Hugging Face Just Sold for a Lower Multiple Than It Raised At

The Information reported Wednesday night that Nvidia agreed to buy Hugging Face for $12.9 billion. Business Insider, which broke the story over the weekend that Hugging Face was taking calls, puts the talks above $13 billion and says nothing has been signed. Neither company has commented, which is mildly unusual for Nvidia, since they normally push back fast on reporting they think is wrong.

Hugging Face is doing roughly $150 million in annualized revenue. That makes this about 86 times revenue, and most of the commentary since has been some version of "well, there's your bubble."

I think the opposite is true. Nvidia got a discount here, and you can see it in the numbers rather than the vibes.

Start with the round nobody is mentioning

In August 2023, Hugging Face raised $235 million at $4.5 billion post-money. Salesforce Ventures led. Google, IBM, Amazon, Qualcomm, AMD, Intel and Nvidia all participated. The company was doing something like $30 million in ARR at the time, which prices that round somewhere between 100x and 150x depending on whose estimate you trust.

Nvidia is paying 86x.

So the multiple came down at the same time that revenue grew roughly five times over, Clem Delangue started telling reporters the company was close to profitability, and the strategic stakes went from "nice open-source hub" to "the layer that decides whether open weights stay competitive." Buyers don't usually get a cheaper multiple and a control premium in the same transaction. When they do, it's normally because the seller wasn't running a real auction, and I haven't seen anyone make that point about this deal yet.

The revenue figure is already stale

Look at how Hugging Face got to $150 million. Sacra has them at about $53 million exiting 2024 and about $81 million exiting 2025. The Information reported $100 million in June of this year, then $150 million in August. That is fifty percent growth in two months.

Obviously that doesn't hold forever. But you don't need it to. If Hugging Face just doubles over the next twelve months, Nvidia bought at 43x forward, and if it does anything close to what it did over the last twelve, this prices in the twenties. Ordinary numbers for infrastructure that everybody in a category depends on.

It's also worth looking at where the money comes from. Storage, GPU rental, hosted inference endpoints, enterprise deployments. Unglamorous stuff that scales with the raw volume of open-weight activity, which is a far better thing to own than exposure to any particular model release.

The comp is eleven days old

The cleanest test of whether $12.9 billion is expensive comes from a deal that happened earlier this month. Bloomberg reported that Stripe finalized an agreement to buy OpenRouter for more than $7 billion. Axios later put the figure in Stripe's investor letter north of $8 billion, mostly stock. The Information had OpenRouter at roughly $140 million annualized in late July. Call it 50x, struck less than three months after a Series B that valued the company around $1.3 billion.

OpenRouter is a routing and metering layer. It sits on top of models, takes a fee on credit purchases, and points traffic at four hundred or so endpoints. It's a good business, and Stripe is right that whoever counts the tokens ends up owning the spending relationship.

Hugging Face is the thing underneath all of that. The Hub holds over two million models and around 1.5 million datasets, and the transformers and datasets libraries are load-bearing imports in a huge share of the ML code written since 2019. When DeepSeek or Qwen or Mistral ships weights, they ship them there first. When an enterprise team pulls a model into production, the canonical URL is a Hugging Face URL.

Nvidia paid about 1.6 times what Stripe paid and got the road instead of the toll booth.

It looks stranger when you widen out. Nvidia paid around $20 billion in December for Groq's technology and people, and roughly $6 billion to license from Poolside. It has now paid less for the distribution layer of open AI than it paid for an inference chip team.

What this actually costs them

Nvidia reported fiscal Q2 the same night the story broke. Revenue of $96.2 billion, up 106 percent. Net income of $59.7 billion, up 126 percent. Data center alone did $89 billion at 75 percent gross margins. They guided Q3 to $108 billion and told analysts to expect something like 70 percent revenue growth next fiscal year.

$12.9 billion is a quarter of one percent of a $5.1 trillion market cap, and 20 days of Q2 net income. For scale, the buybacks and dividends Nvidia paid out last quarter came to $26 billion, twice what it is spending here.

