NVIDIA bought the place where we all keep our models

On September 3, NVIDIA confirmed on its own blog that it is acquiring Hugging Face for $12.93 billion. The platform figures come from that same announcement: more than three million hosted models, more than five hundred thousand datasets, more than one million applications, over eighteen million developers and researchers, and more than two hundred thousand companies. TechCrunch reports annualized revenue of around $150 million, a figure attributed to The Information, and notes that Hugging Face's last round was $235 million in 2023, led by Salesforce Ventures with NVIDIA among the investors. Neither the announcement nor the coverage says whether the payment is cash or stock, or when the deal is expected to close.

Let me do the division before offering an opinion: $12.93 billion over $150 million in annualized revenue is roughly eighty-six times revenue. That multiple does not describe the purchase of a business. It describes something else.

What was bought was not the revenue

Jensen Huang said Hugging Face will remain an open platform for the entire ecosystem, multi-cloud and multi-accelerator, and that NVIDIA compute will not be required to build on or deploy through it. The founders stay and the brand stays. Take the promise seriously for a moment and then ask the uncomfortable question: if NVIDIA isn't going to use the repository to push its GPUs, what exactly did it buy for thirteen billion dollars?

It bought the moment of decision. Hugging Face isn't the most profitable company in the open ecosystem; it's the busiest. It's where a team goes when it has to choose which model goes into production, under which license, with which benchmark sitting next to it. That moment, the choosing rather than the deploying, is the only place left where you can still influence a market that is otherwise already settled by who owns the GPUs. NVIDIA bought the storefront, not the warehouse.

What changes for anyone running a technology company

Nothing, tomorrow. That's the honest answer and it belongs before the alarming one. The weights you already downloaded are still yours, the licenses you accepted still hold, and the promise of openness is signed by the CEO in a public statement.

What changed is quieter: a risk that used to be diffuse now has an identifiable owner. Until this week, the risk of depending on Hugging Face was the generic risk of depending on a startup — that it runs out of money, changes direction, gets acquired. As of this week the risk is specific and has a shape. A public company, with hardware incentives, controls the channel through which open model distribution flows, and answers to a board every quarter.

Openness promises are almost never broken in bad faith. They break under quarterly pressure three years later, when the person who made the promise is gone and the person deciding has a different problem. The only thing that really protects you isn't the promise: it's the architecture.

So the practical question isn't whether to trust NVIDIA. It's this: how many doors does your stack have, and do you know how many?

At Indrox we pull weights from there every week. This week I'm going to do three things, and I'm writing them down because they're cheap and almost nobody does them.

First, an inventory. Which models run in production today, where they came from, under exactly which license, and at which version. Not the list we assume it is: the real list, pulled from the systems.

Second, a copy. The weights of the models already in production, stored on infrastructure we control. That costs disks, not engineering. A remote repository is a runtime dependency only if you decided it would be.

Third, a portability test. That the runtime we use can load those weights from another origin and on another accelerator, and that someone has actually run it at least once. Portability that was never tested isn't portability, it's an assumption.

My read

I think Huang's promise is sincere and I think it's irrelevant, and both can be true at once. Sincere because today it suits him: a closed Hugging Face is worth far less than an open one, and NVIDIA knows it. Irrelevant because the conditions that make it convenient today are not the conditions that will exist when the accelerator market has three serious competitors instead of half of one.

That's the job of whoever runs technology: not to guess whether someone will behave badly, but to build so that the question doesn't matter. This week's acquisition doesn't worry me. What worries me is discovering, while reading it, that I didn't know how many doors I depended on. I'm going to find out. I suggest you do the same — not because something bad is coming, but because the inventory is useful anyway and today you have an excuse to ask for it.