Nvidia in talks to buy Reflection AI, a startup valued at $8 billion

Nvidia is discussing the purchase of Reflection AI, a US startup that builds open frontier-level models, according to the Financial Times. The last known benchmark for the company's value is $8 billion, the valuation it received in October 2025 when it raised $2 billion. According to media reports, Nvidia itself took part in that round. Now it appears ready to move from investor to owner. The deal terms, price and timeline are not disclosed in the available data, and the talks could still fall apart.

What is known about the talks

The story comes from an FT article behind a paywall. We cannot relay the details, so we are relying on the headline and the company's public history. The link collected 84 points and 49 comments on Hacker News. The discussion boils down to one question: what happens to open weights if their author turns out to be the largest seller of accelerators? As far as is known, neither Nvidia nor Reflection AI has officially confirmed the deal.

Who is Reflection AI

The startup was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both Google DeepMind alumni. Laskin worked on reinforcement learning for Gemini, and Antonoglou was part of the AlphaGo and AlphaZero teams. The first product was Asimov, an agent for working with codebases. The company's main ambition is broader: to build an American open model that can compete with China's DeepSeek and Qwen, and to release its weights publicly.

Here is what is known about the company from public sources:

• Founders: Misha Laskin and Ioannis Antonoglou, both from DeepMind.
• October 2025 round: $2 billion at a valuation of about $8 billion.
• Investors, according to media reports: Nvidia, Lightspeed, Sequoia, DST, GIC, Eric Schmidt and others.
• Product: the Asimov agent for understanding and editing code.
• Positioning: open frontier models as an alternative to Chinese weights.

Nvidia's logic

For a GPU maker, open models pay off directly. The more companies run their own models on their own servers, the more accelerators they need. Closed APIs concentrate demand in a few labs, while open weights spread it across thousands of customers. Nvidia already has the Nemotron family, and buying Reflection AI would give it a team that knows how to train top-tier models, not just fine-tune other people's. Nvidia would also gain leverage in its dealings with Washington, where open American models are seen as a matter of technological sovereignty.

There are risks too. Nvidia has already invested in OpenAI, xAI, Mistral and other labs, and acquiring one of them will make its other customers wonder whether the chip supplier is becoming a competitor. Antitrust regulators in the US, the EU and China already watch the company's every move. A deal of this size will almost certainly draw their attention.

What teams building on open models should do

Until the deal closes, there is no need to change anything, but work out a fallback in advance. If you plan to rely on future Reflection AI weights, check which license applies to the releases already published, and keep copies of the checkpoints you need. Build in the ability to swap the base model quickly: moving to Llama, Qwen, Nemotron or DeepSeek should take days, not quarters. For those choosing between buying GPUs and renting cloud capacity, it is worth watching whether Nvidia starts releasing models optimized only for its own hardware. Such a scenario would change the total-cost-of-ownership calculation.