Nvidia, Stripe and the rush to acquire open-weight AI platforms
Nvidia is reportedly nearing a $13 billion acquisition of Hugging Face, the platform used to share open-weight AI models and benchmarks. The reported transaction follows Nvidia’s $6 billion agreement with open-weight model builder Poolside, under which most Poolside employees will move to the chipmaker, and Stripe’s acquisition of OpenRouter for more than $7 billion.
These transactions concentrate attention on companies that help developers find, run and adapt large language models outside the proprietary offerings of frontier labs. Hugging Face has become a central developer space for open models, while OpenRouter provides businesses with access to open-weight models. The reported valuations show that distribution, hosting and model-routing capabilities have become strategically important parts of the AI stack.
Why Nvidia is looking beyond chips
Nvidia has strong commercial ties with hyperscalers and leading model labs, but some of those customers are also developing inference hardware. OpenAI announced the capabilities of its Jalapeño inference chip, and Google is likewise building its own chips. The reported Hugging Face deal would give Nvidia access to a large community of model builders and users that it could steer towards its chips and standards.
Nvidia already offers its Nemotron family of open-weight models, although the source notes that adoption has not been large. Acquiring an established platform would therefore extend the company’s position beyond supplying compute and into the tools and communities used to build and deploy models.
Adoption remains limited, but use cases are clear
Open-weight model adoption is still relatively small. Ramp’s spending-data survey found that 6% of companies use open-weight models, while Jellyfish found that 2% of surveyed software engineers do so. Nik Albarran, AI product lead at Jellyfish, said these models are most often used where products depend on repeated inference workloads, including customer-service chats.
For high-volume and repetitive requests, a model can be tuned to answer common questions at lower cost. Stripe framed its OpenRouter acquisition around the need to make efficient use of scarce compute resources, with tokens acting as a central currency for businesses building with AI.
Coding and agentic work can present a different balance. Requests vary more and may require greater reasoning, leaving frontier models with an advantage through easier access and, in some cases, token subsidies. Albarran said businesses are more likely to consider self-hosting when their AI workflows have matured and when prices from frontier labs rise.
Control and specialization shape the next decision
The policy stakes around model availability remain visible in open-weight AI regulation and deployment pressures as developers weigh control, configurability and the conditions for deploying open technology. In the current market, Albarran said companies primarily choose open-weight models for control and configurability rather than spending alone.
Fireworks CEO Lin Qiao said the company processes 40 trillion tokens a day and is betting on model diversity. Her view is that application companies can use their products and data to build models tailored to individual use cases. For businesses, the immediate implication is to identify repetitive, mature workflows where tuning or self-hosting can provide the required control, while retaining frontier models where variable and reasoning-heavy tasks demand them.