Everyone is on edge waiting for Nvidia to make official what could be the most significant tech deal of the week: a reported $13 billion acquisition of Hugging Face, the go-to platform for sharing open-weight AI models and benchmarks. Hugging Face, which recently gained attention as the target of a group of reward-hacking OpenAI agents, sits at the heart of the ecosystem where developers build and deploy large language models outside the control of frontier labs. You could think of it as a GitHub tailored for the AI era.

This rumored deal follows Nvidia's $6 billion agreement with Poolside, an open-weight model developer, which will see most of Poolside's staff transition to the chip giant. And just two weeks ago, Stripe closed its acquisition of OpenRouter, a leading provider of open-weight models for businesses, for over $7 billion. That's a substantial wave of capital flowing into a space built on giving things away, and it highlights where the AI industry is heading. For Nvidia, the push is about reducing reliance on its partnerships with major hyperscalers and frontier labs—especially as big model creators like OpenAI and Google develop their own inference chips, such as OpenAI's Jalapeño, which was unveiled this week. If model builders are getting into hardware, Nvidia wants a stronger foothold in the model-building side of the business.

Nvidia already offers its own Nemotron line of open-weight models, but adoption has been modest. By taking control of the largest U.S. developer hub for open models, the company could tap into a vast user base and steer them toward its chips and standards. There are also mounting concerns about the cost of AI inference, pushing companies to explore cheaper alternatives from Chinese developers like Moonshot, DeepSeek, and Alibaba. Right now, adoption of open-weight models is still in its early stages—only about 6% of companies use them, based on spending data from Ramp, or just 2% of software engineers surveyed by Jellyfish, a developer tools maker.

Because these models handle high-volume, repetitive tasks, an open-weight model can be fine-tuned to deliver answers at a lower cost. That's exactly how Stripe has framed its OpenRouter acquisition. "Tokens are the central currency for companies building with AI, and it's clear that the real-world economic potential will depend on making good use of scarce compute resources," said Patrick Collison, Stripe's cofounder and CEO, in a statement. For coding and agentic tasks, though, the varied nature of requests and the need for deeper reasoning often give frontier models an edge, partly because proprietary labs offer easier access and sometimes subsidize tokens. Albarran notes that as companies refine their AI workflows, turning to open models will become simpler. Still, the primary motivation for adopting these models today is control and configurability, rather than budget pressures. "When your AI-driven workflows are much more mature, that's when it makes sense to invest in self-hosting models."

Lin Qiao, CEO of Fireworks—a leading router and host for open-weight models used by corporations and frequently mentioned as a potential acquisition target for a tech giant—says her company processes 40 trillion tokens daily, surpassing both Gemini and OpenAI's APIs. Fireworks is betting on model diversity: as LLMs multiply and improve, companies will find it easier to customize them for specific needs. "They can use their product and product data to build their own model. The future is actually specialized intelligence. Literally, every single company should have their own model per use case, and that will happen automatically."

It's easy to overlook how early we are in AI's evolution as both a tool and a business. The dominance of OpenAI and Anthropic isn't a given, though. As tech giants look to hedge their bets against the biggest labs, the appeal of open technology is becoming increasingly hard to ignore.