NVIDIA reportedly in talks to buy or invest further in Reflection AI, five days after Beam launch, raising questions over whether a US open-weight developer will fall under a chipmaker's control

NVIDIA is in talks to either acquire Reflection AI or increase its investment in the company, the Financial Times (FT) reported. Bloomberg Law and Investing.com carried the news, both citing the FT article. Both are secondary reports, and neither NVIDIA nor Reflection has made an announcement. It has not been confirmed whether the talks would lead to an acquisition or an additional investment, or what the valuation or amount would be.

The timing matters. Reflection announced "Beam," its first open-weight model, on October 5, five days before the report. Beam's weights have not been released yet. A company that only just positioned itself as the top US developer of open-weight models could end up controlled by the world's largest chip seller.

What has been reported, and what remains unknown

According to the reports, there are two options: buy the company outright, or deepen the existing relationship by increasing its stake. NVIDIA is already a Reflection investor. At the time, it was reported to have taken part in the October 2025 round in which Reflection raised about $2 billion. The talks are about whether NVIDIA moves from being one of several investors to being the owner.

Much remains unknown: how far the talks have progressed, the terms and valuation, how regulators would review a deal, and whether Beam would still be released under Apache 2.0 after an acquisition. Figures in early reports have sometimes differed sharply from the final deals. Until the terms are clear, this should be treated as a report of talks and nothing more.

What Beam promises, and what it has yet to deliver

According to Reflection's October 5 announcement (a primary source), Beam is a mixture-of-experts (MoE) model with 501B total and 23B active parameters. It was trained on 23.8 trillion tokens and has a context window of 1 million tokens. The company says reinforcement learning used 10,500 GB300s over four weeks and ran more than 100 million rollouts. Its self-reported scores are 80.9 on SWE-Bench Verified, 80.1 on Terminal Bench v2.1, 90.5 on GPQA Diamond and 97.8 on AIME 2026. Reflection says the model matches GLM-5.2's performance with one-third to one-quarter of the inference compute.

For now, however, users can only join a waitlist. Reflection says it will release Apache 2.0 weights and a technical report "this month," but as of October 8 Beam's weights could not be found on Hugging Face. No third-party evals have been published either. Every performance figure comes from Reflection itself, and none has been independently verified.

In other words, Beam as an open-weight model is still only a promise. The company that made that promise could change owners before the weights are released.

Few US developers build open-weight models

Reflection draws attention because hardly any US companies build open-weight models. In OpenRouter's weekly figures for September 14–20 (as reported), Chinese models processed 67.46 trillion tokens, compared with 14.21 trillion for US models. Chinese models have now led for 21 straight weeks. In an Epoch AI test published October 8, in which Epoch handed its own work to AI, open-weight models also lagged clearly behind.

Outside the US and China the field is just as thin. Europe's Mistral Large 4 still has not released its weights, despite saying they would come "by the end of the month." In a September 29 report, Anthropic singled out Z.ai's GLM-5.3 as the open-weight model with the strongest cyber capabilities and pointed to the risk of proliferation. The US clearly needs a competitor to China's open-weight models.

On October 4, Axios described Reflection's aim as "a level that falls short of the US frontier but can compete with China's best open models." It also reported that the company would source compute from Nebius and SpaceX. Reflection was close to the only newcomer that could become that competitor.

Why a chip seller wants it

For NVIDIA, open-weight models help sell its chips. When anyone can get the weights, demand for the GPUs that run them spreads across clouds, enterprises and national data centers. Reflection says Beam used more than 10,000 GB300s for reinforcement learning, which makes it a handy showcase: a US-built model trained on NVIDIA's latest chips.

There are also problems. NVIDIA sells chips to nearly every frontier lab. If it owned a model developer, it would also be competing with its customers. Open-weight models might also be tuned to run best on particular hardware. Open weights are not the same as independence from any one company's interests.

Will the Apache 2.0 promise hold?

The first thing to check is whether Reflection releases the weights and technical report under Apache 2.0 this month, as it promised. If the weights come out, third-party evaluations can test the self-reported numbers. If they don't, people will inevitably suspect that the talks affected the decision to release.

The second is how the deal is structured. If NVIDIA only adds to a minority stake, Reflection stays independent, at least on paper. If NVIDIA buys the company, it will have the final say on licensing, release timing and which hardware gets priority. Either way, until NVIDIA and Reflection make formal announcements and primary sources disclose the terms, the story must be treated as a report.

Just five days ago, this US open-weight developer published concrete figures for the first time. Who ends up owning it will shape how the US competes with China's open-weight models.