Why $12.9 Billion for a Model Repo Might Not Be Crazy

My first reaction when I saw the news this morning that NVIDIA had reportedly agreed to buy Hugging Face for $12.9 billion was pretty simple: WoW, that is crazy money for a model repo.

I use Hugging Face quite a lot. Search for a model, read the model card, find a quantized version, download some weights, look through a dataset, test something and move on. Obviously I know there is a lot more behind Hugging Face than a directory full of models, but mentally I have still always thought of it as something like GitHub for AI models.

And $12.9 billion for that sounds slightly insane.

Especially when you look at it the traditional way. Hugging Face is reportedly doing around $150 million in annual revenue. NVIDIA would therefore be paying roughly 86 times revenue. Not earnings. Revenue. Hugging Face was valued at $4.5 billion only three years ago.

If someone showed me those numbers in isolation I would probably say the valuation makes very little sense. Then I realized I had done something surprisingly similar myself earlier today.

I am currently negotiating to buy secondary shares in a private company operating somewhere around the intersection of finance and AI. I am deliberately keeping the company and the number of shares private for now, but I have spent quite a bit of time going through the numbers. Look at the company using the valuation methods I learned over decades of running companies, reading financial statements and thinking about investments, and I think it looks expensive. Probably very expensive. Today I still said yes. I want to buy X number of shares. Why? Because I am not really buying what the company is today. I am buying what I think it can become., is it 1999 again, yes it kinda is, but not also in the same way.

My base case is that I think the investment can return at least five times my money over the next five years. Obviously I can be completely wrong. That is investing. But if I believe my assumptions about where the market is going, what technology will do to this industry and where this particular company can sit within that change, then today's traditional valuation becomes much less interesting.

That does not mean valuation suddenly does not matter. That is a very convenient argument people have used to lose enormous amounts of money over the years. It means sometimes you are valuing the wrong thing. And I suspect that is exactly what NVIDIA is doing with Hugging Face. NVIDIA is not paying $12.9 billion because Hugging Face currently generates $150 million a year. There is almost no way to make that arithmetic look sensible. But NVIDIA itself just reported $89 billion in Data Center revenue. In one quarter. Once you put those two numbers beside each other, the acquisition starts to look completely different.

The question Jensen Huang and NVIDIA's board are asking is probably not whether they can grow Hugging Face into a SaaS business large enough to justify $12.9 billion. The question is whether owning Hugging Face can help protect and expand the hundreds of billions of dollars NVIDIA expects to make from AI infrastructure over the coming decade. If the answer is yes, $12.9 billion might actually be cheap. Because Hugging Face is not really a model repository anymore. It has become one of the distribution layers of open AI. Millions of developers use it. There are millions of models, datasets and applications on the platform. When a new open model appears, Hugging Face is one of the first places developers expect to find it. Someone fine-tunes it, someone quantizes it, someone benchmarks it, someone creates a strange little derivative of it, and much of that activity ends up back in the same ecosystem. That is not just storage. That is a position in the flow of an industry. And I think positions like that are becoming increasingly valuable. The interesting part is that Hugging Face does not necessarily have to make enormous amounts of money itself for NVIDIA to make enormous amounts of money from owning it. Most of what I download from Hugging Face costs me nothing. But the model eventually has to run somewhere. That is the business NVIDIA cares about.

If open models continue getting better, I actually think that makes Hugging Face more strategically important, not less. Models themselves may increasingly become commodities. There will be hundreds of good models, then thousands, optimized for different tasks, languages, hardware and use cases. When the underlying thing becomes abundant, the layer that helps you find, evaluate, distribute and run it becomes more valuable. If models are abundant, discovery matters. If models are interchangeable, evaluation matters. If models are free, compute still costs money. And NVIDIA sells the compute.

You can pretty easily imagine how the pieces fit together. Find a model on Hugging Face. There is an NVIDIA-optimized version ready. Deploy it. Fine-tune it on NVIDIA infrastructure. Run inference through optimized NVIDIA libraries. Move it onto a DGX, RTX workstation, cloud GPU or whatever NVIDIA hardware comes next.

