Nvidia's $12.9 Billion Bet That Its Next Battle Is in the Models, Not the Chips

On September 3, Nvidia agreed to buy Hugging Face for $12.93 billion.

That is not a chip company's headline. Hugging Face does not make silicon. It is the hub where more than 18 million developers, researchers, and startups share, discover, and download open AI models — the place where a builder decides which model to run before deciding which chip to buy it on. Nvidia paid about three times the $4.5 billion the startup was valued at in 2023, on the heels of its best quarter ever, with the explicit promise that Hugging Face "will remain open" and that its own chips are "not required" to use the platform.

Hold that tension. The company that sells the world the fastest AI accelerator — and runs the software moat, CUDA, that keeps customers locked in — bought the neutral marketplace where developers compare everyone's models, and then said it would not lock anyone to its hardware. That is an odd thing to do. Understanding it requires seeing where Nvidia's moat is weakest, and the answer is not in the chip.

Training built the moat. Inference will test it.

For two years the story was simple. Training an AI model — the one-time, enormous burst of compute that builds it from scratch — runs overwhelmingly on Nvidia's GPUs, and almost no one can leave because of CUDA, the software layer that makes Nvidia's chips the easy default. That training moat is real, and it is why the last quarter was absurd: revenue of $96.2 billion for the quarter ended in late July, up 106% from a year earlier, with $89 billion of it from data centers alone at a 75% gross margin. On the earnings call, Jensen Huang said AI has "reached its inflection point." The company guided the next quarter to $108 billion — assuming, notably, no data-center compute revenue from China.

But the part of the cycle that will decide the next several years is not training. It is inference — running a finished model to answer a question, generate a few words of text (each one a "token"), or drive an agent through a task. Inference is where the money is actually earned, because it happens millions of times a day at massive scale, and it is the part that cares most about one number: cost per token. Inference is also where the CUDA moat is thinnest, because it rewards latency, efficiency, and cost per watt more than raw power — which is exactly the door custom chips from Google, Amazon, and others have been walking through for years.

So when Nvidia unveiled its next generation, Vera Rubin, it led with that number: up to a 10x reduction in the cost of inference per token versus the current Blackwell generation, and 4x fewer GPUs to train the large "mixture-of-experts" models that now dominate. Those are Nvidia's own claims, not independent benchmarks. But one early data point is a customer, not a press release: CoreWeave, among the first to stand up Rubin racks, measured its first performance at 10x more tokens per megawatt than Blackwell. The bet is that by the time Rubin ships in the second half of this year, "cheapest token" — not "fastest chip" — is the number the whole market prices on.

What $12.9 billion actually buys

This is where the Hugging Face deal stops looking like a logo purchase. Hugging Face is the front door of the open-model ecosystem — the 3 million models, 500,000 datasets, and 200,000-plus companies that discover and deploy AI there — and Nvidia is already its single largest contributor of open models. Owning the hub where developers evaluate and ship models puts Nvidia at the center of the exact moment a builder picks a model, in an era where the defining question is the inference economics of open models.

The read that matters is about who is not at that table. The dominant frontier labs — OpenAI, Anthropic — ship closed models, tied to their own stacks and increasingly their own or rented chips. The open model is the part of the AI future that stays neutral, and it can run on anything. By owning the open-model marketplace while the closed labs entrench on their own silicon, Nvidia is trying to make "the open model" read as "runs best on Nvidia" — even while it insists the platform stays multi-cloud and multi-chip. Reported analysis from TechCrunch points to a quieter, harder reason: Nvidia has committed to tens of billions of dollars in cloud compute deals for customers, and if those customers do not use all the capacity they have bought, Nvidia can be on the hook. Hugging Face's developer base is a ready channel to sell into. The demand behind the offense is concrete, not hoped for: on the same day as the earnings, AWS said it will add 2 million more Nvidia GPUs — Blackwell Ultra, Rubin, and Rubin Ultra — for 2027 and 2028, roughly trippling its commitment to about 3 million chips in five months.

None of this is a warning that the chip thesis is broken. It is the opposite: the chip business is compounding so fast that Nvidia can afford to buy its way into the layer it expects to matter next. But the offense does change two things an investor actually tracks. It moves the interesting part of the story from 2026 revenue into 2027 and 2028 — when Rubin and the open-model hub are both running at scale and someone has to prove that cheaper tokens translate into more Nvidia revenue, not just cheaper rivals. And it raises the bar on what the capital is doing. The discipline question is no longer "does Nvidia win training?" It is whether this is the best place to put this money — in open models, in the silicon, or in the next dark horse that appears once inference gets crowded. Nvidia no longer needs to own the market to keep growing, because the market is still expanding under its feet. But it now has to defend the layer where the training moat does not reach. That is where the offense starts to bend.