqui-hire is the most likely outcome, because Nvidia has used it before. In December 2025, Nvidia took a non-exclusive license to Groq's inference technology and hired its founder, Jonathan Ross, and president, Sunny Madra. Groq stayed independent under a new chief executive.12 CNBC reported the deal at about $20 billion, compared with a prior Groq valuation of roughly $7 billion.12 Nvidia reportedly did something similar with Poolside's engineers in August.12 One regional report goes further and calls the acqui-hire the most likely option because it avoids regulatory review.13

The reports make the same regulatory point: according to the FT, a hire-and-license structure would probably avoid a lengthy merger review.17 But that is not the same as avoiding antitrust scrutiny, which one outlet notes still applies.16 Another argues that Reflection's strategic profile and size could invite closer attention than Groq did, especially if regulators ask whether Nvidia would have de facto control.13 Dealroom's take is that the option itself shows how large tech companies are designing AI deals to get around antitrust rules.2

The alternative is a full purchase like Hugging Face. Nvidia signed a definitive agreement in September to buy Hugging Face, with closing expected in the first half of 2027 subject to regulators.12 Even that deal is reported at different prices. Some outlets round it to $13 billion.1115 One breakdown gives about $11.9 billion plus up to $1 billion in retention equity.12 Another says about $12.9 billion.3 A full takeover of Reflection at anything close to its last valuation would be about twice the size of Hugging Face, with a long regulatory process to match.13

Circular financing

The deal also draws attention to how money moves between Nvidia and the labs it backs. Reflection has signed large compute contracts. It agreed to pay SpaceX $150 million a month from July 2026 through 2029 for Nvidia GB300 capacity at Colossus 2 near Memphis, a contract worth up to $6.3 billion. Nebius agreed to sell it another $1 billion of capacity.12 Reflection has also agreed to co-build a 250-megawatt data center with Shinsegae Group.13

One analysis spells out the loop critics point to. Nvidia invests in Reflection. Reflection commits billions to compute that runs on Nvidia hardware. Nvidia could then buy the startup at a valuation its own money helped support.1 That does not mean the valuation is wrong. It does mean the largest buyer of this asset is also the company most invested in seeing it look valuable.

The stake-plus-chips option fits these obligations directly. More Nvidia capital and hardware would help cover a compute bill that is large for any startup.12 But a lab that rents Nvidia chips, counts Nvidia as its lead investor, and then takes more of both would have little room to disagree with its sponsor.12

The strategic case: an American open-weight lab

Nearly every report frames the deal around geopolitics. Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both former Google DeepMind researchers, and is positioned as a US alternative to low-cost Chinese open models such as DeepSeek and Kimi.1611 It released its first open-weight model, Beam, in early October. Reflection says Beam matches Z.AI's GLM-5.2 on advanced reasoning benchmarks.11 Beam is a sparse mixture-of-experts model with 501 billion total parameters, of which 23 billion are active per token.13

Jensen Huang has argued for American open-weight models as a way to reduce dependence on Chinese alternatives.11 The Trump administration's July 2025 AI action plan calls for US open models that could become international standards.15 Reflection reportedly works with the Defense and Energy departments and is co-developing models with allies such as South Korea.13 Its investors include 1789 Capital, where Donald Trump Jr. is a partner.6

A product gap also helps explain Nvidia's interest. Nvidia makes the Nemotron open models but has no frontier-grade model of its own, and Reflection could fill that gap.13 That would mark a real shift for Nvidia, from selling the infrastructure models run on to owning a model.16 It could also create tension with customers such as OpenAI and Anthropic, which rely on Nvidia chips while competing on models.13

The caveat about Beam's weights

One detail undercuts the open-model story. As of mid-October, Beam was available only through a waitlisted API. Reflection had promised to release the weights under an Apache 2.0 license later in the month, and its Hugging Face organization listed no models.12 This matters because an Apache license, once granted, cannot be revoked for weights already published. A promise to publish, however, does not bind a new owner or a team that has moved to Nvidia.12 In July, an investor-focused outlet noted that Reflection had raised large sums without yet shipping a foundation model.8 Beam has changed that only in part.

The verdict

Taken together, the reporting points to Nvidia formalizing a relationship it has already been building: $800 million in equity, a lab whose compute runs on Nvidia hardware, and a model that serves Washington's open-weight goals. My view is that the acqui-hire or the larger stake is much more likely than an outright purchase. That fits the Groq and Poolside precedents and avoids a merger review that would follow so soon after Hugging Face.

For venture investors, the lessons are mixed. A hire-and-license deal at a premium, like Groq's reported $20 billion against a roughly $7 billion prior valuation, can make a cap table whole without a formal sale.12 For other Reflection backers, the outcome will determine how much liquidity they get and when.4 The broader lesson for AI valuations is that a lab's most likely buyer may be its own chip supplier, which also helped set the price. Bullish commentators say a deal would strengthen Nvidia's position, but even they concede it would not solve Nvidia's larger pressures, including custom-chip competition and China export limits.18