Open-weight AI companies are the Valley’s hottest acquisition targets

Nvidia is reportedly closing in on a $13 billion buy of Hugging Face, the open-weight model hub that has become a central meeting point for developers building AI outside the frontier-lab club. If the deal lands, it would be the biggest sign yet that the smartest money in AI is moving toward the software people once treated like a free side alley. Hugging Face is no obscure repo. It’s where open models, benchmarks and the builders around them congregate, and TechCrunch notes it has even become famous as the target of reward-hacking OpenAI agents.

The rumored acquisition fits a pattern. Nvidia already agreed to put $6 billion into Poolside, an open-weight model builder, in a deal that will send most of Poolside’s staff to the chip giant. Stripe, meanwhile, bought OpenRouter two weeks ago for more than $7 billion. That company sits at the top of the market for open-weight models sold to businesses. So yes, a lot of capital is flowing into a sector built around giving models away, at least at the surface.

The strategic logic is pretty clear. Nvidia does not want to lean forever on the hyperscalers and frontier labs that buy its chips today and may build their own alternatives tomorrow. OpenAI and Google are already working on inference chips of their own, and OpenAI announced Jalapeño this week. If model builders want chips, Nvidia wants a slice of model building too. It already has its Nemotron family of open-weight models, but uptake has been limited.

Open models are still a minority play. Ramp’s spending data survey found 6% of companies using them, while Jellyfish’s survey of software engineers put the figure at 2%. Nik Albarran, Jellyfish’s AI product lead, says they fit best in repetitive workloads like customer service chats, where a model can be tuned to answer cheaply. That’s also the story Stripe is telling with OpenRouter: tokens are the real currency, and scarce compute has to be used well.

But coding and agentic work are a different beast. Requests vary more, reasoning matters more, and frontier models often still win, helped by easier access and, sometimes, token subsidies. Even so, Albarran argues companies are drifting toward open models as they get their AI workflows under control. Fireworks CEO Lin Qiao is pushing the idea even harder, saying her company processes 40 trillion tokens a day and that every app company should think about having its own model. Whether that turns into a durable business or just another round of AI copycat shopping is the real question.