China's open weight AI models are winning on one stat. That stat means less than it sounds like.
Ben Werdmuller says China's open weight strategy is beating closed US labs. His core data point turns out to describe a much smaller slice of startups than advertised.
A widely cited a16z stat says most startups pick Chinese AI models. Its own source says that's not quite true.
Ben Werdmuller argued this week that China's strategy of shipping open weights, not just better benchmarks, is what is actually winning developer adoption away from OpenAI and Anthropic. His clearest piece of evidence turns out to be less clear than he presents it.
The argument
Werdmuller's July 20 essay on werd.io points to Moonshot's Kimi and Alibaba's Qwen as proof that open weights beat closed APIs on a dimension US labs are not optimizing for: the ability to self host, fine tune, and swap models with zero commercial licensing friction. His central data point is a stat from a16z partner Martin Casado, that around 80 percent of startups pitching a16z who use open source models are using Chinese ones.
The stat is narrower than the headline
Casado later clarified on X that the 80 percent figure describes a subset: roughly 16 to 24 percent of all startups a16z sees are using open source models at all, and it is that group where Chinese models dominate. The 80 percent number is real. It just is not "80 percent of startups use Chinese AI," which is how it keeps getting quoted.
What is actually corroborated
The part of Werdmuller's case that holds up independently is pricing. Moonshot's Kimi is priced well below Anthropic and OpenAI's equivalent tiers, per Axios reporting, and open weights let a cost sensitive team self host instead of paying per token indefinitely. That is a real, structural advantage, separate from the contested stat.
Why a build studio cares
We choose models for client builds on cost, latency, and control, not nationality. If cost sensitive teams keep gravitating to cheaper open weight Chinese models regardless of the vendor lock in tradeoff, that is a real signal about where the default stack is heading, caveat and all. We would rather track the honest version of this argument than repeat the flattering one.
Next step: read Werdmuller's essay and Casado's own clarification before citing the 80 percent number anywhere. If you're weighing open versus closed models for a build, write to us at hello@gattyworks.com.