Nvidia’s Hugging Face Buy Defends Against Its Biggest Customers

Nvidia agreed to buy Hugging Face for $12.93 billion, the chipmaker confirmed Thursday in a blog post from CEO Jensen Huang — a deal that puts the world's largest hub for open AI models under the ownership of the company whose GPUs run most of what gets built on it.

The number matters less than what it is paying for. At roughly $150 million in annualized revenue, the price works out to about 86 times sales, a multiple several financial outlets converged on this week. Hugging Face isn't yet profitable. What Nvidia is actually buying is control of the pipe through which developers discover, fine-tune and deploy open-weight models — at a moment when its own biggest customers are trying to need it less.

That's the mechanism worth sitting with. OpenAI is developing custom silicon with Broadcom. Anthropic has its own chip effort underway. Google and Amazon have run in-house accelerators for years. Each one that works is GPU revenue Nvidia doesn't collect. A thriving open-model ecosystem cuts the other way: it gives every company that doesn't want to build a frontier lab in-house — or pay one for API access — a credible alternative, and the hardware people reach for to run those open weights is overwhelmingly Nvidia's. Owning the platform where that ecosystem lives ties developer habits to Nvidia hardware more directly than any chip roadmap can.

Nvidia isn't new to this logic. It bought Israeli orchestration startup Run:ai for roughly $700 million in 2024, a deal EU regulators cleared unconditionally after an eight-month review; it struck a $6 billion agreement with open-weight builder Poolside; Groq assets went for $20 billion in December. Stripe paid over $7 billion for model router OpenRouter earlier this summer, up from a $1.3 billion valuation in May. Model-hosting and orchestration layers are being bought up across the industry, and Hugging Face — with 18 million developers, 3 million models and half a million datasets, according to Nvidia's own count — is the largest prize in that category yet.

It's also the riskiest, because Hugging Face's value depends on being nobody's subsidiary. Futurum analyst Nick Patience noted the price only makes sense as a bet on future distribution leverage rather than current financial performance, since nearly every open model already runs on Nvidia hardware without Nvidia owning the storefront. Rival chipmakers AMD and Intel, along with Google, Amazon and OpenAI's own custom-silicon efforts, have a direct interest in Hugging Face staying neutral and are expected to raise exactly that concern once the deal reaches antitrust review. Unlike Nvidia's Groq and Poolside arrangements, an outright acquisition of this size triggers a mandatory U.S. waiting period and likely EU scrutiny; the company itself is targeting a close in the first half of 2027, which gives regulators room to make this take longer than Nvidia wants.

There's a Greater Bay Area stake in this that goes beyond chip geopolitics. Alibaba's Qwen family and DeepSeek's models are, by download volume, now among the most heavily used open-weight systems on Hugging Face — Qwen alone accounts for roughly 40% of new model derivatives on the platform, and DeepSeek has passed a billion cumulative downloads there. Moonshot AI's Kimi models ship through the same hub. For Chinese labs that depend on Hugging Face as their main channel to global developers, an American chipmaker now owns that channel, at the same time Washington continues to restrict which Nvidia chips Chinese buyers can access. Huang's post pledges that "NVIDIA compute will not be required to build on or deploy through Hugging Face" and that the platform will keep serving every model builder — a commitment that will be tested the first time Nvidia has to choose between platform neutrality and steering more of that traffic toward its own hardware and cloud partners.

Yahoo Finance quoted Nvidia investor and BD8 chief executive Barbara Doran calling the move a hedge for a future when raw chip demand eventually cools, not a signal that it's cooling now. The bigger question the deal leaves open is a structural one: Hugging Face's model today is built around developers downloading weights, which is exactly the behavior that keeps them buying GPUs. If distribution keeps shifting toward renting inference through an API rather than pulling weights down to run locally — the direction closed labs have pushed for years — owning the download hub buys Nvidia less than the price tag suggests. Regulators reviewing the deal into 2027 will be the first real test of how much room Nvidia has to act on what it just bought.