Chinese open-weight models accounted for 41% of Hugging Face downloads this spring, edging past U.S. releases for the first time, while the top six slots on OpenRouter by usage went entirely to Chinese open releases from DeepSeek, MiniMax, Tencent, Xiaomi, and Z.ai. Anthropic's Claude Opus 4.7 sits in seventh. The number that best describes the shift, though, arrived on July 13, when Microsoft CEO Satya Nadella published a blog post telling enterprises that proprietary APIs make them "pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful."

That's not the CEO of a company that resells OpenAI capacity making a rhetorical flourish. It's the closest thing the enterprise software industry has to an official endorsement of the open-weight default.

The economics were doing the work well before Redmond gave them a voice. On OpenRouter, Anthropic's Opus 4.8 lists at $1.37 per million tokens; DeepSeek V4 Flash lists at six cents. That's a 23x spread. Justin Summerville at OpenRouter told CNBC that open-source Chinese models run "60% to 90% cheaper" than the leading Anthropic and OpenAI systems, and the routing data reflects it. Since February 8, U.S. companies' token share on Chinese models via OpenRouter has stayed above 30% every week, peaking at 46%. The 12-month trailing average is 11%. The first-half-of-2025 baseline was 4.5%.

Vercel's numbers tell the same story from the infrastructure side. In June the company routed 29% of gateway traffic to open models. Z.ai's GLM 5.2, released that same month, exploded on the platform. "In its first full week after launch, daily token volume grew about 27x and the number of customers using it grew about 80x," said Harpreet Arora, Vercel's head of agentic infrastructure. His diagnosis was blunt: "Price is doing the work here. When a task doesn't need the best model, teams are beginning to route it to the cheapest one that's good enough."

Regulated-industry buyers, once assumed to be a permanent moat for the closed labs, aren't behaving that way either. LaunchLemonade, an agent platform for regulated industries, now lists Chinese open models alongside ChatGPT in its top five.

Hugging Face, where a new repository is created every seven seconds and nearly three million public models now live, has become the distribution layer for this rerouting; half the Fortune 500 are deployment customers. CEO Clem Delangue framed the strategic argument the closed labs least want to answer: "You don't really make it safe by keeping it behind closed doors for just a few players. You make it more dangerous because you create asymmetry of power."

The 2017 open-source-database inflection, when Postgres and MySQL quietly displaced Oracle as the boring default for new production systems, took about five years to become obvious to CFOs. This one is compressing into two quarters, and the CEO of the largest enterprise software company on earth just published the memo.

Chinese open-weight models pass 41% of Hugging Face downloads as Nadella tells enterprises they're paying twice

Sources