NVIDIA Bought the Shelf: Why Hugging Face Was the One Deal It Couldn't License

Two weeks ago NVIDIA agreed to buy Hugging Face for $12.93 billion — the second-largest acquisition in its history, behind only what it paid for Groq's technology at the end of last year. This morning, the New York Times and Bloomberg reported that the DOJ has opened an antitrust investigation into that Groq deal, specifically asking whether NVIDIA structured it to avoid merger review.

Those two facts belong in the same paragraph. Almost nobody is putting them there.

Here's the thing most of the coverage got wrong, and I include the version I heard from three different founders last week: NVIDIA did not acquire Poolside. It paid $6 billion for a non-exclusive license to Poolside's "Model Factory," made offers to 109 employees, and put $1 billion in at a $12 billion pre-money valuation. The founders stayed. The company still exists. NVIDIA said plainly it was neither an acquisition nor an acquihire — and structurally, it wasn't.

That distinction isn't pedantry. It's the whole story.

Three deals that weren't acquisitions — and one that is

Target	Date	Structure	Reported value	Merger review?
Enfabrica	2025	Non-exclusive license + team (AI networking)	~$900M	No
Groq	Dec 2025	Non-exclusive license + founder & staff hired; company continues under a new CEO	~$20B	No — and now under DOJ investigation
Poolside	Aug 2026	$6B license to "Model Factory" + 109 hires + $1B equity at $12B pre	~$7B	No
Hugging Face	Aug 2026	Outright acquisition	$12.93B	Yes

Roughly $27 billion across three deals in nine months, all built on the same template: pay a licence fee, absorb the people, leave the shell of the company standing. Senators Warren and Blumenthal sent Jensen Huang a letter in March asking whether the Groq structure was designed to evade antitrust scrutiny. NVIDIA's answer, then and now, is that this is the American system working as designed. The DOJ apparently wants a closer look.

And then, with that template working beautifully and a federal investigation already circling it, NVIDIA turned around and did the one thing the template exists to avoid. It bought a company outright — filed, reviewable, and on a deal where the antitrust argument writes itself.

Why break the pattern here?

Because the template only works on a certain kind of asset.

You can license a chip architecture. You can license a model factory — Poolside's Model Factory is, at bottom, a training pipeline: code, weights, methods, and the people who know how to run it. Hire the 109 engineers who built it and you have functionally acquired the capability, no matter what the paperwork says.

You cannot license 18 million developers.

Hugging Face isn't IP. It's a place. NVIDIA's own announcement puts the platform at more than 3 million models, 500,000 datasets, 1 million applications, and 200,000 companies using it to find and deploy AI. None of that transfers with a licence and a hiring spree. Habit doesn't license. Defaults don't license. The muscle memory of from transformers import doesn't license. If you want the distribution layer, you have to own the distribution layer — and owning it means filing the paperwork and standing in front of a regulator.

So NVIDIA did. That tells you how badly it wanted this, and it tells you the licence playbook has a ceiling. Capability, yes. Community, no.

The detail I keep coming back to: per Clément Delangue's own account to CNBC, Hugging Face approached Huang, weeks before the deal. The neutral ground went looking for the buyer. Make of that what you will about the economics of running the world's model registry on someone else's margins.

The part that actually lands on engineering software

Here's where this stops being an AI-infrastructure story and starts being our story.

ThreadMoat tracks 1,089 companies across the engineering and industrial-AI stack — $33.9B in disclosed VC funding, $131.2B in combined valuation, 815 investors, 49 countries. Sort them by what they're actually built on and a pattern shows up that predates this deal by two years.

The AI-native simulation cohort — roughly 100 CAE/CFD/FEA companies in our incumbent-exposure model — is largely a surrogate-model business. Train a neural network on solver output, replace the solve with an inference call, collapse a six-hour run into six seconds. That's the thesis. And that thesis runs on a stack that looks like this:

GPUs — NVIDIA.
The physics-ML framework — NVIDIA PhysicsNeMo, open-source, with the reference checkpoints published on Hugging Face. Ansys is integrating it into SeaScape for semiconductor design; Luminary Cloud and nTop have both built on it.
The simulation runtime and synthetic-data engine — Omniverse and Isaac Sim. NVIDIA again.
The distribution shelf — Hugging Face. As of two weeks ago: NVIDIA.

