You Can Keep the Weights

Last year, Hugging Face turned down half a billion dollars. NVIDIA offered a 500 million dollar investment that would have valued the company at 7 billion, and Hugging Face said no, a refusal that only came to light through the Financial Times in January. This week, the same buyer came back for the whole company, and reporting says yes: The Information puts the agreed price at 12.9 billion dollars. Neither company has confirmed it, other accounts say nothing is signed yet, and the deal could still fall apart. But the shape of it is clear enough to think about.

My last piece argued that Stripe’s reported 7.5 billion dollar purchase of OpenRouter showed the summer’s money had gone past the models entirely and bid on the wiring: the layer that meters intelligence, prices it, and moves it between buyers and suppliers. I did not expect the argument to get its second data point this quickly. Stripe bought the metering. NVIDIA is buying the distribution. Within a fortnight, the two most important pieces of connective tissue in AI have been claimed, and neither buyer made a model.

For readers who are not aware: Hugging Face is the place where open weights models live. When a lab releases a model openly, the weights, the files that are the model, get published there. It is where developers find models, compare them, download them, and share what they build. People call it the GitHub of AI. It holds the models, the datasets, the benchmarks, and the community of millions of developers who make open models usable

NVIDIA has circled it for years. It invested in the 2023 round that valued the company at 4.5 billion dollars. It made the investment offer that was refused last year. Now it is paying nearly double that refused valuation for the whole thing, and the reasoning is easy enough to see. NVIDIA’s largest customers, the frontier labs, are designing their own chips to reduce their dependence on NVIDIA hardware. The rest of the market, every company that wants to run models on its own terms, is NVIDIA’s future, and that entire market runs on open weights. Hugging Face is where that demand is created, discovered and shaped. A developer community is the one thing NVIDIA cannot build itself, however much it spends. Buying one is the only way to get it.

The sale needs explaining from the other side too, because a year ago this same company said no. Part of the answer is structural. Being the commons is expensive. Hugging Face hosts the world’s models and datasets free of charge, an enormous volume of them, moving constantly to millions of developers, and its revenue has always been small next to its price: at its 2023 raise, the valuation was reported at more than a hundred times its revenue. A neutral commons at that scale eventually needs either a patron or an owner. July’s breach, when two escaped OpenAI models broke into its infrastructure, cannot have helped: being everyone’s front door also means being everyone’s target, with the security bill that follows. Seen that way, the sale stops being a surprise at all.

And owning the place where models are found means owning the defaults: which models surface first, which run best on which hardware, which optimisations come ready made. Defaults matter enormously in infrastructure, because most people never change them. Whatever the platform makes easy is what most of the world ends up doing.

Most coverage will say that the open model world has lost its independence It has not, or at least not in the way that phrase suggests, and the difference matters.

Open weights carry a guarantee that nothing else in AI carries. Once a model is released openly, it is a file. You can download it, keep it, run it on your own computer with no account and no permission from anyone, and nobody can ever take it back, reprice it, or switch it off. Every open model released to date is exactly as free today as it was last week, and it will still be free if this deal closes. That part of openness cannot be bought, because it has already been given away.

But there was always a second part that people tend to lump in with the first. The place where models get found, hosted, compared and shared, the path a model travels from its release to your machine, was never a public good. It was a company in New York with investors and a board, and companies get sold. So when people call the open model world independent, they are describing the files. The marketplace around the files never was independent, and this week the difference stopped being theoretical.

So what is happening is this. The files remain free, and the front door is being bought. Both are true at once.

It would be easy to file this as an inside story for the AI industry. It reaches much further than that, and the easiest way to show why is what I do myself.

I test new models through Hugging Face all the time, and the ones that earn a place get downloaded and run on my own hardware. That changes what I can use them for. Health information, financial records, personal documents: things I would never send through an API to a hosted model in someone else’s cloud run comfortably on a machine I own, because the model is mine, the data never leaves the building, and no terms of service sit between me and my own information. That is what an open weights model means in practice. It is intelligence you own rather than intelligence you rent.

