Nvidia Weighs Buying or Deepening Its Bet on Reflection AI, the $25 Billion Open-Weight Startup It Already Backed With $800 Million

Nvidia is in early-stage talks to either acquire Reflection AI or significantly deepen its investment in the open-weight model developer, the Financial Times reported on Saturday, citing people with direct knowledge of the matter. The discussions could conclude within weeks, though the newspaper cautioned they may still fall apart.

The FT said a deal could take several forms. One option is an acqui-hire, in which Nvidia would hire Reflection staff and license its technology rather than pursue a full acquisition, a structure that could avoid a lengthy regulatory review. Other paths under consideration include an agreement to supply the startup with more chips and compute, or simply increasing Nvidia's equity stake.

Nvidia is already one of Reflection's largest financial backers, having invested $800 million in the company, according to the FT. Reflection was founded in 2024 by two former Google DeepMind researchers, Misha Laskin and Ioannis Antonoglou, and builds tools that automate software development, one of the fastest-growing use cases for AI.

The valuation has moved fast. Reflection's chief executive Laskin told CNBC in April that the startup was raising fresh capital at a pre-money valuation of $25 billion. Earlier this year the company also signed a deal with SpaceX for additional computing capacity at the Colossus 2 data center.

The talks come days after Reflection launched its first open-weight model. Beam, unveiled on Monday, contains 501 billion total parameters but activates only 23 billion per task, a mixture-of-experts design meant to make inference faster and cheaper. The company claims Beam is competitive with Chinese AI startup Z.ai's GLM-5.2 on coding and agentic tasks, though those benchmark results are self-reported and have not been independently verified.

The strategic context is a widening price-performance race with Chinese open-weight rivals such as DeepSeek, Kimi and Qwen, which offer cheaper and more customizable models that generate code nearly as well as leading systems from OpenAI and Anthropic. Backing a US open-weight champion is one way for Nvidia to keep that demand anchored to its own hardware.

For Nvidia, the logic is circular in a useful way: every open-weight model deployed by enterprises drives inference workloads, and inference workloads drive GPU demand. Whether the outcome is an acquisition, an acqui-hire or a bigger check, the effect is the same. The company that supplies the picks and shovels is now trying to own a bigger piece of the gold rush itself.