QuarryWhy is an American chip company paying $11.9bn for a site full of free, often Chinese, AI models?
When AI models are free, who gets paid?
NVIDIA reports on 17 November with no data-centre chip sales to China in its forecast. Yet its own filing says many of the world's most popular open models are Chinese, and it is paying $11.9bn for the site that distributes them. Free models, wherever they come from, sell American chips.
Jaidev Karthickeyan7 min readData: NVIDIA Forms 10-K, 10-Q and 8-K; Vercel; Artificial Analysis
What happened
On 2 September NVIDIA agreed to buy Hugging Face, the site where developers share open AI models, for about $11.9bn. Buried in the filing is a sentence you would not expect from an American chip company: "Many of the world's most popular and successful open-source models originated in China and are then downloaded, revised, fine-tuned, and tested by developers in the United States and worldwide."
NVIDIA reports its third quarter on Tuesday 17 November, and its forecast holds the other half of the story. It expects about $108.0bn of revenue while "not assuming any Data Center compute revenue from China". Its licensed chip shipments there were under 1% of data-centre revenue last quarter, held back by both governments. So China isn't much of a customer. It has turned into something more useful: one of the biggest suppliers of free models that make everyone else buy chips.
My read is simple. NVIDIA wants models to be worth as little as possible, because the less a model costs, the more of the bill lands on the chip. It doesn't matter to NVIDIA who gives the models away. It matters to Washington, because many of the most used ones are Chinese.
The chart shows who is buying. NVIDIA splits its data-centre revenue between the hyperscalers and everyone else: AI clouds, industry, enterprise and governments. In the quarter to July, the hyperscalers bought $48.7bn, up 102% in a year. Everyone else, the buyers most likely to run a downloaded model, bought $40.3bn, up 138%.
The paper trail
What NVIDIA's filings said about open models in 2026
February, annual report: open-source AI, "if deployed on our competitors' platforms, … could reduce demand for our products and services"
A risk
August, quarterly report: "We believe the demand for open-source foundation models and applications promotes use of our products worldwide"
A driver
September, deal filing: agrees to buy Hugging Face for about $11.9bn
A purchase
Same filing: "Other parties are actively lobbying the U.S. Government" to restrict open-source models
A fight
In seven months open models went from a risk to manage to a market to own and defend. The annual report already said the quiet part: NVIDIA speeds up its own open-model releases "in support of market development".
The strategy, unpacked
Four levers NVIDIA pulls on open models, from its own filings
Supply. It releases its own open models, Nemotron among them
"in support of market development"
Distribution. It is buying Hugging Face, where open models are shared
about $11.9bn
Demand. It funds and guarantees the companies that buy its chips
$99bn of stakes, $25bn committed, up to $105bn guaranteed
Defence. It argues against rules that would restrict open models
"would favor our foreign competitors"
The payoff. Gross margin on the hardware all of it runs on
75.0%
Read down the column and one idea repeats. Each lever makes the model cheaper or easier to get, and none of them touches the price of the chip. An AI bill has two parts, the model and the compute to run it. Push the first towards zero and the second is all that is left to pay for.
Going to ground
First order
Open models come from many places. China's carry the volume. Not all open models are Chinese. OpenAI publishes gpt-oss, Meta, Google and Mistral publish their own, and NVIDIA ships its Nemotron family. Of the four open models with the most providers on Artificial Analysis in early November, three were Chinese and one was OpenAI's.
The volume is where China stands out. On Vercel's AI gateway, one Chinese model, DeepSeek-V4.1-Flash, processed nearly half of September's tokens. Open models as a group handled about 71% of tokens for 23% of the spend.
Why would Chinese labs give so much away? I can't see inside their strategy, so this is my reading. A lab that can't buy the best chips can't win by selling the best closed model. It can win by making everyone's model cheap. Every free model, Chinese or not, takes pricing power away from the labs that charge for theirs.
Second order
NVIDIA sells the chips that run them. A free model still needs hardware, and much of the running happens outside China. In September nearly 70% of gateway spend on Chinese-origin models went through US-based providers, up from under 10% in December. Chinese models, American data centres, NVIDIA's chips.
NVIDIA took a $4.5bn charge on chips it couldn't sell to China last year and $0.4bn more this year. It lost the Chinese customer and kept the Chinese product. The same holds for a free model from California or Paris: when good models cost nothing, the scarce thing is compute, and NVIDIA earns a 75% gross margin on it.
Third order
Enterprises get cheap AI with a political risk attached. For a company buying AI, this is mostly good news. One study estimates open models cost about 90% less than closed ones of similar ability, and a free alternative is the best negotiating tool against a closed-model vendor. AI starts to look less like a software licence and more like a capacity decision.
