Nvidia is in talks to buy Reflection AI, FT reports

The best-funded American bet on open weights may end up inside the company that sells everyone the chips.

Nvidia is in talks to acquire Reflection AI or deepen its investment in the startup, the Financial Times reported on Saturday, citing people with direct knowledge of the matter. Neither company has confirmed it.

What the report says, and what it doesn't

Per Bloomberg's account of the FT report, Nvidia is weighing either buying Reflection or putting more money in. Deal terms could not be established. There is no price, no stake size and no timeline in the reporting.

Reflection was last valued at $25 billion. The Wall Street Journal reported in March that it was raising $2.5 billion at that pre-money valuation, Reuters reported at the time. Reuters already called the company "Nvidia-backed" then.

The timing is notable. On Oct. 5 Reflection announced Beam, its first open-weight model, which TopFive covered last week. Reflection describes Beam as a sparse mixture of experts with 501 billion total parameters and 23 billion active. It says the weights, a technical report and a model card come "later this month" under the Apache 2.0 license. Today it is waitlist only.

Reflection says Beam was pretrained in under four weeks on 6,144 Nvidia GB300 NVL72 GPUs.

Why a chipmaker wants a model lab

Reflection is a US startup building large open-weight models. Its Beam announcement measures the model against Chinese open models, GLM 5.2, Qwen 3.8-Max and Kimi K3, and concedes Kimi K3 is still ahead on raw capability.

For Nvidia, a strong open model is a sales tool. Every company that downloads open weights and runs them itself needs GPUs to do it. Closed models run on the lab's own cloud. Open ones run on yours.

That is the case for the deal. The case against it is the one you should care about. A lab owned by the dominant chip supplier has every reason to tune its models for that supplier's hardware first. Open weights stay open under Apache 2.0, but where the roadmap goes is another matter.

Nothing changes for you until the weights ship

This is a report of talks, and talks fall apart.

If you were planning to build on Beam, the license promise is what counts. Apache 2.0 weights, once released, cannot be taken back. Download them when they ship, whoever owns the company by then.

If you pick open models partly on who is behind them, add ownership to the list of things you track, next to the benchmarks.

Two dates decide it

The Beam weights, promised for later in October. And any statement from Nvidia or Reflection confirming or killing the talks.

Ship date first. Owner second.