The model can remain completely open. NVIDIA does not need to charge for it. It just needs the model to eventually consume compute.

This is also why I think looking at NVIDIA purely as a chip company increasingly misses what Jensen is building. CUDA was probably the original masterstroke. The GPU was important, but the ecosystem around the GPU made NVIDIA incredibly difficult to replace. Hugging Face potentially gives them another ecosystem layer much further up the stack. There is also a defensive element to this.

The companies buying the largest quantities of NVIDIA hardware have an enormous incentive to eventually stop buying so much NVIDIA hardware. Google has TPUs. Amazon has Trainium. Microsoft has Maia. Frontier AI companies are investigating custom silicon. When your customers are some of the richest technology companies in history, you should probably assume they will eventually try to build whatever you are selling them.

The open ecosystem is different. Millions of developers, startups, researchers and smaller companies are much harder to vertically integrate. If NVIDIA can become the natural infrastructure underneath that entire fragmented ecosystem, it creates another enormous market outside the handful of hyperscalers currently writing the biggest checks. And then there is the part I find even more interesting. We still think about Hugging Face primarily as a website humans visit. I am not sure humans will be its most important users five years from now. Agents will increasingly choose models themselves. An agent given a task may decide it needs a particular vision model, reasoning model, embedding model, speech model or some tiny specialist model we would never bother installing manually. It will need a machine-readable place to discover what exists, understand its capabilities, compare alternatives and deploy it.

Hugging Face already looks suspiciously like the beginnings of that registry. If AI agents eventually make millions or billions of those decisions automatically, owning one of the places they naturally go looking becomes incredibly interesting. That part is obviously speculation. But when I am looking at investments now, I find myself spending more time thinking about these second and third-order effects than I ever did before. It is actually very similar to what I was thinking about my own investment today. If I build a spreadsheet using today's revenue, today's margins and a sensible multiple, I can argue myself out of buying the shares quite quickly. But that spreadsheet also assumes the world behaves reasonably normally over the next five years. I do not think it will.

AI is changing the cost of building companies, the number of people required to operate them, how software gets created and increasingly how entire industries are structured. I have written recently about how this is already changing the way I run my own companies. I can now build things with a handful of people and AI agents that I would previously have needed much larger teams to attempt.

So when I look at a company positioned directly in front of one of those changes, I cannot only ask what its current numbers justify. I have to ask what happens if the thesis is right. What happens if this market becomes ten times larger? What happens if AI removes 70 percent of the operational cost? What happens if a company that looks expensive today becomes the infrastructure everyone else builds on? What happens if something that currently looks like a feature turns out to be a platform? That does not mean paying any price. The hardest part of investing in technological transitions is distinguishing between genuinely seeing around a corner and simply inventing a very convincing story for yourself. Both feel remarkably similar at the time. Maybe my investment goes nowhere. Maybe NVIDIA has just paid $12.9 billion for something the open-source community routes around within three years. There is a particularly interesting risk in NVIDIA's case because Hugging Face's neutrality is part of what makes it valuable. If NVIDIA squeezes it too hard or turns it into an NVIDIA sales channel, it could destroy exactly the thing it paid for.

The clever version of this acquisition would probably be the opposite. Keep Hugging Face open. Make it more open. Spend ridiculous amounts of money making it better. Let AMD models live there. Let Intel models live there. Let everyone use it. Then quietly make the path from an open model to NVIDIA compute incredibly good. That seems much more like Jensen. So I have changed my mind since reading the headline this morning. $12.9 billion for a model repository sounds crazy. $12.9 billion for one of the places where the future AI ecosystem discovers what to run, and potentially where billions of AI agents eventually discover what to run, sounds very different. And strangely, after agreeing to buy my own ridiculously valued shares earlier today, I am probably not in much of a position to criticize him anyway. Sometimes the numbers tell you something is expensive because it is expensive. Sometimes they tell you it is expensive because you are still looking at what it was, rather than what it might become.

Figuring out which of those two you are looking at is where it gets interesting.