Two catches come with it. The price of a model can fall to zero, but the price of running it can't fall below the cost of compute, margin included, so the lock-in moves down from the model to the chip. And some of the cheapest models carry a passport. NVIDIA's own filing warns that the US government "is also considering restrictions" on supporting models built on open models "originating in China".
That second catch is the part that worries me. A workload built on a model that regulators later restrict has to be rebuilt, and no price list shows that cost. The comfort is that open alternatives from outside China exist. The switch would cost money, not the whole strategy.
How I think it plays out
Three paths, and what each means for an enterprise buyer
The alliance holds (my base case). Chinese and other open models take routine work and run mostly on NVIDIA outside China. Closed models keep the hardest work.
Cheaper tokens, bigger compute bill, one dominant supplier
The door half closes. Washington restricts Chinese-origin models, or blocks the Hugging Face deal. Other open models remain.
The cheapest tier goes, rebuild costs, closed vendors regain some pricing power
Open runs anywhere. Open models are portable, so rival and custom chips, Chinese ones included, serve them just as well.
Cheapest outcome for buyers; NVIDIA's own stated risk
My hypothesis is the first path for two or three years, with a real chance of the second. NVIDIA now has $11.9bn of reasons to defend open models in Washington, and on this one question its interest matches China's. If it wins, enterprise spending on AI rises as token prices fall, and more of that spending reaches the chip. If it loses, open models survive without their cheapest tier, and the price of routine AI work goes up.
Either way the practical step is the same. Know where your model came from, and keep the workload portable enough to swap it.
What Tuesday can tell us
Three tests for the results on 17 November
Last quarter
Everyone else's share of data-centre revenue. The thesis needs it to keep rising
45%, growing 138% against 102%
Gross margin. If free models move value to the chip, it holds. NVIDIA guided to 74.0%, give or take half a point
75.0%
China. Any data-centre chip revenue there is a bonus; the forecast assumes none
under 1% of data-centre revenue
This note is written before the numbers, so it can be checked against them. If the share falls and the margin slips below guidance, the thesis is in trouble, and an update here will say so.
Ears up: what to watch
What NVIDIA says on the call about open models, China and the Hugging Face timetable.
US rules on Chinese-origin models, and who lobbies for them.
The Hugging Face deal: regulatory approval, due before a close in the first half of 2027.
Where Chinese models run: the share served by US providers, and any shift to non-NVIDIA chips.
Your own stack: which of your workloads depend on a Chinese-origin model, directly or through a fine-tune.
Bone of contention
Here is where I could be wrong. Not all open models are Chinese, and NVIDIA's strategy would survive without the ones that are. "Alliance" is my word. Nothing shows NVIDIA and Chinese labs coordinating; their interests simply point the same way, and the account of why the labs give models away is a reading, not a fact. The chart shows who buys NVIDIA's chips, not which models they run, and "everyone else" includes labs running closed models.
The evidence is narrow too. The token shares come from one gateway, Vercel, not the whole market. The 1% figure covers licensed data-centre chips; NVIDIA still books other revenue from customers based in China. The $99bn covers all of NVIDIA's equity stakes, not only those in customers. And open models cut both ways for NVIDIA: they run on any chip, which is the risk it writes down itself.
China writes many of the models and can't buy the chips. NVIDIA sells the chips and wants every model free. For now, that suits them both.
NVIDIA, Form 8-K, 2 September 2026: agreement to acquire Hugging Face (about $11.9bn, retention programme of up to about $1.0bn, expected close in the first half of 2027), the commitment to keep the platform open, and the risk factor on open-source modelsconfirmed
NVIDIA, Form 10-Q for the quarter ended 26 July 2026: data-centre revenue by market platform, growth rates, gross margin, equity investments ($99bn, with $25bn committed; non-marketable stakes of $47.9bn against $3.8bn a year earlier), the SB Energy guarantees, licensed China shipments under 1% of data-centre revenue, the H20 and H200 charges, proposed restrictions on Chinese-origin open models, and the open-source statementconfirmed
NVIDIA, Form 10-K for the year ended 25 January 2026: the earlier open-source risk statement, open-model releases "in support of market development", investments in AI model makers, customers' in-house alternatives, and NVIDIA AI Enterpriseconfirmed
CNBC, 11 August 2026: NVIDIA's open Nemotron models and Jensen Huang's public case for open modelsreported
Artificial Analysis, provider leaderboard, observed 4 November 2026: the open models listed with the most providers (GLM-5.3, GLM-5.3-Flash, DeepSeek V4.1 Flash, gpt-oss-120b)confirmed
Data for the chart: nvidia-datacenter-mix.csv. The three paths, and the account of why Chinese labs release open models, are my hypothesis, not a forecast or a finding from any source.