(L/R) Dell Technologies CEO Michael Dell, Nvidia CEO Jensen Huang
(L/R) Dell Techologies CEO Michael Dell, Nvidia CEO Jensen Huang and entrepreneur Sergey Brin attend the "Golden Age of American Innovation Summit" at the Institute of Peace in Washington, DC, on October 8, 2026.

Nvidia (NASDAQ: NVDA) is in talks to acquire US open-weight model developer Reflection AI, or to provide additional funding on top of its existing investment. The Financial Times reported on Saturday (October 10), citing people familiar with the matter, that the discussions remain at an early stage. An agreement could be reached within weeks, but the talks could also collapse. Reflection declined to comment, while Nvidia did not immediately respond.

The structure of any deal has yet to be decided. The options reportedly include a full acquisition, an increased equity stake — potentially alongside the provision of additional chips and computing resources — and an "acqui-hire", under which Nvidia would hire Reflection's core employees and obtain technology rights without buying the company outright. The Financial Times said the latter arrangement could avoid lengthy regulatory scrutiny.

Valuation: US$25 billion before its first model was released

When Reflection was in talks for a new funding round in March, The Wall Street Journal reported that the company planned to raise US$2.5 billion at a pre-money valuation of US$25 billion. Chief executive Misha Laskin later told CNBC in April that the round had been completed at that valuation. In October 2025, Reflection raised US$2 billion at a valuation of US$8 billion, with Nvidia leading the round and contributing about US$800 million. Sequoia Capital, Lightspeed Venture Partners and DST Global also participated.

Reflection was founded in March 2024 by Laskin and Ioannis Antonoglou, two former DeepMind researchers. The company initially focused on autonomous coding agents before shifting towards open-weight models. It has yet to report significant revenue.

Beam: launched just five days ago

Reflection released its first open-weight model, Beam, on October 5 (Monday), targeting coding, reasoning and AI agent tasks. Beam is a text-only model built on a mixture-of-experts (MoE) architecture, with 501 billion total parameters and 23 billion activated for each computation. It supports a context window of one million tokens.

The company said Beam performed comparably with Chinese developer Z.ai's GLM-5.2 on advanced reasoning tests, while requiring three to four times less inference compute. Its coding and AI agent performance was "close to" Alibaba's Qwen3.8-Max, although the company acknowledged that Moonshot AI's Kimi K3 still led in raw capability.

The above figures were released by Reflection and have not yet been independently verified, TechCrunch reported. The South China Morning Post, citing early third-party testing, said Beam could be one of the most token-efficient open models in its class. European technology publication Trending Topics, however, said Beam lagged behind Kimi K3 and GLM-5.3 on the DeepSWE and Terminal Bench tests. The company said the model weights and technical report would be released later this month under the Apache 2.0 licence.

Beam was trained using about 10,500 Nvidia GB300 chips, with computing power rented from SpaceX's Colossus 2 data centre in Memphis. Reflection has been paying SpaceX US$150 million a month since July, under a lease running until 2029 and worth up to about US$6.3 billion in total.

Not Nvidia's first deal of this kind

Nvidia has adopted the "technology licensing plus talent hiring" approach several times in recent months:

  • Groq: In December 2025, Nvidia obtained a non-exclusive technology licence from AI chip start-up Groq for about US$20 billion and hired most of its employees.
  • Poolside: In August this year, Nvidia paid US$6 billion for a non-exclusive software licence, hired more than 100 engineers and invested US$1 billion at a pre-money valuation of US$12 billion to strengthen its own open-weight model family, Nemotron. Poolside's three founders remained with the company.

Bloomberg reported that such arrangements could again prompt US lawmakers to accuse Nvidia of evading merger scrutiny. Technology giants have used similar structures before, including Microsoft and Inflection, and Google and Character.AI.

Nvidia's calculations

In an open letter titled "Open Weights and American AI Leadership" published on July 24, Jensen Huang argued that US leadership in AI depends on building an open ecosystem spanning industries. In March, Nvidia also joined Mistral, Thinking Machines Lab and Perplexity in forming the Nemotron Coalition, of which Reflection is a member.

Analysts believe that the wider adoption of open models will lead more companies to deploy them independently, driving up demand for Nvidia GPUs. Reports have said Nvidia's own Nemotron range still lacks a frontier model developed independently in-house. Nvidia's equity investment portfolio already includes OpenAI, Anthropic, CoreWeave (NASDAQ: CRWV) and Nebius (NASDAQ: NBIS), as well as optical communications companies Coherent (NYSE: COHR) and Lumentum (NASDAQ: LITE).

Acqui-hires carry risks: if key members of an acquired team leave during integration, Nvidia could end up with little more than a technology licence rather than the research and development capability it sought.

What to watch next

  1. Deal structure: A full acquisition would mean Beam's revenue being included in Nvidia's financial statements, while a licensing-and-hiring arrangement or increased stake would point more towards securing a position in the ecosystem.
  2. Release of Beam's weights and independent testing: Whether Reflection's in-house data can be validated by third parties, and the level of adoption among developers.
  3. Regulatory and congressional response: In particular, whether similar arrangements face renewed scrutiny.
  4. Whether Reflection brings in new investors during the negotiations: If it does, this could affect Nvidia's stake and add complexity to the talks.