Nvidia Paid 86× Revenue. It Wasn't Buying Revenue — It Was Buying the Default.

Hugging Face's ARR passed $100 million in July. Nvidia agreed to pay roughly $12.9 billion. That's about 86× revenue.

No financial model produces that number. So the honest question isn't whether Nvidia overpaid — it's what Nvidia thinks it bought, because it clearly wasn't a P&L.

Here's my answer: it bought the place where developers decide which hardware their models run on.

And one commitment worth reading carefully, because it's in the filing rather than a press release: Nvidia has committed to keep the platform open — to continue permitting model makers, developers and users to upload and download models and datasets of their choosing, and to support other silicon vendors.

That is a more specific promise than most acquirers make. Hold onto it; we'll come back to it.

A correction on scale, and it works in your favour

"Over a million models" undersells this considerably.

Hugging Face hosts more than 2.9 million public model repositories, alongside roughly 1 million datasets and 1.44 million hosted demo Spaces, used by more than 13 million developers.

That's not a repository. It's the index of open AI — the thing a from_pretrained() call resolves against, the first search result when anyone goes looking for a model that already exists.

What "the default" actually means

Owning Hugging Face doesn't let Nvidia block anyone. It doesn't need to.

It lets Nvidia decide ordering. Which runtime a model card mentions first. Which quantisation is offered by default. Which inference provider sits one click away from the download button. Which optimised kernels surface in search.

None of that is exclusion. All of it is influence over the moment a developer makes a hardware-adjacent choice without realising they're making one.

Nvidia's competitors don't compete for chip sales in a purchasing meeting. They compete for them in a developer's first hour with a model. Nvidia just bought that hour.

The Groq deal makes it a pattern

Look at the last nine months together.

→ December 2025: Nvidia buys Groq's assets for about $20 billion — its largest deal ever, taking out an inference-chip challenger founded by the creators of Google's TPU → September 2026: Nvidia agrees to $12.9 billion for Hugging Face — the distribution layer for open models

Roughly $33 billion in nine months, and neither deal buys a better GPU.

For context on how large a break this is with Nvidia's own history: before Groq, its biggest acquisition was Mellanox at about $7 billion in 2019.

Nvidia isn't defending its chip lead. It's buying the layers above and below it — the places where substitution decisions get made. Seen that way, $12.9 billion for $100 million of revenue is not a valuation error. It's insurance priced against the possibility that models stop caring which silicon they run on.

The neutrality question — and the precedent

Your instinct about neutral ground is the right one. There's a recent case study that makes it concrete.

In June 2025, Meta took a 49% non-voting stake in Scale AI. Google moved to cut ties. OpenAI began phasing out its work. Not because the labelling got worse overnight — because Scale's actual product was neutrality, and one investment ended it. Mercor inherited the customers.

Now apply that. AMD, Google, Amazon, Microsoft, Meta and Qualcomm all build silicon that competes with Nvidia. Nvidia has committed, in a public filing, to support other silicon vendors on the platform.

I take that commitment at face value. I also think it may not be the variable that matters. Neutrality isn't decided by the owner's conduct — it's decided by the counterparty's perception. The question isn't whether Nvidia behaves. It's whether AMD wants its optimised kernels discovered through Nvidia's storefront, and whether a Chinese lab wants its weights hosted on infrastructure owned by a US chip company subject to export controls.

Those are answered by other people, and mostly not in public.

The risk nobody's pricing

This deal is signed, not closed. H1 2027, subject to regulatory approval.

Nvidia has been here before. Its $40 billion Arm acquisition collapsed in early 2022 under FTC and CMA opposition, on essentially this theory: a dominant chip company shouldn't own the neutral layer that its competitors depend on.

A dominant AI chip company buying the neutral distribution hub for the models that run on AI chips is a cleaner version of that same argument. I'd put meaningful probability on this taking longer than H1 2027, arriving with behavioural conditions attached, or not arriving at all.

My read

This is the most strategically sophisticated acquisition of the year, and the multiple is the tell. Nvidia paid 86× revenue for something that generates almost no revenue, because the asset was never the revenue — it was the position.

The open-source community's concern is legitimate but slightly misaimed. The risk isn't that Nvidia closes the platform. It's that "open" and "neutral" turn out to be different properties, and this deal only guarantees the first one.

Question for you: if a platform stays completely open but its owner competes with half the people using it, is that still neutral ground? Scale AI's customers answered that question in about a month, and they answered it with their feet.

Address

Boston, Massachusets

Contacts

news@techproduit.com

Get your weekly newspaper