Nvidia Negotiates Acquisition of Reflection AI to Integrate Beam Sparse MoE Team via Acqui-Hire Model

Multiple sources confirm that Nvidia is in early negotiations to acquire or invest in Reflection AI, a startup valued at $25 billion. The primary objective is to secure the team behind its newly released open-weight model, Beam, thereby accelerating the construction of a US-centric open-source large model ecosystem.

To bypass potential antitrust scrutiny associated with full acquisitions, the deal may utilize an "acqui-hire" structure similar to previous Nvidia moves involving Groq. This approach involves hiring the core team while licensing the technology. Beam features a 501B parameter sparse Mixture-of-Experts (MoE) architecture, claiming performance comparable to GLM 5.2 in coding and agentic tasks while reducing inference compute costs by 3-4 times. This follows Nvidia's acquisition of Hugging Face and absorption of Poolside's team, reinforcing its dominance in the open-source AI landscape.

This strategic shift marks Nvidia's transition from a pure infrastructure provider to a vertically integrated "chip + model" solution provider. If finalized, this move will reshape the competitive dynamics between open-source communities and commercial closed models, signaling Nvidia's intent to build the strongest open-source foundation model system in the US.