Hugging Face raises $235 million at a $4.5 billion valuation

Background

Hugging Face is best known as a place where machine-learning developers publish and download open models and datasets; for many teams it is the first stop when they want to try someone else's model rather than train one from scratch. That role puts the company in the middle of the AI supply chain: the models hosted there run on chips made by hardware vendors, and they are often deployed through cloud services.

The round

On 24 August 2023 the company raised a $235 million Series D ^[1]. The round valued Hugging Face at $4.5 billion post-money, more than double the $2 billion valuation it had in 2022 ^[2]. Axios reported that Salesforce Ventures led the round ^[2].

Who put money in

The investor list went well beyond the lead: it included Google, Amazon, Nvidia, AMD, Intel, Qualcomm, IBM and Sound Ventures ^[1]. Nearly every major AI chipmaker and cloud provider took part ^[1]. News of the deal was first reported by The Information and then appeared to be confirmed by Salesforce chief executive Marc Benioff on X ^[3].

Reading the numbers

A few terms help make sense of the announcement. A Series D is conventionally a company's fourth large round of venture funding, money raised after several earlier rounds rather than at the start. A post-money valuation is the value of the company once the new money is counted in, so it is the figure that includes the cash just raised. Annualized revenue is a run rate: a shorter period's sales scaled up to a full year, useful for comparing fast-growing companies but not a forecast.

Coverage at the time noted that the $4.5 billion figure was reportedly more than 100 times Hugging Face's annualized revenue ^[3], a multiple that says more about expectations for the market than about current sales. The company did have paid products: AutoTrain, a service for training models with less manual setup; the Inference API, for running a model over a network instead of on your own machine; and Infinity, aimed at faster inference in production ^[3].

What it signaled

Two things stand out. First, the investor list is unusual for a single startup round: chipmakers that normally compete with one another, and cloud providers that sell overlapping compute services, all bought in ^[1]. Second, the size of the valuation relative to revenue shows where optimism was pointed in this period: not only at the labs that train large models, but also at the layer where models are shared, adapted and served ^[3].

For Hugging Face itself, the round brought capital and a set of partners whose hardware and cloud services its users already rely on ^[1]. The company's position is therefore easy to describe and hard to hold: it sits between the hardware vendors, the clouds and the model builders, and depends on all of them continuing to treat the platform as common ground.