Nvidia bid $12.93bn for a company last priced in 2023

Nvidia agreed on 3 September to acquire Hugging Face for $12.93bn, a deal it expects to close in the first half of 2027. In the same handful of days, three companies that rent out compute raised more than $5bn between them, and four frontier labs shipped new models. Only one side of that came with a price attached.

What got priced

Crusoe raised more than $3B co-led by Atreides Management and Valor Equity Partners, valuing the Denver data centre operator at roughly $30B. **Fluidstack** closed a previously unreported $1.5B round led by Jane Street at over $18B. Gimlet Labs raised a $300M Series B led by Andreessen Horowitz at $3B, for an inference cloud that spreads workloads across different chip types. iPronics, a spin-off from the Universitat Politecnica de Valencia, raised a $125M Series B with Nvidia joining, to scale optical circuit switching inside AI data centres.

None of the four trains a model. Two of them are buildings and power; one routes work between chips; one switches light between them.

The Fluidstack mark, against the record

This ledger already carried the Fluidstack number, in a weaker form. On 14 April, TechCrunch relayed a Bloomberg report that the company was in talks for a $1B round at an $18B valuation, and noted that no such round had been confirmed. Five months on it is confirmed, and it is bigger than the report: $1.5B, not $1B.

The two accounts of the previous round do not line up. Fluidstack's own announcement records an $830M Series A closed in January at $7.5B, led by Situational Awareness; Forbes describes a $750M July round at the same mark. They agree on $7.5B, which is the figure that carries the argument: the company has more than doubled it inside a year, on the back of an Anthropic contract to build the first sites of a $50B American compute programme.

What Nvidia is buying

The record makes the Hugging Face price legible. Its last priced mark was $4.5bn in August 2023, and nothing has repriced it since. It has raised $395.2M in total. It crossed $100M in annual run rate in June, on 13 million registered users of whom about 97% pay nothing. Roughly half the Fortune 500 use it.

So $12.93bn is about 2.9 times a three-year-old mark, and more than thirty times the company's lifetime fundraising. Clem Delangue has argued that enterprises are done renting frontier models and are moving to running open weights themselves. Nvidia has now paid for the place those weights live.

The models, meanwhile

The same week produced xAI's Grok 4.5, aimed at coding and autonomous agents rather than consumer chat; Anthropic's Claude Fable 5.1 and Claude Mythos 5.1; and, inside a single week, Meta's Muse Spark 1.3, Google's Gemini 3.8 Flash and OpenAI's GPT-6 Astra, a pace engineers quoted by CNBC called model fatigue.

Four labs, one week, no disclosed prices, and one shared dependency: the layer underneath, which is exactly where this week's money went. The test of whether that is a repricing or an overshoot is not the next model launch. It is whether the Crusoe and Fluidstack marks survive the first lab that trims a capex commitment.