Nvidia has already written one large check to Reflection AI: about $800 million for a stake in a two-year-old lab that gives away its models and sells the enterprise service around them. Now the chipmaker is negotiating to buy the rest of the company, or to buy more of it.
The talks, reported on October 11, cover either an acquisition of the startup or a further investment, according to people familiar with the matter. They are at an early stage, and a deal could be reached within weeks, though the people cautioned the discussions could also collapse.
The reports build on previous reporting of Nvidia’s $800 million position and of the chipmaker’s interest in deepening it. The talks have since widened to a possible outright acquisition, a step the earlier coverage did not describe in detail.
Three structures are on the table. Under one, Nvidia would hire Reflection’s staff and license its technology rather than buy the company outright, the arrangement the industry calls an acqui-hire. Under another, Nvidia would agree to supply the lab with more of the chips it needs to train larger models. The third is a plain increase in Nvidia’s equity stake.
The acqui-hire route carries the lightest regulatory footprint. A full purchase of a lab whose private valuation has run into the tens of billions would invite a lengthy antitrust review in Washington, where regulators have spent two years circling the largest deals of the AI boom. Hiring the team and licensing the weights would let Nvidia capture most of the value without opening that process.
That template is well known to antitrust enforcers. Microsoft hired most of the staff of Inflection AI and licensed its technology in 2024, and Google reached a comparable arrangement with Character.AI. Both stopped short of a purchase. Enforcers have questioned arrangements of that kind in the years since without blocking them, which has left the structure available to companies that want the people and the models without the ownership.
Reflection was founded in 2024 by two former Google DeepMind researchers, Misha Laskin and Ioannis Antonoglou, to automate software development. It came out of stealth in March 2025 with $130 million in financing, then raised $2 billion that October at an $8 billion valuation. By June 2026 the company was worth $25 billion, based on its own fundraising.
Its first open-weight model, Beam, arrived on October 5: a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active at once, built for coding, reasoning and agentic work. Reflection says it is competitive with China’s leading open releases at a lower cost per token, and the company has said it plans to publish the full weights later in October.
That position is why a chip company cares. Open weights have become the volume business of AI. Chinese labs pushed prices down, Western labs answered with free releases of their own, and every company that downloads a model still has to run it somewhere. Those somewheres tend to be clusters of Nvidia chips.
The logic runs through Nvidia’s order book rather than its preferences. A thriving open-model ecosystem is a demand engine the company does not have to build or staff. Nvidia has funded the other side as well, anchoring OpenAI’s record financing round, and sells to closed and open labs alike.
That neutrality has been its most valuable asset. The claim of supplying every camp has kept Nvidia outside the regulatory crossfire that has found cloud providers and model makers. Owning a frontier lab would complicate the story, which is part of why the acqui-hire structure sits on the table.
Reflection also fits a pattern. Labs of its size are short of capital and silicon at the same time, and Nvidia can supply one while exercising influence over the other. Reflection has signed a computing agreement with SpaceX that runs to 2029 and was reported at up to $6.3 billion, giving it access to Nvidia’s GB300 chips at the Colossus 2 data center near Memphis, hardware that a larger owner would help guarantee.
The talks could still end with nothing more than a larger equity check. Nvidia’s $800 million already sits on the books, and a follow-on investment would deepen the tie without the burden of ownership, the regulatory review, or the responsibility that comes with running a lab whose mistakes become your own.
What is less ambiguous is the direction. The supplier that built its position by selling to everyone is moving, lab by lab, from vendor to owner. Whether Nvidia wants the job that follows, from the safety debates to the disclosure duties and the capital demands of training models it can no longer blame on someone else, is the part no term sheet answers.
For now the early-stage talks say something about the middle ground in artificial intelligence, and about how few companies can afford to hold it.
