NVIDIA is currently in negotiations for a deal that could significantly enhance its access to a state-of-the-art AI model. According to a report by the Financial Times on October 11, the tech giant, which holds the title of the world's most valuable company, is engaged in discussions with Reflection AI, a developer of open-weight models. The potential deal options include a direct acquisition, an additional equity investment, or a unique structure known as an "acqui-hire."

The latter option would enable NVIDIA to recruit the entire Reflection AI team and secure a technology license, all without undergoing the comprehensive antitrust review typically associated with a full merger.

NVIDIA is a leading provider of GPUs, a critical component in the AI landscape. Despite its development of large models, the company has yet to establish a premier open-source model of its own. While it has introduced the Nemotron series, their performance has not reached the pinnacle of the industry. Reflection AI, on the other hand, recently unveiled its inaugural model, Beam, on October 5. Notably, Beam was developed using NVIDIA chips throughout its architecture and training phases.

Beam is a sparsely activated expert model boasting a total of 501 billion parameters, with only 23 billion activated at any given time. Reflection AI claims that Beam can rival Zhipu's GLM-5.2 in code and agent tasks, despite GLM-5.2 having approximately 744 billion total parameters and 40 billion activated. In essence, Beam utilizes fewer active parameters to compete with larger models, thereby reducing inference costs.

It's important to note that this assertion is based on the vendor's own statement and has not undergone independent verification. Nevertheless, Beam's training data is substantial, having undergone reinforcement learning on around 10,500 GB300 graphics cards and pre-training utilizing 6,144 GB300NVL72 units. A company whose survival hinges on NVIDIA's computing power has now developed a product that can stand toe-to-toe with Chinese open-source models.

Reflection AI positions Beam as a cost-effective solution. The company claims that Beam achieves comparable scores to GLM-5.2 on advanced reasoning benchmarks while utilizing only one-third to one-fourth of the inference computing power. For businesses, open-source weights enable them to modify and deploy the model on their own hardware, resulting in significantly lower operational expenses compared to closed-source interfaces from OpenAI or Anthropic. This is why NVIDIA is keen to allocate its scarce GB300 computing power to Beam first.

Over the past two years, open-source weight models have gained traction, primarily driven by Chinese labs. DeepSeek, Kimi, Qianwen, and Zhipu have elevated capabilities to new heights while maintaining competitive pricing. In contrast, most Western majors have opted to conceal their best systems behind paid interfaces. NVIDIA is eager to promote a Western flagship model to prevent enterprises from opting for Chinese models to cut costs.

Thus, Beam aligns seamlessly with NVIDIA's interests from the outset.

Reflection AI was established in 2024 by two former Google DeepMind researchers, Mikhail Raskin and Yannis Antonoglou. The company secured a new funding round in April of this year, aiming to raise $2.5 billion with a pre-money valuation of $25 billion. Less than a year ago, its valuation hovered around $8 billion, with the surge attributed to capital market bets on Western open-source models. NVIDIA led a $2 billion funding round in October of the previous year, having already invested $800 million, making it one of Reflection AI's most significant investors.

To secure the necessary computing power for training, Reflection AI also inked a deal worth up to $6.3 billion with Elon Musk's SpaceX to gain access to cards at the Colossus 2 data center. Additionally, it forged a multi-billion-dollar partnership with cloud service provider Nebius.

The term "acqui-hire" is not novel. On December 24 of the previous year, NVIDIA acquired the inference technology of chip company Groq through a non-exclusive license valued at approximately $20 billion. Simultaneously, it recruited founder Jonathan Ross and roughly 90% of the employees, leaving Groq as an independent entity. From an economic perspective, NVIDIA effectively acquired Groq's core assets.

Legally, Groq remains an independent company. However, this structure is now under scrutiny by the U.S. Department of Justice. Bloomberg reported in September that the DOJ is investigating whether the deal was specifically designed to circumvent merger reviews. Furthermore, former Groq employees who were not absorbed into NVIDIA filed a lawsuit this week.

In essence, NVIDIA aims to replicate its strategy with Groq with Reflection AI. However, the regulatory environment has become more stringent compared to a year ago. A member of the U.S. Federal Trade Commission publicly cautioned earlier this year that large companies might exploit hiring arrangements to bypass merger filing requirements under the HSR Act. Given Reflection AI's strategic significance and $25 billion valuation, this review is expected to be more rigorous than that of Groq.

By laying out several deals, NVIDIA's strategy becomes apparent. In September, it acquired the model and dataset platform Hugging Face for approximately $13 billion and is now closing in on Reflection AI. While its chips remain the lifeblood of the entire AI economy, NVIDIA is not content with merely selling hardware; it aspires to dominate the model layer as well.

Once this vertical integration of chips and models is realized, NVIDIA will transition from a computing power supplier to a shaper of the open-weight ecosystem.

Jensen Huang, NVIDIA's CEO, stated in July that the U.S. must establish its own domestic open AI ecosystem, which is crucial for fostering innovation and prosperity opportunities nationwide. Trump's AI action plan also views open-weight models as a potential global standard. The value of companies like Reflection AI lies in this geopolitical context—the West requires a flagship model capable of competing head-on with China's affordable and effective open-source models.

The specific terms, price, and whether the deal will be an acqui-hire or a company acquisition remain uncertain. However, NVIDIA's intention to integrate open-weight model capabilities into its own system is already evident; the only real question is how lenient the regulatory hurdles will be.