250 Employees, 5 Million Users a Day

By October 2024, Hugging Face had 250 employees serving over five million daily active users. The company was profitable. It hosted three million models. New models were uploaded every ten seconds. And Clément Delangue, the CEO, kept saying the same thing in interviews: “This is the extent that I can see in terms of scale. The way we operate is the same as when we were 40 or 50 people — it might work if we were 250, but not 1,000.”

That sentence is the most interesting thing about Hugging Face’s organizational design. The company’s growth ceiling is not capital, not market, not technology. It’s the operating model itself — and Delangue is fine with that, because the operating model is what makes the leverage possible.

For executives building AI-era organizations, the structural choices behind that ratio are worth studying.

What They Don’t Do

Most companies of 250 employees have department heads, formal job ladders, marketing teams, sales teams, defined OKRs cascading down a hierarchy, and synchronous all-hands meetings.

Hugging Face has none of these in the form most executives expect. The team organizes around three focal areas — revenue, visibility, usage — each owned by a different group with different incentives. There is no single marketing department; the engineer who builds a feature is also expected to tweet about it and talk to journalists about it. There is no community manager who responds to user feedback; that responsibility is shared across whoever is closest to the conversation.

Lysandre Debut, who leads open-source at the company, described the staffing model: “If we have ten experts in machine learning, the field moves so fast that our experts cannot implement everything that comes out. Enabling our community means that we can keep up to date, because an interested community member can work on new technology, provide it as a tool, and make it make sense to the rest of our community.”

This is the structural insight: at Hugging Face, the headcount inside the company is not the actual labor force. The headcount is the orchestration layer for a much larger volunteer labor force outside the company.

The Generalist Bias

Delangue described the cultural orientation early: “We started with the idea that people are much more multifaceted than large organizations allow them to be. We try to turn that into a strength for the company.”

In practice, this means hiring people who would be uncomfortable at a more structured organization. Roles are fluid. A research engineer might run a sales conversation. A community lead might write a research paper. The downside is real — onboarding takes longer, and the company has had to push back on contributors who saw volunteering as a hiring audition. The upside is that the team can absorb new domains (audio in 2021, vision in 2021, robotics in 2024) without restructuring.

Most enterprises looking at this from the outside see a startup quirk. It isn’t. The generalist orientation is what makes the leverage ratio sustainable. A specialist organization would need 1,000 employees to cover the same surface area Hugging Face covers with 250. That extra 750 in headcount is what most growing companies pay for, and it’s what slows them down.

The Decentralized Decision-Making That Actually Works

The temptation, when reading about flat organizations, is to assume they are chaotic. Hugging Face’s version is not chaotic — it’s opinionated.

Thomas Wolf, the Chief Science Officer, on project proposals: “People come to me ten times a day with a proposal for a new project, and I say yes to maybe 1% of it. We also do a lot of small bets with maybe one or two people, to see if they can prove that it will work. We want people to have a plan — what happens if they succeed? What happens if they fail? What are the lessons?”

Note the phrasing. Management doesn’t approve projects in the formal sense. Management gives feedback. The decision-maker is the person doing the work. But the feedback is direct enough that bad ideas die quickly, and the small-bets discipline ensures that the company can run dozens of experiments without committing real resources to most of them.

This is the structural difference between flat and decentralized. Flat organizations remove the hierarchy and hope for the best. Decentralized organizations remove the hierarchy and replace it with fast, honest, asynchronous feedback loops. The first one collapses past 50 people. The second one scales to 250 and Delangue thinks possibly 500.

The Lesson That Doesn’t Translate

A reasonable reaction to all of this is: we couldn’t run our company this way. That’s correct, and it’s the right lesson.

Hugging Face’s operating model is not a template. It’s a specific solution to a specific problem: how do you build a company whose primary leverage comes from a community of contributors, most of whom don’t work for you? Most organizations don’t have that problem. Most organizations are trying to ship a product to customers, not orchestrate a global open-source community.

But the principle generalizes. The specific structures Hugging Face built — generalist teams, fast feedback over formal approvals, clear focal areas with autonomous owners — are answers to a question every AI-era company has to answer: how do we move fast enough to keep up with how fast AI is changing the work itself?

Companies that try to answer that question with traditional hierarchy will struggle. The work changes faster than the org chart can be redrawn. Hugging Face’s bet is that the operating model itself has to be designed for a world where the field moves faster than any individual team’s expertise.

So far, the bet is paying off. Five million users a day, with 250 employees, suggests the leverage ratio is not theoretical. The next question — whether the same operating model can scale past 500 — is one Delangue has openly said he doesn’t yet know the answer to. That is, in itself, the most candid thing a CEO has said about organizational design in years.