NVIDIA Is Not Buying the Models

I made a video version of this piece, if you'd rather watch it.

What's confirmed, and what isn't

Two outlets did the original reporting. The Information says NVIDIA agreed to acquire Hugging Face for $12.9 billion, citing one anonymous source. Business Insider says talks are real but nothing is signed. Bloomberg ran a headline on August 26 that was really a story about other people's reporting: "Nvidia in Talks to Buy AI Startup Hugging Face, Reports Say."

Everyone else is just repeating those two.

Neither company has confirmed anything. On August 27, Bloomberg asked Hugging Face co-founder Thomas Wolf directly. He declined to comment and said only that the company has always attracted interest and offers.

Hold onto that line. It matters more at the end of this piece than it does here.

The backstory is confirmed, and it's the more interesting half anyway. The Financial Times reported that Hugging Face turned down a $500 million investment from NVIDIA in late 2025, at a $7 billion valuation, because it didn't want a single dominant investor in the room.

A company that said no to half a billion dollars on principle is now reportedly selling outright for roughly twenty six times that amount.

Something changed. Not the models.

One number reframes the entire deal

Hugging Face is running at roughly $150 million a year in revenue. That's an outside estimate, not a company figure. Treat it as directional, not precise.

Take it at face value anyway. $12.9 billion divided by $150 million comes out to about 86 times revenue.

For context on who's paying it: NVIDIA reported second quarter fiscal 2027 revenue of $96.2 billion, up 106% from a year ago, with data center revenue of $89.0 billion, up 117%. It's sitting on roughly $99 billion in cash and marketable securities.

This isn't a company that overpays by accident. When an operator with that balance sheet pays 86 times revenue, the revenue isn't the asset. Something else on the platform is worth the money, and the price is the tell.

The obvious explanation is right, but it doesn't explain the price

NVIDIA sells the compute. Hugging Face is the Google Drive of developers, hosting roughly 2.5 million public models and more than 700,000 public datasets across tens of thousands of organizations. NVIDIA owns the factory. It just bought the storefront.

That's true. It's also the version you get from the headline, and it doesn't explain 86x.

Three things do. None of them are the models.

Reason one: it buys intent, months before the purchase order

Every download on Hugging Face is a developer telling you exactly what they care about, and by direct implication, how much GPU they're going to need.

NVIDIA never had those numbers. It sees a purchase. It sees an order. Both of those show up at the end of a decision made months earlier, by some engineer trying a model out on a laptop.

Buying Hugging Face moves NVIDIA from the invoice to the intent. It stops selling shovels and buys the map of where everyone is digging.

People wave at "data" here without saying what they mean. So be specific: the value isn't data in the abstract. It's a forward indicator on demand for a product with an eighteen month manufacturing lead time. If you're allocating fab capacity, knowing which architectures are gaining traction two quarters early is worth more than most acquisitions on the market.

Reason two: betting on the open-weight wave, and winning early

Jensen Huang has spent this year publicly pushing free, open weight models. People read that as generosity.

It isn't generosity.

Look at what happened on Vercel's AI Gateway. In June, open weight models were 28% of tokens. By August, they were 62%.

That's one gateway, not the whole market. But the direction isn't subtle.

NVIDIA doesn't care which model wins. It cares that the world keeps building more of them, because every one of them runs on something, and NVIDIA sells that something. A proprietary model concentrated inside three labs is three customers. Two and a half million open models are a market.

Owning the place those models live is owning the on-ramp to the market you're betting on.

Reason three: defense

Google has TPU. Amazon has Trainium. Microsoft has Maia. OpenAI is designing chips with Broadcom. Anthropic has confirmed its own silicon effort.

NVIDIA's biggest customers are all quietly trying to stop being customers.

You can't stop that with a better GPU. The whole point of a custom accelerator is that it doesn't have to be better in general. Only better for one workload, at one company's scale.

What you can do is make the ecosystem so large, and so anchored to your hardware, that leaving costs more than staying. Buying the distribution hub is how you do that. It's the Apple move: the silicon is the moat, but the ecosystem is what makes the moat expensive to cross.

I used to tell VCs there was no moat in software. I was wrong.

Before MeDucation AI, I sat in a lot of rooms with investors and said the same thing with total confidence.

A founder would pitch a software product, and I'd think: there's no moat here. Anyone can build this. The code isn't defensible.

I was wrong, and it cost me time I'd like back.

Building software is the easy part. It was easy then. It's close to free now. What's hard, and what almost nobody can copy, is distribution: being the place people already go, already trust, and already have their work sitting in.

That's the whole lesson sitting inside this deal, and it's why I couldn't let it go.

NVIDIA isn't paying $12.9 billion for two and a half million model files. You can download those for free right now. It's paying for the position those files sit in, and for being the default door a developer walks through on the way to buying compute.

Distribution was the moat the whole time. It just took NVIDIA spending $12.9 billion to prove it to me.

The 2023 round was designed to prevent exactly this

In 2023, Hugging Face raised $235 million. Salesforce led the round. Look at who else got in: NVIDIA, AMD, Intel, Qualcomm.

