Twelve Point Nine Billion for the Open Model Commons: Nvidia Buys Hugging Face - September 5, 2026

0:00Welcome to the DX Today podcast. Your daily deep dive into the AI ecosystem. I'm Chris and joining me is always as Laura. Thanks Chris and I have been waiting all week to talk about this one because on September 3rd, the chip company that already owns the picks and shovels of this entire industry went out and bought the general store. That is a great way to frame it. But let's put the actual number on the table first. Because I think the number is the part that makes people sit up straight when they hear it. $12.93 billion. That is what Nvidia is paying for hugging face. The platform where basically the entire open model world goes to publish, download and discover machine learning models and data sets. It is the closest thing this industry has to a public square. And it just changed hands to the company that supplies almost everyone's compute. And my first honest reaction was confusion. Because Nvidia sells graphics processors and networking gear and software that runs on them. Why does a hardware company need to own a website where people share model files? Because it is not a website where people share model files.

1:01That framing undersells it enormously. Hugging face hosts roughly 3 million models, about 1 million applications and more than 500,000 data sets on a single platform. Okay, those are big numbers. But big numbers on their own do not justify a $12 billion check. What is the number that actually explains the price to you when you look at this? More than 18 million developers. That is the number. Nvidia is not buying storage. It is buying the default habit of 18 million people who reach for that platform first when they want a model. Let's stress test that against the financials, though. Because I looked at the reported revenue and it does not obviously support this valuation by any traditional measure that a public market investor would recognize. You are right. And it is worth saying out loud. Hugging face was reported last month at roughly 150 million dollars of annualized revenue. That works out to something like 86 times revenue. 86 times revenue. For context, a healthy fast growing software company might trade at 15 or 20 times.

2:02And people already complain that those numbers are stretched. So what are they actually paying for? They are paying a position multiple, not a revenue multiple. The question Nvidia is answering is not, what does this business earn today? It is what does it cost to never be locked out of distribution for a company whose market value is measured in trillions. 12.9 billion is a rounding error against the risk of arrival owning that choke point. Walk me through the structure because the way a deal is built usually tells you more about what the buyer is worried about than any press release quote ever will. Reported as roughly 11.9 billion dollars in cash plus about one billion dollars in retention with closing expected in the first half of 2027 subject to regulatory approval. That retention equity line is the tell isn't it? A billion dollars set aside specifically to keep people from walking out the door in the 18 months after the ink dries. Exactly right. When you buy infrastructure, you buy servers. When you buy a community platform, you are buying the maintainers, the moderators, and the culture.

3:06And none of that is on the balance sheet. If half the core engineering team resigns in protest six months after closing, Nvidia has paid 12 billion dollars for a very expensive file server and a logo. There is a detail in the reporting I keep coming back to, which is that hugging face reportedly turned down a $500 million offer from Nvidia only last year. And then according to reporting from CNBC, chief executive Clement D'Alang was the one who went to Jensen Huang weeks ahead of this deal. The seller initiated the conversation the second time around. So in roughly a year, the valuation move from a rejected 500 million to an accepted 12.93 billion in the direction of the phone call reversed. What changed in between? D'Alang's own explanation is refreshingly plain. He said the platform needs more compute, more support, more collaboration, and more visibility. And that is why they went to talk to Jensen in the first place, which points at something people underestimate about that platform. Serving 3 million models to 18 million developers is not a cheap hobby.

4:08It is a serious bandwidth and storage bill every single month. It is a genuinely brutal cost structure. You are running what amounts to a global content delivery network for enormous binary files. Mostly for free, funded by a comparatively small enterprise business on the side. Every time a popular open model gets released, millions of people pull tens of gigabytes each. And somebody has to pay that egress bill without ever sending an invoice to the person downloading. So the charitable read is that the comments had an unsustainable funding model and found a patron. The uncharitable read is that the comments just got acquired by the single most interested party in it. Both readings can be true at once. And I think honest analysis requires holding them together rather than picking the one that fits your priors about big technology companies buying smaller ones. Let's take the commitments seriously first. Because Nvidia did make specific public promises rather than the usual vagria sureances that acquires hand out and then quietly abandoned a few years later. Jensen Huang said that hugging face will remain an open platform for the entire AI ecosystem.

5:12And then added the sentence that actually matters, which is that Nvidia compute will not be required. That last clause is the one I would print out and tape to the wall because it is specific enough to be tested and specific enough to be visibly broken later. And that is why it is a smart thing for them to have said a vague promise cannot be audited. A promise that names a concrete behavior can be checked by anyone paying attention. But here's my pushback and it is not about honesty, it is about defaults. Nobody has to break a promise to change an ecosystem. You just have to make one path slightly smoother than the others. That is the single sharpest concern in this whole story and you have put your finger right on it. Neutrality is rarely destroyed by a ban. It gets eroded by documentation, by tutorials, by which button is bigger. Right. If the one click deploy path happens to target one company's inference stack and the alternative takes four configuration steps in a forum search, you have shaped behavior without violating a single word you said.

