China now drives 41% of all Hugging Face downloads, eclipsing the US

The numbers are staggering and they tell a story of a power shift most people missed. Hugging Face just dropped its Spring 2026 ecosystem report, and the topline figures show a platform doubling in size—13 million users, over 2 million public models, and more than 500,000 datasets. But raw growth hides a brutal concentration of attention. The top 200 models, a microscopic 0.01% slice, vacuum up nearly half of all downloads. Half the models on the platform can’t even crack 200 downloads. It’s a superstar economy, plain and simple.

Dig into the geography, and the narrative takes a sharp turn. China has not just caught up; it has blown past the United States in monthly and overall downloads. Chinese models now account for 41% of all downloads, a plurality that reshapes assumptions about where AI’s center of gravity lies. Meanwhile, the old guard of industry labs is losing its grip. Before 2022, companies were responsible for roughly 70% of development. By 2025, that number collapsed to 37%. Independent developers and small, unaffiliated collectives have filled the vacuum, rising from 17% to 39% of downloads. These are the tinkerers—the people quantizing, adapting, and redistributing base models. They’re not just users anymore; they’re the new distribution channel.

Big Tech is absolutely still in the game, but they’re playing differently. Over 30% of the Fortune 500 now have verified Hugging Face accounts, and the report highlights NVIDIA as the single most prolific contributor among the giants. Legacy companies are upgrading their subscriptions, and startups like Thinking Machines are building entire product lines on open weights. The report suggests a pragmatic truth: the downstream value created by all this tinkering and adapting likely dwarfs the initial cost of creating the models. Companies chaining themselves exclusively to closed APIs are paying a premium and sacrificing flexibility, and they seem to know it.

What does this actually look like on the ground? It’s not one ecosystem. It’s a messy collection of overlapping sub-ecosystems built around specific languages, domains, or problems. A model with modest global downloads might be a vital piece of infrastructure for a niche community, showing sustained engagement that a top-line chart would never reveal. The big, lingering question is about control. If individuals and Chinese labs are driving both creation and distribution, the power to set the agenda for what gets built—and who it works for—has quietly changed hands. The map of AI influence has been redrawn, and most of the old landmarks are obsolete.