Nvidia is reportedly closing in on a $13 billion acquisition of Hugging Face, the platform developers use to share open-weight AI models and benchmarks, according to TechCrunch AI. The deal isn’t confirmed yet, but TechCrunch AI reports that everyone in the Valley is waiting on Nvidia to make it official. If it lands, it caps off a wild few weeks of capital flooding into open-weight AI, a corner of the industry built on giving software away for free.
What stands out here is the pace. Three big moves in a matter of weeks, all pointed at the same target.
The deals stacking up
- Hugging Face: The rumored $13B Nvidia buy. Think of it as GitHub for the AI era, the hub where developers build and deploy models that frontier labs don’t own.
- Poolside: Nvidia already struck a $6B agreement with this open-weight model builder, moving most of its staff to the chip giant.
- OpenRouter: Two weeks ago, Stripe bought the top provider of open-weight models to businesses for more than $7 billion.
That’s north of $26 billion chasing a sector that hands out its core product for free. So why the rush?
Why Nvidia wants in
Nvidia’s problem is dependence. Right now it leans on deals with hyperscalers and frontier labs to sell chips. But those same labs are building their own silicon. OpenAI just announced the capabilities of its Jalapeño inference chip this week. Google is doing the same. If model builders are making their own chips, Nvidia wants a piece of the model-making business in return.
The company already ships its own Nemotron family of open-weight models, but uptake has been thin. Buying the biggest U.S. developer space for open models fixes that overnight. It hands Nvidia a mass of users it can steer toward its chips and its standards. That’s the real prize.
There’s a cost angle too. As inference bills climb, companies are eyeing cheaper models from Chinese firms like Moonshot, DeepSeek, and Alibaba. Nvidia would rather keep that demand inside its own ecosystem.
How small this actually is (for now)
Here’s the part worth sitting with: adoption is tiny. Just 6% of companies use open-weight models, per spending data from Ramp cited by TechCrunch AI. Only about 2% of software engineers touch them, according to developer-tools firm Jellyfish.
Nik Albarran, Jellyfish’s AI product lead, told TechCrunch that open models mostly show up in high-volume, repetitive work. Customer service chats are the classic case. You tune a cheap model once, then let it answer the same questions millions of times. For coding and agentic tasks, frontier models still win, partly because proprietary labs make access easy and sometimes subsidize the tokens.
Stripe framed its OpenRouter buy in exactly those economic terms. “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,” CEO Patrick Collison said.
But Albarran notes the main draw today isn’t price. It’s control and configurability. “There are not many companies where that is the case yet… [but] if the prices continue to go up from the frontier labs, more and more companies will be forced to at least consider it,” he told TechCrunch.
What comes next
The bigger bet is on specialization. Lin Qiao, CEO of open-weight host Fireworks, says her company already processes 40 trillion tokens a day, more than Gemini’s or OpenAI’s APIs. Her pitch: every company should eventually run its own model per use case.
“Every single app company should consider hiring an in-house researcher,” Qiao told TechCrunch. “The future is actually specialized intelligence.”
That’s the thesis behind all this dealmaking. The dominance of OpenAI and Anthropic looks locked in today, but it isn’t guaranteed. Tech giants are hedging, and open technology is proving hard to resist.
For practitioners, the signal is clear. Watch frontier pricing. The moment self-hosting pencils out, the companies that own the open-model rails, and the chips underneath them, will be waiting. You can read the full breakdown at the original source.