Nvidia agreed this week to buy Hugging Face, the site where the open-source AI world keeps its models, for a reported $12.9 billion. On its face the price is strange. Hugging Face generates something like $150 million in revenue, and Nvidia already builds and gives away open models of its own. It is not short on weights. So what does the money actually buy?
The cleanest answer came from analysts doing the arithmetic: $12.9 billion is about 12 days of Nvidia's sales, and it purchases an insurance policy against Nvidia's own best customers. OpenAI is designing chips with Broadcom. Anthropic trains on Amazon's Trainium. Google is a decade into its own TPUs. Every one of Nvidia's largest buyers is building an escape route. Open models are the counterweight, because a downloaded model gets fine-tuned and served on Nvidia's CUDA software by default. Broadcom and Amazon can build a chip. Neither can make ten thousand code repositories target it. As one product analyst put it, Nvidia spent 12 days of revenue to make sure the open-source rival to its own customers never dies.
Hugging Face is where that rival lives. It is the default place a new model lands, the leaderboard, the datasets, and the transformers library a good slice of the industry runs on. Owning it is the weights equivalent of Microsoft buying GitHub in 2018, and for the same reason: control the surface developers already stand on, and you shape everything they do next.
The timing tracks a genuine shift. Chinese open models have taken more than 30 percent of US token usage on OpenRouter every week since February, peaking near 46 percent. Making them easier to run keeps lowering the switching cost. A recent Ollama release lets Claude Desktop quietly serve Qwen, DeepSeek or Kimi through Anthropic's own model picker, and Apple's newest Macs are configured to run sizeable models on a laptop. The default is drifting toward free, one convenience at a time.
The other half of the story is that "free" is starting to find a price. Open weights are not the same as free labour, and Chinese labs are testing routes to revenue. Moonshot AI is reportedly negotiating with Microsoft, Amazon and Google to put its Kimi K3 model on their clouds, seeking as much as 30 percent of the resulting revenue. Enterprises are paying in subtler ways: Thomson Reuters built an in-house model on an adapted Qwen, and the legal AI firm Harvey post-trained Kimi K3 into a system it says runs at under a quarter of the cost of leading foundation models. The money moves to the distributor and the integrator rather than to a per-token meter.
For the two labs preparing to go public, Anthropic and OpenAI, this is the uncomfortable backdrop. Their pitch is a premium product in a market increasingly treating free as the default. Closed models will still win plenty of work, and the point is not to swap Claude for Qwen tomorrow. The veteran investor Bill Gurley's line is that open models are "water running downhill." The real lesson, The New Stack argued, is to stop assuming today's default will still be tomorrow's. Nvidia just spent $12.9 billion betting on exactly that.