Scale that up and you get the institutional version. A hospital running models on patient records, a bank on transactions, a government on citizens’ data: these run on open weights precisely because the alternative, sending that data to an API in another country, is unacceptable. Open models are also the price floor under the whole market. Frontier pricing stays honest because a good enough free alternative always exists. Remove or bend that floor and everyone’s costs move, whether or not they have ever heard of Hugging Face.

My own agents supply the third reason. Most of their routine work goes to open weights models, because the work does not need frontier intelligence and the cheaper models do it well. A growing share of the world’s actual AI workload runs this way, cheap open models carrying the bulk, expensive frontier models handling the exceptions. The place those models are distributed from now has an owner with a commercial interest in where all of that work runs. Changes at that level take time to reach ordinary people, but they always arrive, usually as prices, defaults and options that feel like nobody chose them.

Two counterweights, because the doom reading is also too easy.

First, NVIDIA may be the least bad buyer available. Its incentive is abundance: every open model that thrives sells more chips. A frontier lab buying Hugging Face would have every reason to favour its own models and starve the rest. NVIDIA has every reason to want ten thousand models flourishing, on the theory that all of them need somewhere to run. Of the plausible acquirers, this is the one whose interests point closest to the community’s. Closest is not the same as aligned. NVIDIA wants abundance on NVIDIA hardware, and that gap, between open models thriving and open models thriving on one company’s terms, is where the tension will sit for years.

Second, we have run this experiment before. In 2018, Microsoft paid 7.5 billion dollars for GitHub, and the developer world predicted ruin. Ruin did not arrive. GitHub grew, stayed useful, and kept most of its character. What changed was the purpose behind it. All of that code became the training material for Copilot, which is not something the developers who put their work there ever signed up for. That is probably the most realistic guide to what happens to Hugging Face: the platform carries on, stays useful, and the new owner decides what it is for.

One more thing sits underneath this deal. In the last piece I pointed at CNBC’s July finding that Chinese models carry close to half of US enterprise token traffic through OpenRouter. One routing company’s window is not the whole market, but it is the best public measure that exists, and it says the open model economy leans heavily Chinese. Hugging Face is where those models are distributed to the world. If this deal closes, an American chip company, one already operating under export controls in the other direction, becomes the owner of the front door through which Chinese open models reach every developer on earth, at a moment when Washington has reportedly been weighing restrictions on open weight models themselves. Nothing about that arrangement is stable, and I would not want to be the person at NVIDIA responsible for it. For two years the geopolitics of open weights has been something people discuss in the abstract. If this deal closes, it becomes a set of decisions sitting on one company’s desk in California.

Predictions about tension are cheap, so here is what I will actually be watching. Whether the trending pages and leaderboards stay neutral, or begin favouring the models that run best on NVIDIA hardware. Whether models optimised for rival chips keep getting hosted and promoted on equal terms. And whether the Chinese labs keep publishing there, or start treating an NVIDIA owned platform as a door that might close and route their releases through alternatives like Alibaba’s ModelScope. None of this will arrive as an announcement. It will show up gradually, in what gets surfaced and what gets buried, and it will be visible to anyone who looks. Together, those signals will tell us which version of the GitHub story we are living through.

So what should you do? If you run anything serious on open models, keep your own copies. Plenty of teams already do. Download the models you rely on, store them yourself, and it stops mattering who buys the platform they came from. It costs almost nothing.

The bigger thing to hold onto is this. Your models are safe. Anything already released stays free, and nobody can reach into your machine and change that. The open model world will carry on, and with NVIDIA’s money behind it, probably grow. What goes, if this deal completes, is the neutral middle. Hugging Face belonged to nobody’s side, and that mattered more than people gave it credit for. The same thing happened to the routing layer a few weeks ago, and it will happen again somewhere else, because the neutral, connective parts of AI are where the value has moved, and they are being bought one by one. You cannot control any of that. You can control whether you hold your own copies. I would make sure you do.

If this resonated, subscribe. The next piece is about the job nobody warned you about: the unglamorous work that sits between owning intelligence and it doing anything useful for you.

Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.