Nvidia weighs buying Misha Laskin's Reflection AI after its Beam debut
Nvidia's early talks could lead to an acquisition, acqui-hire, larger investment or computing deal with Reflection AI. Nvidia has already invested $800 million; Reflection's Beam model weights are due later this month.
By Ryan Merket · Published
Primary source: Financial Times
Why it matters
Nvidia is already a major Reflection investor and supplied the GPUs behind Beam's reported training run. An acquisition, additional investment or computing deal would tie the chipmaker more closely to a model team whose weights are not yet public, putting Nvidia on both the hardware and model sides of the open-weight market.
Reflection AI, the open-weight model startup founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, is in early talks with Nvidia about a possible acquisition or deeper investment, the Financial Times reported on October 10th. The talks come five days after Reflection introduced Beam, its first open-weight model. Reflection says it will release the weights and related materials later in October.
The discussions remain at an early stage. The FT reported that a transaction could be a full acquisition, an acqui-hire in which Nvidia hires Reflection staff and licenses its technology, or a larger equity investment. Nvidia could also deepen the relationship through a deal for more chips and computing capacity. Multiple people told the FT a deal could be reached in the coming weeks, while cautioning that discussions could fall apart. Reflection and Nvidia declined to comment to the newspaper.
Nvidia is already one of Reflection's largest shareholders and has invested $800 million in the company, according to the FT. Reflection's latest funding round closed at a $25 billion pre-money valuation on April 23rd, 2026, according to Reflection AI's newsroom. The FT said it could not establish the terms under discussion; that valuation is not a proposed deal price.
A model announcement, not a weights release
Reflection introduced Beam on October 5th in a company announcement. It is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active at a time, built for coding, reasoning and agentic workloads. Reflection says it trained Beam on 23.8 trillion tokens and ran more than 100 million reinforcement-learning rollouts using 10,500 Nvidia GB300 GPUs over four weeks. Those are company-reported figures.
The weights are not yet available to users. Reflection says Beam is undergoing final red-teaming and evaluations, and that it will release the weights, technical report, model card and developer materials later this month. Reflection offers a sign-up for early access. Until the release, outside developers cannot independently examine the weights or reproduce Reflection's reported performance.
Beam puts a concrete model behind the research ambition Laskin and Antonoglou have pursued since leaving DeepMind. Laskin led reward-modeling work for Gemini, according to TechCrunch's 2025 profile of the founders. Antonoglou helped create AlphaGo and AlphaZero and worked on post-training for Gemini, according to Sequoia Capital.
Their stated focus is on agents that can complete multi-step tasks. In a Sequoia podcast, Laskin said that "someone needs to solve the depth problem too." Reflection began with coding agents, where tests and successful code execution can provide feedback on whether a task worked. Beam extends that reinforcement-learning approach to a model Reflection says is designed for coding, reasoning and agentic work.
What Nvidia would gain
Nvidia has a model business of its own, including the Nemotron family, but the FT reported that it lacks a frontier-scale model. Buying or investing further in Reflection would give the chipmaker a closer tie to a startup building one. Reflection says its high-compute reinforcement-learning run used 10,500 Nvidia GB300 GPUs over four weeks. That does not establish who supplied the compute. Reflection secured compute access through providers including SpaceX and Nebius, which provide access to Nvidia chips, according to TechCrunch.
The commercial logic runs in both directions. Nvidia supplies the chips Reflection uses, while Reflection's models could give Nvidia customers another option alongside its hardware and software. The FT reported that Nvidia has called for a domestic open-weight ecosystem, while Reflection says it wants enterprises, governments and sovereign institutions to customize models and run them under their own control. Neither Nvidia nor Reflection AI has confirmed the talks or Nvidia's motives.
Nvidia was among the investors in Reflection's $2 billion October 9th, 2025 funding round, which valued the startup at $8 billion, according to TechCrunch's report and investor list. Reuters later corrected its account to remove a description of Nvidia as a lead investor, as Investing.com reported. Sequoia Capital and Lightspeed Venture Partners also participated in the round.
In October 2025, Laskin told TechCrunch that Reflection planned to release model weights while keeping much of its training data and full training pipeline proprietary. Reflection says Beam was trained in part on proprietary licensed datasets. The planned release will let developers inspect and run the weights, but it will not make the full training process or data public.
A full acquisition would put Reflection's team and model operation under Nvidia. A deeper investment, an acqui-hire or a computing agreement could strengthen the relationship without the same organizational change. For Laskin and Antonoglou, Beam is the first public model from a company built around their research credentials; its practical value will depend on what developers can do with the weights once Reflection releases them.