Nvidia in Talks to Acquire Reflection AI, the Open-Source Lab It Backed With $800 Million

Nvidia Weighs Full Acquisition of Reflection AI

Nvidia is in early talks to acquire or significantly deepen its stake in Reflection AI, the open-source artificial intelligence laboratory it backed with an $800 million investment last year, according to reports from Bloomberg and the Financial Times. The options on the table include an outright acqui-hire, in which Nvidia would absorb the startup's researchers and license its technology, an expanded chip supply arrangement, or simply a larger equity investment in a company it already owns a piece of.

Reflection AI was valued at $25 billion in its last funding round, nearly triple the $8 billion mark it reached in October 2025, when Nvidia first wrote its $800 million check as part of a $2 billion raise. The company was founded in March 2024 by Misha Laskin and Ioannis Antonoglou, both veterans of Google DeepMind, with the stated goal of building frontier-grade open-weight language models that keep intellectual property onshore in the United States rather than ceding the open-source lane to Chinese labs such as DeepSeek.

The Beam Model and a Massive Compute Bill

On October 5, Reflection released Beam, a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, trained on 23.8 trillion tokens with a 1 million token context window. The company says Beam matches Z.ai's GLM-5.2 on reasoning benchmarks while using three to four times less inference compute, though those efficiency claims have not been independently verified. Full weights and a technical report are scheduled for release under an Apache 2.0 license later this month.

Training Beam required a four-week reinforcement learning session across 10,500 Nvidia GB300 GPUs. Reflection secured that capacity through a $6.3 billion, multi-year compute deal with SpaceX's Colossus 2 facility that has the startup paying roughly $150 million per month through 2029, and the company has also signed a computing agreement worth more than $1 billion with Nebius for Nvidia GPU access through 2029.

Strategic Stakes Beyond the Deal

The proposed acqui-hire structure matters because it is the version of the deal least likely to trigger a Federal Trade Commission review. Nvidia would not need to file for a formal merger if it is technically just hiring researchers and licensing code, a playbook well understood in Silicon Valley after the FTC's scrutiny of Microsoft's handling of Inflection AI's team in 2024.

The talks follow DeepSeek's R1 release in early 2025, which rattled Washington because an open-weight Chinese model rivaled American closed models at a fraction of the training cost. Reflection positioned itself from the start as a domestic alternative, giving enterprises and governments a credible option outside Chinese-origin models. The reports note that none of the described deal structures are finalized, and talks at this stage could still fall apart or shrink into something smaller than an acquisition.

Sources: startupfortune.com · Ft · Bloomberg · Reuters