American developers are now routing as much as 46 percent of their A.I. token traffic to Chinese models, according to OpenRouter data reported by CNBC on Tuesday, a figure that has held above 30 percent every week since Feb. 8. The trailing twelve-month average sits at 11 percent. The first-half 2025 figure was 4.5 percent. Hours after CNBC published, Reuters reported that China’s Ministry of Commerce has spent the past month meeting with Alibaba, ByteDance and Z.ai about curbing foreign access to their systems.

The two stories are the same story, told from opposite ends of a supply chain that wasn’t supposed to exist.

The immediate driver is price. Justin Summerville, who works on data and analytics at OpenRouter, described open-source Chinese models as “60 percent to 90 percent cheaper” than the leading systems from Anthropic and OpenAI. Harpreet Arora, head of agentic infrastructure at Vercel, was blunter: “Price is doing the work here.” Z.ai’s GLM-5.2, released in June 2026, closed to within a percentage point of Anthropic’s Opus 4.8 on one widely cited agentic benchmark while costing roughly a fifth as much. In its first full week, GLM-5.2’s daily token volume grew about 27-fold and its customer count about 80-fold.

The migrations aren’t confined to indie developers. Nikkei Asia reports that Coinbase and Uber are running significant workloads on GLM-5.2 and DeepSeek, and that the shift accelerated after the Trump administration imposed and then partially lifted export controls on Anthropic’s frontier Mythos and Fable models. Lindy moved all of its traffic from Claude to DeepSeek in June. Flo Crivello, the startup’s chief executive, said the switch would save his company millions.

Kyle Chan, a fellow at the Brookings Institution, framed the behavioral shift plainly: U.S. companies that were “previously prioritizing A.I. adoption regardless of model” are now “getting more cost-conscious.”

Beijing has noticed. According to three people familiar with the discussions, officials from the Ministry of Commerce and the National Development and Reform Commission have been meeting with the three labs about limits on frontier model release, treating theft of A.I. technology as a national-security offense, and new restrictions on funding domestic A.I. startups. Two of the three said any rules might apply only to future models. The Commerce Ministry, N.D.R.C., Alibaba, ByteDance and Z.ai didn’t respond to Reuters’s requests for comment.

The symmetry is the point. Washington’s export controls were designed to keep American frontier capability out of Chinese hands. What they produced instead was a Chinese open-source stack cheap enough that American enterprises adopted it voluntarily, and a Chinese state now weighing whether that adoption has become a strategic asset worth withholding.