Nvidia is negotiating to buy Reflection AI, the open-weight lab it already funded

On October 10 the Financial Times reported that Nvidia is in early talks to acquire Reflection AI or to deepen its investment in the startup. Bloomberg and Reuters carried the same report. The deal could take three shapes: an acqui-hire, extra chips and compute, or a bigger equity stake. In an acqui-hire, Nvidia would hire the staff and license the technology instead of buying the company. A deal could land within weeks, the paper said, and it could also fall apart. Nvidia has already put $800 million into Reflection. FT Reuters Bloomberg

Reflection AI was founded in 2024 by two former Google DeepMind researchers. Its June round valued the company at $25 billion. On October 5 it shipped Beam, its first open-weight model. The model carries 501 billion total parameters and 23 billion active per token, with a context window of 1 million tokens. Reflection says Apache 2.0 weights will arrive later this month. Reflection

Nvidia agreed in September to buy the model hub Hugging Face for $13 billion. Before that it licensed technology from the chip startup Groq, in a deal reported around $20 billion. The company that sells the chips now owns more of the models that run on them. Open weights that circulate widely sell more GPUs, so Beam’s release looks likely to survive a deal. What changes is who sets the release date and the license terms. The Straits Times

Whether Beam’s weights actually land this month is the first test

If you build on open weights, this changes who is making the promise. Suppose you had planned to self-host Beam. The terms would now be set by a company in the middle of an acquisition. The concrete check is this month’s release: whether the weights appear as promised, and whether the license lets you redistribute and fine-tune. Keep one second path, either another open model in the same class or an API fallback. Run it once so you know it works.

The deal structure matters too. An acqui-hire moves staff and licenses the technology without the merger filing a full acquisition would trigger. The regulatory review some people expect may not happen. The Straits Times

The rest of today’s news

  • Yandex’s Vladimir data centre goes down after a drone attack: the third strike on its sites in a week, taking more than 80 services with it, including Yandex Cloud and the YandexGPT API. Al Jazeera Kyiv Post
  • The EU says the AI Act is enough: tech chief Henna Virkkunen said the law covers the whole life cycle of frontier models, and the Commission is reviewing replies from more than 30 providers it sent information requests to. InnovationAus
  • A Berkeley study finds 10 minutes of AI use erodes persistence: among 1,222 participants, those who got AI help stuck with hard problems for less time once the tool was removed. Berkeley News arXiv
  • Qwen-Image-2.1-Turbo ships: a 7B open-weight image generation and editing model that produces 2K images in 8 steps. MarkTechPost
  • Hone raises a $60 million seed: Benchmark and Index led at a $285 million valuation for its persistent business agents. Hone
  • A Super Micro contractor pleads guilty: the first plea in the $2.5 billion scheme to divert Nvidia AI servers to China. explainx
  • OpenAI resets Codex and ChatGPT Work usage: the reset landed worldwide on October 11 and does not cover API credits. AI Industry Today
  • Odyssey-3 opens a public preview: a free research preview of a world model that turns text prompts into interactive environments. TPS