Four layers. One owner. I wrote in July that simulation was the most exposed layer in the whole stack because AI surrogates break the per-solve licensing model that Ansys and its peers are priced on. That's still true. What's changed is who the exposure runs to. The threat to Ansys was never really a Swiss startup with 40 people. It was the substrate underneath every one of those startups consolidating into a single vendor — and that vendor now controlling not just the compute and the framework but the place your customers go to find your model.

NVIDIA has committed to keeping Hugging Face open — brand intact, neutral to frameworks, clouds and inference providers, no requirement to use NVIDIA compute. I take that commitment as sincere and I'd still tell any founder to read it the way Forrester's Charlie Dai put it: be alert to any shift in its open stance. Neutrality isn't a policy. It's a structure. And the structure just changed.

What founders should actually do about it

Three things, none of them panic.

1. Know your substrate risk, and price it. If your product is a surrogate model trained with PhysicsNeMo, hosted on Hugging Face, running on CUDA, you have a single-vendor dependency across four layers and no leverage at any of them. That's not fatal — plenty of great businesses run on AWS — but it needs to be a line in your board deck, not a surprise in due diligence. The question an acquirer will ask is: what part of this is yours?

2. The exit menu changed shape, and it might be good news. The old menu for an AI-native engineering startup was: get bought by Dassault, Siemens, PTC, Autodesk or Schneider, or grind. NVIDIA has now demonstrated a fourth option — the licence-out: your technology gets a nine-figure-plus cheque, most of your team gets hired, your investors get marked up, and your company keeps existing. Poolside's founders stayed and raised at $12B pre. That is a genuinely new liquidity path for a category where "you're too small to buy but too dangerous to ignore" has been the default position for years.

3. Understand that it also breaks how exits get counted. This is a data problem I'm now living with. When Autodesk paid $2.6B for MaintainX, that was an acquisition — clean row in the database. When NVIDIA pays $6B for a licence and takes 109 of your people, what is that? Not an exit. Not a raise. Not an acquisition. Every market-intelligence dataset in this space, ours included, is going to under-count consolidation if the licence-out becomes the norm. If you're reading exit counts as a proxy for how consolidated a category is, that proxy is degrading in real time.

The tell

NVIDIA has spent nine months proving it doesn't need to buy companies to absorb them. Then it paid $12.93 billion, in public, under review, with the DOJ already asking questions about the last one — for a website where people put files.

That's not a chip company diversifying. That's a company that has concluded the scarce asset in AI is no longer silicon or weights or even talent. It's the shelf — the default place developers go, and the telemetry of what they reach for. NVIDIA can see download curves on every rival backend now. Justin Boitano called Hugging Face a "deconcentration platform." He may even be right about the effect on the developing world. He is definitionally wrong about the effect on concentration in the stack our companies build on.

The verdict

For the incumbents in engineering software, this doesn't change the near-term picture: Dassault, Siemens, PTC and Autodesk still own the design core, and switching costs still measure in careers. For the challengers, it changes the picture more than the headlines suggest — the layer they thought was neutral infrastructure now belongs to a competitor's supplier.

And for everyone watching the M&A tape: stop counting acquisitions. Start counting the deals structured so you won't count them. Over the last nine months, those were four times bigger.

The Arm bid died in 2022 on exactly this argument — a platform everyone depends on shouldn't be owned by one of the parties depending on it. NVIDIA lost that one. This time it isn't buying the architecture. It's buying the storefront. My honest read is that it gets through, because the remedy regulators know how to demand — keep it open, keep it neutral — is the thing NVIDIA has already volunteered.

Which is a fine outcome, right up until the quarter where being neutral costs more than it earns.

ThreadMoat tracks the whole board.

1,089 engineering and industrial-AI companies, $33.9B in disclosed funding across 49 countries, and the incumbent-exposure map for all nine layers of the stack — updated as the deals land, licensed or otherwise.

Subscribe to ThreadMoat → for the full dataset and the weekly signal, or browse more Insights →.

Deal values, dates and structures reflect public reporting as of 10 September 2026; the Groq figure has been reported at both $17B and ~$20B. Dataset figures are ThreadMoat's Q3 2026 snapshot (data last synced 9 September 2026). Analysis and judgments are mine, not financial advice.