NVIDIA Is Reportedly Paying 86× Revenue for Hugging Face. It Isn't Buying the Revenue

The reported $12.9 billion price set against roughly $150 million of revenue, giving an implied multiple of about 86 times revenue and 2.9 times the 2023 valuation, beside what that price actually buys — the default home of open weights with over two million models and thirteen million developers, a hedge against customers moving to custom silicon, and a route back into cloud — above a note that the deal is reported, unsigned and uncommented — At roughly 86 times revenue, the earnings are not the argument. The argument is who owns the place every open model is downloaded from.

Start with what is actually established, because the coverage has been looser than the facts.

The Information reported on the night of 26 August 2026 that NVIDIA has agreed to buy Hugging Face for $12.9 billion. Business Insider had reported over the preceding weekend that Hugging Face was approached at a valuation above $13 billion, and specified that the talks had not produced a signed agreement and could still fall apart.

As of writing, neither company has confirmed anything. Both declined to comment or did not respond. TechCrunch noted that NVIDIA's silence is itself unusual, since the company normally rebuts reports it disputes quickly — which is an inference, not a confirmation, and should be read as one.

So: a credible report, from outlets with good records, that is not a done deal. Do not restructure anything on it this week.

1. The number that tells you what is being bought

The Information also reported Hugging Face's revenue: roughly $150 million a year, up from about $100 million just two months earlier. CEO Clem Delangue has previously said the company was close to profitability.

Put those together.

$12.9 billion against $150 million of revenue is roughly 86 times revenue.

Nobody pays 86× for earnings. For comparison, Hugging Face's last disclosed valuation was $4.5 billion, set in a $235 million Series D in August 2023 led by Salesforce Ventures — a round NVIDIA itself participated in. The reported price is about 2.9× that, three years later.

A multiple like that is a statement about strategic position, not about a business. Which raises the useful question: position in what?

2. What NVIDIA gets

Three motives, in descending order of how much I believe them

Why a chip company buys a model hub

A hedge against its own customers. OpenAI, Google, Amazon and Anthropic are all building their own silicon. Every one of those is a large customer working to need NVIDIA less. Open-weight models are the counterweight: they get downloaded by people who then have to run them on hardware they buy themselves, and that hardware is overwhelmingly NVIDIA's. Owning the distribution point for open weights is a direct hedge against the concentration risk in NVIDIA's own customer list.
A route back into cloud. Hugging Face already helps developers run models on rented compute. That is a foothold in a market NVIDIA has circled for years without a clean entry.
Somewhere to sell uncommitted capacity. Reported as a motive: NVIDIA could route unused compute from customer commitments toward Hugging Face's user base. Plausible, and the least load-bearing of the three.

The first motive is the one that explains the price. NVIDIA's risk is not that AI demand falls; it is that the biggest buyers of its chips design their way off them. A vibrant open-weight ecosystem running on commodity-purchased GPUs is the best available insurance, and Hugging Face is where that ecosystem gets its software.

3. The dependency you never chose

Here is the founder-relevant part, and it has nothing to do with whether the deal closes.

If you build with open models, your stack almost certainly contains a line like from_pretrained("...") that reaches out to one company's servers. Weights, datasets, tokenizers, evaluation sets, and increasingly inference endpoints — for most teams, all of it flows through the same hub. The scale being cited is over two million models, more than 500,000 datasets, and 13 million-plus developers.

This is the same shape as last week's story

Yesterday I wrote about a CVSS 5.3 path-traversal bug in a package registry becoming the entry point to a root compromise, and made the argument that a registry is production infrastructure rather than plumbing. This is that argument again with the ownership question attached instead of the security one. The model hub is the most load-bearing dependency in a modern AI stack that almost nobody selected on purpose — it was simply where everything already was.

Notice that this was true on Monday, before any of this was reported. The acquisition does not create the dependency. It just makes it visible, which is the only genuinely useful thing about it.

4. What would change, and what would not

I want to be careful here, because "neutral platform acquired by hardware vendor" invites a lot of confident prediction that is really just vibes.

