Nvidia chases developers with a wave of open-weight model acquisitions

The rumored $13 billion purchase of Hugging Face — roughly 48 billion shekels — is only the tip of the iceberg. Two weeks earlier Nvidia closed a $6 billion deal, about 22 billion shekels, for Poolside, a builder of open-weight models, with most of its team moving to Nvidia. At the same time Stripe acquired OpenRouter, the leading enterprise gateway for open-weight models, for more than $7 billion, or roughly 26 billion shekels. Together the three transactions signal a massive capital flow into a sector whose business model rests on giving technology away for free.

For Nvidia the logic is clear: reduce dependence on frontier labs and hyperscalers that are starting to produce their own inference chips. This week OpenAI unveiled Jalapeño, a proprietary inference chip, and Google is moving in the same direction. If model builders become hardware makers, Nvidia wants a piece of the model-building business. Nvidia already has its Nemotron family of open-weight models, but adoption remains low. Controlling the largest developer platform for open models would give it direct access to a user community it can steer toward its own chips and standards.

In the background, the price of inference is pushing companies to examine cheaper models from Chinese firms such as Moonshot, DeepSeek and Alibaba. Adoption is still narrow but growing: a Ramp spending survey shows only 6 % of companies use open-weight models, and Jellyfish, which builds developer tooling, puts engineer usage at just 2 %. Nik Albarran, Jellyfish's AI product lead, said the primary use case today is companies whose products rely on repeated inference workloads — customer-service chats, for example. In those high-volume, repetitive tasks an open model can be tuned to answer cheaply.

Stripe framed the OpenRouter acquisition around the same logic. "Tokens are the primary currency for companies building with AI, and the real economic potential depends on efficient utilization of scarce compute resources," said Patrick Collison, Stripe's co-founder and CEO. Yet for coding and agentic tasks, where requests are varied and complex reasoning is required, frontier models still win — partly because proprietary labs offer easier access and, in some cases, token subsidies. Albarran noted that as AI workflows mature it becomes easier to switch to open models, but the main driver today is control and configurability, not cost savings.

Fireworks, a leading enterprise router and host for open-weight models often mentioned as an acquisition target, disclosed that it processes 40 trillion tokens a day — more than the Gemini or OpenAI API. Fireworks is betting on model variety: as LLM proliferate and improve, it will become easier for companies to fine-tune them for specific needs. "Every application company should consider hiring an internal researcher," said CEO Lin Qiao. "They can use their own product and data to build a model tailored exactly to their problem."