The Advantage That Leaves No Document

Nvidia agreed on 3 September to buy Hugging Face for $12,930,300,000. In the post announcing it, Jensen Huang writes one precise sentence: Nvidia compute will not be required to build on or deploy through the platform. That is a promise about a requirement. Requirements were never the thing that mattered on that platform.

On 24 July, Hugging Face had signed a letter from Nvidia. It is called Open Weights and American AI Leadership, Huang promoted it with the first post of his life on X, and twenty-five companies signed it: Nvidia, Microsoft, Meta, IBM, Dell, Palantir, the Linux Foundation, Hugging Face. OpenAI, Anthropic and Google were absent. Tom's Hardware reduced the document to one line worth the whole of it: none of the signatories sells access to a closed frontier model. The letter asks Washington to avoid premature restrictions on downloadable models, and in a single paragraph, the only one that is not about open weights, asks that distillation not be treated as misappropriation. Six weeks later, one of the twenty-five signatures bought another.

My hypothesis is that the letter was part of the operation, and I declare it as a hypothesis because no source supports it. The sequence makes it hard to avoid. A year ago Hugging Face turned down a $500 million investment from Nvidia. On 24 July it signs the letter. Over the summer, by his own account to CNBC, Clément Delangue goes to Huang to propose selling the whole company.  Asking Washington to leave downloadable models alone and then buying the place they are downloaded from are two moves that hold each other up: the first sets the regulatory weather in which the second is worth more. That they hold each other up is no proof they were planned together. Treating them as coincidence requires believing that the lead signatory of a letter on the strategic value of open weights had not yet priced their warehouse.

The models stay where they are. What changes hands is the infrastructure around them: the most complete documentation of model development anywhere, the Transformers library that vLLM and SGLang depend on while competing with Nvidia's own TRT-LLM, the local inference engine llama.cpp, which joined Hugging Face this year, and a compute layer served by AMD, Cerebras, SambaNova and Groq. None of that produces an act.

An advantage that leaves no document is delivered through documentation, the example that works on the first try, the quickstart that runs without errors, the compute credit for the university course. It requires excluding nobody. It does not even require a decision: it is enough that the best resources keep arriving, month after month, where they are cheapest to make. Europe has just settled that a high-risk credit decision must be documentable and contestable by 2 December 2027, as I wrote about the algorithm that has to show its work. About a default, nobody has settled anything.

My second declared hypothesis concerns the target, and it cuts against the current reading. Nvidia is said to have moved against the proprietary APIs of OpenAI and Google. Look instead at who buys from Nvidia. In the quarter it reported to the SEC in August 2025, two unnamed customers accounted for 23 and 16 percent of total revenue and four more for 14, 11, 11 and 10: six buyers, 85% of the quarter, every one of them building silicon of its own. Whoever downloads open weights and serves them on enterprise clusters, regional clouds and racks on the premises has no alternative programme to put into production. Buying the warehouse buys the channel to the customers who cannot defect. Yesterday I wrote that 80 % of OpenAI's and Anthropic's enterprise revenue comes from 1 % of their customers, and called that a patron list. Nvidia's own filings describe the same shape one layer down. Nobody lined these numbers up for me, and the reading is mine.

Then again, Hugging Face's costs are real. Storing petabytes of models and datasets costs money, serving them to eighteen million people costs more, and a warehouse that runs out of cash would be concrete damage to anyone writing software today. For ten years the people building on it treated it as public infrastructure, and for ten years it had investors and a board: it could have been sold on any day of those ten years. Funding and owning are written in two different instruments, and only one of them was ever signed.

The way neutrality changes hands I had already seen in Anthropic's oversight trust, which bars its trustees from owning shares and says nothing about being hired by the company they watch. The form here is the same and the scale is larger. According to The Register, at present the only outlet reporting it, one billion dollars is earmarked for the Hugging Face employees moving to Nvidia. The people who hold the neutral ground go onto one party's payroll, with an incentive to stay there.

Someone writing software will pick a card two years from now because the first example they saw ran on it, and will not know they chose. The regulator who opens a file in 2028 will look for an act. An advantage that leaves no document produces none. They will find better documentation.