Four direct competitors. They don't normally show up on the same cap table. Each took a small piece specifically so none of them could own it.

That was the whole design, and it worked, right up until it didn't. It was a handshake, not a contract. Neutrality that lived on a cap table. And cap tables are for sale.

The damage won't look like sabotage

Nobody expects NVIDIA to wake up and delete support for AMD chips. It won't happen, and framing the risk that way makes it easy to dismiss.

A popular model on Hugging Face has dozens of variants, each tuned to run fast on different hardware. Whichever one surfaces first is the one most developers use. That's a ranking decision, and the owner of the platform makes ranking decisions.

The tooling that makes models run well on AMD and Intel doesn't disappear. They just rank lower. The recommendation that says this runs better on NVIDIA hardware sits above the one that doesn't.

No announcement. No blog post. No policy anyone can point to in a complaint. Just a default, applied a few million times.

That's the part I'd want a regulator to ask about. It's also the part that's hardest to write a remedy for.

NVIDIA has already lost this exact fight once

In September 2020, NVIDIA announced it was buying ARM, the British company whose chip designs sit in nearly every smartphone on Earth, for about $40 billion.

Regulators hated it, for a reason that should sound familiar. ARM licensed its designs to everyone, including NVIDIA's competitors, so NVIDIA owning ARM meant NVIDIA controlling something its rivals depend on.

On December 2, 2021, the FTC sued to block it, calling it a vertical deal that would let the combined firm stifle competing next generation technologies. On February 14, 2022, seventeen months after announcing it, NVIDIA walked away. The FTC issued a statement on the termination.

Now read the Hugging Face deal against that. Same buyer. Same structural objection: a supplier acquiring the neutral layer its competitors depend on. The main difference is that ARM's neutrality was contractual. Hugging Face's was a gentleman's agreement between four investors.

There's a real chance this never happens

NVIDIA has spent roughly $27 billion on AI deals this year and has carefully avoided the word acquisition. The Groq deal, around $20 billion, was structured as a license plus people. Poolside was roughly $6 billion to license the software, plus about a billion for equity, plus around a hundred engineers hired.

That structuring isn't an accident, and regulators have noticed. Two senators wrote to Jensen Huang in March asking whether the Groq deal was built to delay the antitrust clock. The FTC chair had already said the agency was examining whether deals like it were structured to escape review.

Against that backdrop, NVIDIA now wants to buy the distribution layer of open source AI outright. An actual acquisition. A price attached.

My honest expectation: if this is confirmed, it spends most of 2027 in regulatory review before anyone learns whether it survives.

The question I can't answer

Everything above is analysis I'm reasonably confident in. Here's the part I'm not.

Do regulators see a model hub as infrastructure, or as a website?

If it's infrastructure, this is ARM again, and it doesn't close. If it's a website, it's a large company buying a popular developer tool, and it closes without much drama.

I don't know which way that goes, and I don't think anyone does yet. The honest answer is that the framing gets decided by lawyers arguing about what Hugging Face fundamentally is, not by anything in the technology.

That single classification question decides a $12.9 billion deal and, more importantly, sets the precedent for every platform acquisition that follows it.

If you work in antitrust, infrastructure policy, or you've watched a platform ranking change reshape a market from the inside: I want to hear which way you think that classification lands, and why. Not whether the deal is good or bad. Just infrastructure or website.

That's the whole thing.

The full video breakdown is here.

Frequently asked questions

No. Neither company has confirmed anything. The Information says NVIDIA agreed to acquire Hugging Face for $12.9 billion, citing one anonymous source, and Business Insider says talks are real but nothing is signed. Everyone else is repeating those two. Asked directly by Bloomberg on August 27, 2026, co-founder Thomas Wolf declined to comment and said only that the company has always attracted interest and offers.

Because the revenue is not the asset. Hugging Face runs at roughly $150 million a year by outside estimate, so $12.9 billion is about 86 times revenue. Three things justify that price, and none of them are the models: the platform is a forward indicator of GPU demand months before a purchase order, it is the on-ramp to the open-weight wave NVIDIA is betting on, and it is a defense against NVIDIA's largest customers building their own chips.

Intent, months earlier than the invoice. Every download is a developer signalling what they care about and, by implication, how much GPU they will need. NVIDIA currently sees only the purchase order, which arrives at the end of a decision made months earlier. For a product with an eighteen month manufacturing lead time, knowing which architectures are gaining traction two quarters early is worth more than most acquisitions on the market.

There is a real precedent. NVIDIA announced a roughly $40 billion acquisition of ARM in September 2020, the FTC sued to block it on December 2, 2021 as a vertical deal that would stifle competing next generation technologies, and NVIDIA walked away on February 14, 2022. The structural objection is the same here. The main difference is that ARM's neutrality was contractual while Hugging Face's was a gentleman's agreement between four investors.

Not sabotage. A popular model on Hugging Face has dozens of variants tuned for different hardware, and whichever one surfaces first is the one most developers use. That is a ranking decision, and the platform owner makes ranking decisions. Tooling for AMD and Intel would not disappear, it would simply rank lower. No announcement, no policy to point to in a complaint, just a default applied a few million times.