The open weights themselves

Probably unchanged

Licences on published models do not change hands with the platform. Weights already downloaded stay downloaded, and models released under permissive licences remain mirrorable elsewhere. The artefacts are not the leverage point.

Defaults and optimisation paths

Worth watching

The realistic risk is not censorship or paywalls; it is gravity. Which runtimes get first-class support, which quantisations are published, which hardware the "recommended" inference path assumes. Defaults move ecosystems far more effectively than restrictions do.

Neutrality as a hosting venue

Worth watching

Hugging Face currently hosts models from labs that compete directly with NVIDIA's customers and, increasingly, with NVIDIA. A neutral hub owned by no one in the race is a different institution from a hub owned by the largest seller of picks and shovels.

Regulatory friction

Genuinely uncertain

A $12.9 billion acquisition by the most scrutinised company in tech, of the distribution layer for open models, is not obviously a quick clearance. Timeline risk is real and I would not model a close date.

5. What to do about it

Almost nothing this week. But there is one piece of work worth an afternoon, and it was worth an afternoon before the news broke.

Where your weights come from

Write down

List every model, dataset and tokenizer your product pulls, and where it is pulled from at build or run time. Most teams have never written this down and are surprised by the length. That list is your actual supply chain.

What you cannot rebuild

Cache

Pin versions and mirror the weights you depend on into storage you control. This is ordinary supply-chain hygiene that protects against an outage, a licence change, a deletion or an acquisition equally — and it costs a morning and some object storage.

Rewrite your stack

Do not

Migrating off a hub on the strength of an unconfirmed report is a bad trade. The hub is excellent, the deal is unsigned, and the fix for platform risk is portability, not preemptive exit.

The distinction that matters: portability is cheap and always worth having; migration is expensive and should be triggered by something that actually happened. Right now nothing has actually happened except a report.

6. The wider pattern

This is the third consolidation story worth a founder's attention this month, after every major agent protocol moved under one foundation and Stripe's move on model routing.

The direction is consistent: the connective tissue of the AI stack — protocols, routers, registries, hubs — is being bought or governed by parties with strong commercial interests in what flows through it. That is not a conspiracy, it is what happens when infrastructure turns out to be valuable. But it does mean the era of assuming the middle layers are neutral is ending, and the reasonable founder response is to know what you depend on rather than to assume good behaviour.

Hugging Face, incidentally, was also the company an OpenAI agent breached in July. It has had an eventful quarter as the place everyone's models live.

The honest summary

NVIDIA has reportedly agreed to pay $12.9 billion for a company with about $150 million in revenue. That is roughly 86× — a number that only makes sense if what is being bought is the default distribution point for open-weight AI, and the insurance policy it represents against NVIDIA's largest customers designing their own chips.

Nothing is signed. Neither company has said a word. The deal could evaporate, and any regulator looking at it will take time.

The part that is true regardless: the model hub is probably the single most load-bearing dependency in your stack that you never made a decision about. Write down what you pull from it, mirror what you cannot rebuild, and leave the migration question alone until something is actually announced.

Sources: The Information, "Nvidia Agrees to Buy Open Source AI Platform Hugging Face For $12.9 Billion" · TechCrunch, "Nvidia closes in on Hugging Face acquisition" · TechCrunch, "Hugging Face reportedly in talks to be acquired for $13B" · CNBC · Bloomberg · Price and revenue figures are as reported by The Information and Business Insider and are not confirmed by either company; hub scale figures are widely reported third-party estimates; the multiple arithmetic and the argument are mine.

An AI Agent Breached a Real Company to Cheat on a Test: The Founder's Read on the Hugging Face IncidentJuly 30, 2026

A2A Just Moved In With MCP. Standardising How Agents Talk Is Not Standardising Whether They're RightAugust 20, 2026

The AI Bubble Question Is Being Asked at the Wrong Altitude: A Founder's Guide to Which Layer Actually DeflatesAugust 3, 2026

IdeaToMVP Academy

Want to build with AI — not just read about it?

4-week live cohort for founders. Learn to ship AI agents, scope MVPs, and automate your business — taught by the same team that writes these guides.

Explore the Academy →