Chinese open-weight models are quietly eating US enterprise AI

What happened

New platform data shows Chinese-origin AI models now account for between 30% and 46% of enterprise API token usage flowing through US developer platforms — and they’ve held above the 30% line every week since early February. The current poster child is GLM-5.2 from Z.ai (Zhipu AI’s international brand), which Vercel says saw the fastest adoption of any model it has tracked in 2026: roughly 27x growth in daily token volume and about 80x growth in customer count in its first full week. Vercel’s CEO said he was “genuinely impressed, almost shocked” at its coding ability, and the model landed on Vercel’s AI Gateway within days of release.

The pitch is not subtle: open-weight Chinese models like GLM-5.2, DeepSeek V4, and Kimi run roughly 60-90% cheaper than the leading Anthropic and OpenAI models while landing within about a percentage point on agentic benchmarks. Coinbase reportedly cut its AI spend nearly in half by switching, even as its token consumption went up.

Why it matters

For two years the frontier-lab story was that quality justified the premium. That gap is now around a rounding error for a lot of workloads, and CFOs have noticed. GLM-5.2 also ships open weights under an MIT license, so the fallback isn’t just “a cheaper API” — it’s “run it yourself.” Meanwhile the supply-chain and security questions about routing enterprise workloads through Chinese-origin models are getting louder in exact proportion to the cost savings. Nobody said procurement was going to be boring this year.

The wry bit

US labs spent the week launching models cleared by government security reviews. US enterprises spent the week routing up to 46% of their tokens to models from Beijing. The invisible hand has apparently read the pricing page.