Nvidia is holding early discussions about acquiring Reflection AI, lifting its stake, or completing an acquihire, the Financial Times reports. A deal might land within weeks, said people briefed on the talks, though it could equally unravel.
Three shapes are under consideration. Nvidia might purchase the open-weight model developer outright, add to the investment it already holds, possibly with extra chip supply attached, or take on Reflection’s staff and license the technology while leaving control untouched. That last route would probably sidestep a drawn-out regulatory review.
Both approaches have recent precedent at Nvidia. Inference chip maker Groq licensed its technology in a December arrangement reported at $20B, with founder Jonathan Ross moving over as Groq carried on running its own cloud business. In September, Nvidia agreed to hand over $12.9B for Hugging Face.
Reflection already lists Nvidia among its major backers. The chipmaker led a $2B round in October 2025 and has committed $800M in total, the FT said. A $25B pre-money valuation was placed on the company by its most recent round, chief executive Misha Laskin told CNBC in April.
Reflection had unveiled Beam just five days before news of the talks broke. The model carries 501 billion parameters and uses a mixture-of-experts design. Against lower-cost open models from China, DeepSeek and Kimi among them, the company positions Beam on coding and agentic work, the area its coding agents first targeted.
Laskin and Ioannis Antonoglou started Reflection in 2024 after doing research at Google DeepMind. What attracts Nvidia is a model lab that reaches for its chips first rather than treating them as one option among several.
