Open-weight AI: Nvidia targets Hugging Face for $13 billion

The open-weight model market is attracting heavy investments: Nvidia is reportedly on the verge of acquiring Hugging Face for $13 billion, following a $6 billion deal with Poolside, and Stripe has invested over $7 billion in OpenRouter. Adoption remains limited but is increasing, with targeted uses for repetitive inference loads and promises of greater control.

Colossal Volumes and Limited Adoption

Fireworks claims to process 40 trillion tokens each day, a volume that its CEO Lin Qiao says surpasses that of the APIs from Gemini or OpenAI. Despite these significant uses, the adoption of open-weight models remains low: only 6% of companies use them according to Ramp, and just 2% of software engineers surveyed by Jellyfish. In light of rising inference costs, some companies are turning to cheaper models offered by Moonshot, DeepSeek, or Alibaba. Patrick Collison, co-founder and CEO of Stripe, emphasizes that tokens are essential for companies developing with AI, and that economic value will depend on the effective use of scarce computing resources.

Why Shift to "Open-Weight" (or Stick with Proprietary Models)

According to Nik Albarran, open-weight models are primarily adopted for high-volume repetitive inference loads, such as customer service chats, where they can be tuned for cost-effective responses. He notes that, for now, companies are choosing these models mainly for control and configurability, rather than to cut costs. In contrast, for coding or agent tasks, proprietary models remain favored, particularly because they offer easier access and sometimes token subsidies.

Multi-Billion Dollar Acquisitions in the "Open-Weight" Ecosystem

Acquisition deals are multiplying: Nvidia has finalized a $6 billion agreement with Poolside, the majority of whose employees are set to join the group. Two weeks prior, Stripe acquired OpenRouter for over $7 billion, touted as the leading provider of open-weight models for enterprises. In this context, a reported acquisition of Hugging Face by Nvidia for $13 billion is awaiting confirmation. Hugging Face, a platform for sharing open models and benchmarks, occupies a central position for developers building outside proprietary labs.

Strategic Calculations Around Chips and Models

OpenAI and Google are developing their own inference chips, and OpenAI showcased the capabilities of a chip called Jalapeño this week. In light of this evolution, Nvidia is looking to limit its dependence on hyperscalers and cutting-edge labs while aiming for a share of the model creation market. Nvidia already has Nemotron, a family of open-weight models whose adoption remains limited. Gaining control of the leading open model development space in the U.S. would give it access to a large user base to steer towards its chips and standards. In this landscape, Fireworks, led by Lin Qiao and frequently mentioned as a potential target, positions itself as a router and host of models for enterprises.

What Albarran and Qiao Envision for the Future

Nik Albarran believes that the adoption of open models will become easier as AI workflows become more structured, and that self-hosting will gain importance as processes advance sufficiently. He also indicates that rising prices at cutting-edge labs could push more companies to consider these alternatives. Lin Qiao bets on the increasing diversity of LLMs, which she sees as becoming more suited to specific needs, and recommends that publishers hire an in-house researcher to develop models based on their products and data. She advocates for the idea of specialized intelligence, with one model per use case. In this context, the dominance of OpenAI and Anthropic is not considered a given, while tech giants diversify their investments and show interest in open technology.