Nvidia Eyes Reflection AI in Talks That Could End in Acquisition, Acquihire or More Investment

Quick Facts

Nvidia is in early-stage discussions to fully acquire, acquihire or invest more in Reflection AI, according to the Financial Times and Bloomberg.
Reflection AI, founded in 2024, raised a $2 billion round led by Nvidia in October 2025 at a $25 billion pre-money valuation, with Nvidia having put $800 million into the company in total.
Reflection launched Beam on October 5, a 501-billion-parameter open-weight model targeting enterprise and sovereign AI customers, with weights due under an Apache 2.0 license later this month.

Nvidia Corp. is in early discussions to buy Reflection AI Inc., bring on its engineers through an acquihire, or increase its existing stake in the startup, the Financial Times first reported Sunday. Bloomberg corroborated the story. No terms have been set, and the talks could still collapse.

Three deal structures are reportedly on the table. A full acquisition could come alongside a larger supply of chips and computing capacity. An acquihire would see Nvidia hire Reflection's staff and license its technology without taking ownership, a structure the FT noted would likely sidestep a lengthy regulatory review. A straight equity investment would leave Reflection operating independently.

Nvidia already led Reflection's $2 billion funding round in October 2025 and has put $800 million into the company in total. That round drew participation from former Google CEO Eric Schmidt, Citi, Lightspeed, Sequoia and 1789 Capital, the private equity firm backed by Donald Trump Jr. CEO Misha Laskin told CNBC in April that the round closed at a $25 billion pre-money valuation.

Reflection was founded in 2024 by Laskin and Ioannis Antonoglou, both former Google DeepMind researchers. Laskin led reinforcement learning reward modeling for Gemini 1.5. Antonoglou was a founding DeepMind engineer and one of the co-creators of AlphaGo, the first AI system to defeat a world champion in Go. The Brooklyn-based company now employs roughly 320 people.

On October 5, Reflection unveiled Beam, its first frontier model. Beam is a sparse mixture-of-experts model with 501 billion total parameters, though only 23 billion are active on any given token. The company says Beam was pretrained on 23.8 trillion tokens from curated web and licensed data. Reflection claims it matches the performance of GLM 5.2 on reasoning benchmarks while using three to four times less inference compute.

Beam targets coding, reasoning and agentic workloads. Reflection is positioning the model against closed labs such as Anthropic and OpenAI, open Chinese models from DeepSeek and Qwen, and Western open-weight players including Mistral, Meta and Cohere. The company plans to release Beam's weights under an Apache 2.0 license later in October, along with a full technical report and tooling for running and fine-tuning the model.

Reflection has built a substantial compute base to train and serve Beam. Starting July 1, the company pays $150 million a month for access to Nvidia GB300 systems inside SpaceX's Colossus 2 data center in Memphis under a deal worth up to $6.3 billion through 2029. Nebius Group agreed in July to sell Reflection an additional $1 billion in AI capacity.

The startup has also secured government business. In May, Axios reported that Reflection was partnering with the U.S. Department of Energy on the Genesis Mission federal research initiative and would serve as the AI model provider for the U.S. National Laboratories. Reflection has also begun testing a sovereign AI factory model with South Korean conglomerate Shinsegae Group.

A deal with Reflection would extend an active Nvidia acquisition strategy. On September 3, Nvidia announced a $12.93 billion agreement to acquire Hugging Face, which will retain its brand and open platform status. Nvidia also committed $6 billion to license Poolside's Model Factory software and made offers to hire 109 Poolside engineers tied to its open-source model work.

This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.