Latest / Minds, Bodies, and Terawatts / Nvidia Is Buying the Library Where the Free AI Lives, for $12.9 Billion
Transcript
- 0:00[Host] This is Minds, Bodies, and Terawatts, the show where somebody finally put a price on free. The Information reported Wednesday that Nvidia has agreed to buy Hugging Face for twelve point nine billion dollars.
- 0:12[Guest] Hugging Face is the public library of open AI. Two and a half million models sit on it, free to download, and it is where most of the open-weight world gets handed out.
- 0:23[Host] Nvidia sells chips. Why does a chip company want a library whose entire product is given away?
- 0:29[Guest] Because every one of those files still has to run somewhere, and somewhere is almost always Nvidia hardware. The file is free. The machine underneath it is the business.
- 0:40[Host] Be concrete. If I download a frontier open model tonight, what am I actually running it on?
- 0:46[Guest] Not a gaming card. The standard configuration is eight data center GPUs, roughly a hundred thousand dollars of machine. Free file, six-figure doorstop.
- 0:54[Host] So the free download was never charity. It was the demo.
- 0:58[Guest] And you can see it in the margin. Nvidia's data center business books about seventy-five cents of gross profit on every dollar, and it posted a record quarter one day before this deal leaked.
- 1:10[Host] One thing doesn't fit. Last month Jensen Huang was the loudest voice in the industry defending open models.
- 1:17[Guest] He made the first post of his life on X to publish an open letter arguing that open models strengthen safety and let countries build their own. Hugging Face signed it with him.
- 1:28[Host] The buyer and the seller co-signed the same manifesto, then did this.
- 1:33[Guest] Both things can be sincere. He wants open models everywhere, and every copy running in a garage is demand for his chips.
- 1:40[Host] There's a harder fact. Hugging Face already turned him down once.
- 1:44[Guest] Late last year Nvidia offered five hundred million dollars for a stake. The Financial Times reported why they refused: they did not want a single dominant investor able to sway their decisions.
- 1:56[Host] So the objection was never the principle. It was the number.
- 2:00[Guest] Neutrality was the product. Hugging Face was worth something precisely because it sat between the chipmakers and the model labs and belonged to neither.
- 2:08[Host] Open source has always carried its own insurance against a bad owner: copy the work and leave. Does that still hold here?
- 2:16[Guest] The weights are portable by design. You could mirror the whole library tonight and nobody could stop you. Whether that is enough is what the next year tests.
- 2:26[Host] Start there. What changes for someone downloading a model on Monday morning?
- 2:30[Guest] On Monday, nothing. Over a year: which models get featured, which download fastest, which get tuned to run best on which chip.
- 2:38[Host] And who feels that first?
- 2:40[Guest] The school district and the small clinic running an open model instead of a per-seat subscription. Their floor price is now set by the company that sells the floor.
- 2:50[Host] Regulators have a name for a chokepoint like that.
- 2:54[Guest] Essential facilities. Railroads had to haul every farmer's freight, and Bell had to let you plug a rival's telephone into the wall.
- 3:01[Host] Nobody has ruled that a model repository belongs on that list.
- 3:05[Guest] Not yet. Nvidia is already under antitrust scrutiny in Washington and Beijing.
- 3:10[Host] Which is why your escape hatch matters. Unscarcity's law is that power must decay, and Patrick Deglon's exit is the open-source one. When the owner changes the rules, you copy the work and walk. Don't fight, fork.
- 3:23[Guest] So mirror the weights tonight. But you can fork the library, and you cannot fork the machine it has to run on. For more on who owns the on-ramp to intelligence, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.