The first offer was $500 million. It was rejected.
The November price is $12.93 billion. Twenty-six times, same year, same company.
The agreement was signed September 2, with closing expected in the first half of 2027. At an all-hands meeting on November 18, management called it a plan and promised Hugging Face would stay open.
The 26x is not about technology
Priced on technology, 26 times makes no sense.
Hugging Face builds no models. It hosts, indexes, retrieves: millions of files on a shelf where a developer picks one by task, parameter count or license. That layer produces nothing new. It decides which model gets seen, compared, chosen. The inventory belongs to other people. The ordering belongs to the shelf. The code is open source; the entrance is not. Code can be rewritten. Habits cannot.
A developer editing a model opens this shelf before an editor.
An entrance is worth nothing as a URL and a great deal as a default. Defaults come from the shortest path, not from advertising. Search a model name, it is first. Open a leaderboard, the link goes there. Hunt for a quantized build, the repository is still there. Three paths, one domain, and the default exists. No contract is signed. It only has to save two clicks.
$500 million buys a team and revenue. $12.93 billion buys the default. Same asset, two ways of booking it, one as headcount, one as distribution. Nothing changed in the accounts. Everything changed in the strategy.
The shelf holds other people's goods
NVIDIA (NASDAQ: NVDA) sells compute, and compute is not near the developer. The GPUs sit in a data center; the models sit in a browser.
Whoever decides which model gets downloaded decides which toolchain, precision and kernel library come next. All of it lands on one question: what hardware it runs on. That is the platform move: from selling parts to setting the standard.
The timing says something too. NVIDIA's stock is down about 37% from its 52-week high, and Jensen Huang stepped down as president, CEO and director on November 12 with no successor named. A company in transition buying an entrance this size is pinning the next chief executive's direction.
The long tail matters more. Millions of models cannot all be optimized; the ones that get used decide what inference demand looks like. The shelf does not write that list, it bends it: homepage placement, leaderboard order, default precision. Wherever the list leans, optimization follows.
Download counts are the earliest signal of inference demand. Nobody tunes the long tail. Whoever tunes it first takes the default.
Half a month of gross profit
Last quarter NVIDIA reported $115.84 billion of revenue, $107.6 billion from data centers, at a 74.8% gross margin: about $86.6 billion of gross profit a quarter, near $28.9 billion a month. The deal costs less than half a month of it.
Total guarantees are about $118 billion, nine times the deal. The chip financing platform launched with institutional partners in August is $500 billion, thirty-eight times. Buying an entrance with 11% of the guarantee book is cheap. The rounding error buys position.
To the seller the same money is a 26x premium. To the buyer it is 13 days of gross profit. Both work, which says the purchase is not about cash flow. The premium buys the default adaptation slot at every future model release.
The filings carry no revenue figure for Hugging Face. Payback cannot be calculated, only position valued. This is not how a financial acquisition is priced.
Six years, and a closing date
One clause is easy to miss: the three founders signed six-year retention agreements.
Six years is longer than most lock-ups. What is bought is not the code but those three people and their judgment. The real risk to a shelf is a new operations team changing review standards, recommendation logic and community rules. Developers will not protest; they will go elsewhere.
Six years is a stability statement for developers and a signal to regulators: the people are still there, the rules will not be rewritten overnight.
Signing to closing spans three quarters or more. Hugging Face stays independent, and its customers include NVIDIA's competitors. A chip company buying the door developers use to find models will draw antitrust attention. The outcome turns on how regulators define the market: silicon, model hosting, or the developer entrance. Three definitions, three verdicts. Every quarter of delay adds uncertainty and gives rivals room for a second shelf.
I will not guess whether it clears.
What an entrance costs
An openness pledge does not settle ranking. Ranking is power. Homepage placement, search order, default precision: none of it is covered by "open," and all of it decides whether a model is seen. Developers believe they choose freely. The options were ordered before they arrived. Entrances are sticky as well. Weights can move; download history, version history and discussion threads cannot. Nobody wants to move twice.
A door can be monetized three ways: hosting fees from enterprises, priority optimization fees from model makers, and insight into developer behavior. The first two fit on a price list. The third is quietest and worth most: whoever knows what developers download, modify and deploy holds the next generation's chip requirements in advance.
Admission has limits. Charge too much and labs move their repositories. Charge too little and the payback on $12.93 billion stretches out. The line between open and paid is not in the contract. It is in every ranking.
$12.93 billion buys the key. The goods are not NVIDIA's. The door is.
NVIDIA says the platform will stay open. How open will be visible after closing.
A deal that has not closed has already changed what developers expect the default entrance to be.
That may be the most expensive part.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