Nvidia also disclosed it's sitting on $47.9 billion in private company stock, with another $18 billion committed to equity investments before the year is out. This isn't a bet-the-company allocation, and judging it like one misses what's going on. What matters is whether it moves anything at the scale Nvidia operates at now, not whether it clears some revenue-multiple hurdle.

The insurance policy

The strategic case has very little to do with Hugging Face's P&L. Every one of Nvidia's biggest customers is building silicon to depend on them less. OpenAI and Broadcom were at Hot Chips this month claiming their inference part beats Blackwell-class racks. Google has TPUs. Amazon has Trainium. Anthropic is spreading its bets. Customer concentration is the single biggest threat to Nvidia's own multiple, and everyone inside the company knows it.

Open weights are the counterweight. Every enterprise that fine-tunes an open model instead of buying an API from a lab with custom silicon is demand that stays diffuse and stays pointed at CUDA. Sovereign programs, regulated industries that can't ship data to a closed endpoint, startups that self-host. That's Nvidia's long tail, and open weights are how it gets fed.

They have not exactly been hiding this. Huang's own quote in Wednesday's earnings release lists "a thriving open-model ecosystem" among the things driving demand. His first post on X, last month, was a defense of open models. He signed the industry letter asking Washington not to restrict open weights, and Delangue signed the same letter. Nvidia has already spent tens of billions building its own open models. What that money couldn't buy was the place developers go to find them.

So price it as insurance. Nvidia's data center revenue is running near $356 billion annualized. If owning the distribution point for open models protects one percent of that, the purchase price comes back inside four years, and one percent is a conservative guess at what a hostile shift in open-model gravity would cost them.

Two assets riding along for free

A couple of things come with this that I haven't seen priced anywhere. One is a cloud business Nvidia already tried and failed to build. They scaled back DGX Cloud about a year ago. Hugging Face runs a compute rental and managed endpoint business that already has demand attached to it, which is a way back into that market without starting from a standing stop.

The other is a hedge on a liability they already carry. Nvidia has agreed to help cover tens of billions of dollars in customer cloud commitments, and if those customers underconsume, Nvidia eats it. Owning a platform where a million-plus developers rent GPUs gives them somewhere to put stranded capacity. That option is worth real money and it didn't cost anything extra.

The seller set the price

Nine months ago Hugging Face turned down $500 million from Nvidia at a $7 billion valuation. Per the FT, Delangue didn't want a dominant shareholder who could push the company around. The business wasn't in trouble then and it's in less trouble now. It was growing fast, near breakeven, and under no obligation to do anything at all.

Which means $12.9 billion is roughly the floor that moved a founder who had already proved he'd rather stay independent. Nowhere near the ceiling Nvidia would have paid to keep the open-weight hub out of a competitor's hands. When the buyer's reservation price sits that far above the clearing price, the deal is cheap almost by definition.

Where we could be wrong

The big risk is neutrality. Hugging Face works because it's Switzerland, and a chip company with a direct financial interest in developer lock-in owning that hub is a contradiction that has to resolve one way or another. If AMD or Google or the community routes around it with a fork or a foundation-run registry, Nvidia bought a brand that starts depreciating on day one. The moat here is trust, and it won't survive the first time a hardware preference shows up in the rankings.

There's a regulatory question too. Rival chipmakers and clouds have an obvious argument about a dominant supplier owning the neutral distribution layer for models that run on its hardware.

And the revenue quality is worth a look. A large chunk of Hugging Face enterprise revenue reportedly comes from managed deployments sold to Amazon, Microsoft and Nvidia itself. Some of that gets awkward after close.

Standing caveat: there's no signed agreement as I write this, and Business Insider says the talks could still fall apart.

86x is the wrong number

It prices Hugging Face as a SaaS company, which it has never been. What Nvidia bought is the registry for open AI, at a quarter of a percent of market cap, funded by 20 days of profit, as a structural hedge against its four largest customers going vertical. No revenue multiple settles whether that was smart. What settles it is whether the people who made Hugging Face what it is decide to stay.

Microsoft paid $7.5 billion for GitHub in 2018 and took the same criticism, and that one aged fine. But GitHub didn't have a community that could stand up a credible mirror over a long weekend. Hugging Face does. If they walk, Nvidia spent 20 days of earnings on a logo.