6:13Forester analyst Charlie Dye made essentially this argument from the structural side. Nvidia already consolidates across graphics processors, kudia networking inference software and a frameworks and this adds the developer, distribution and community layers. Let me play devil's advocate on that because a list of layers is not automatically an antitrust problem. Plenty of companies are vertically integrated and the customers are perfectly happy about the convenience. That is a completely fair challenge. Vertical integration is often good for users. The concern is specifically that this particular layer, distribution and discovery is the layer that competitors need in order to reach anyone at all. So the theory of harm is not that Nvidia charges more. It is that a rival chip or a rival runtime becomes marginally harder to find and marginal friction compounds across 18 million people. Precisely and the honest caveat is that this is a theory about the future, not an observed behavior today. Nothing has happened yet. The deal is not even closed and will not

7:14until well into next year. I think it is important to criticize the structure without pretending we have already seen the misconduct because that distinction is what separates analysis from grievance. There is a broader criticism floating around this deal too from the financial side rather than the competition side and I found it uncomfortable in a productive way. You mean Nigel Green calling the American AI trade dangerously circular where the same capital keeps moving between a tight circle of suppliers and gets counted as fresh revenue at every stop along the loop. That one lands harder when you line it up next to the other commitments because this acquisition follows a reported six billion dollar investment in poolside and roughly 50 billion committed to frontier labs. The pattern is that the company selling the compute is also funding the customers who buy the compute and now owns the storefront where the resulting models get distributed. It is a very tight loop. Although I want to defend the other side here because there is obviously real end demand underneath all of this. Actual businesses are actually paying for inference.

8:17This is not purely capital chasing its own tail. Agreed and I would not want anyone to hear this as a claim that the demand is fake. My concern is narrower. Circularity makes it much harder to tell how big the genuine demand actually is. What happens on the regulatory side between now and the first half of 2027? Is this the kind of deal that agencies actually try to stop or is it more likely to sail through with conditions? Historically, this shape of deal is difficult to block because the two companies do not compete with each other in any direct sense. Nvidia does not run a model hub of its own and hugging face has never sold a single piece of silicon to anybody. So the traditional overlap analysis finds nothing and the regulator would have to argue a foreclosure theory instead. Which is a harder case to make and takes considerably longer to build. Which is why I would expect behavioral remedies rather than a block if anything happens at all. Commitments about equal treatment of competing chips and run times probably with some monitoring period attached to them. Let me bring this down to the person listening who just wants to know whether their workflow breaks on

9:20Monday. What actually changes for a working developer in the near term? Honestly, nothing. The deal does not close for the better part of a year and even after it closes there is no reason to expect an abrupt change to a platform whose entire value is trust. So what are the signals worth watching instead? The things that would tell you the trajectory is bending before anyone announces that the trajectory has bent. Watch the default deployment targets and the documentation. Watch whether models optimize for competing hardware, still surface and search rankings. Watch whether the leadership team is still there two years after closing. I want to give Nvidia real credit on one point because they are not a stranger to this community. They have published over 500 models and 250 open data sets on that platform themselves. That is a genuinely strong track record and I do not want to wave it away. Nvidia has been one of the more consistent corporate contributors of open weights and open data sets in the industry. Though contributing to a commons and governing a commons are different jobs with different incentives and being excellent at the first does not automatically make you trust

10:25the second. That distinction is the whole episode compressed into one sentence. A generous neighbor is not the same thing as a Fairland Lord and the entire open model ecosystem just woke up to discover it has a new landlord who also happens to own the utility company. If you had to name the thing that would make you say three years from now that this deal was good for open source, what would that thing be? Open models getting materially cheaper and faster to host and serve, competing hardware still treated as first class on the platform and the maintainers who built the culture still there and still genuinely empowered to say no to their new corporate parent when it matters. And the failure signature, the thing that would make you say this was the moment the open ecosystem quietly lost its independence without anyone noticing it happening in real time. Quiet attrition of the original team, documentation that slowly assumes one vendor stack, an arrival hardware ecosystem building its own competing hub because it no longer feels welcome on this one. That last one is interesting because a competing hub would arguably prove the ecosystem

11:25is healthy even while it proved that the neutrality of the original platform had been lost. A wonderfully uncomfortable paradox to end on. Fragmentation would be evidence of resilience and evidence of failure at exactly the same time and I suspect we will find out which reading wins somewhere in the middle of next year when this deal finally closes. That's all for today's episode of the DX Today podcast. Thanks for listening and we'll see you next time.