For the past five months, American software developers have been running a quiet experiment in cost arbitrage — and the results are reshaping the global AI market. Since February 8, 2026, the share of tokens processed by US companies on Chinese AI models via the OpenRouter gateway has exceeded 30 percent every single week. The peak reached 46 percent. In the twelve months before that, the average was just 11 percent. In the first half of 2025, it was 4.5 percent.
The numbers, reported by CNBC, tell a story that is less about ideology than economics. Open-source Chinese models cost 60 to 90 percent less than leading offerings from Anthropic and OpenAI, according to Justin Summerville of OpenRouter. For companies spending millions of dollars a year on AI inference, that gap is not an abstraction but a line item that determines whether a product is profitable.
The Lindy Moment: One Startup’s Full Migration to DeepSeek
The clearest illustration of what is happening came in June, when Lindy, a US AI startup, moved 100 percent of its traffic from Anthropic’s Claude to DeepSeek. Lindy CEO Flo Crivello described the moment in terms that resonated across the developer community: “We did it, and you could see that cost curve go down, like, crash to the ground.” Crivello said the switch would save Lindy millions of dollars within months.
Lindy is not alone. Coinbase has claimed to have nearly halved its AI spending by defaulting to cheaper models including Chinese alternatives. Uber reportedly exhausted its entire 2026 budget for Anthropic models in only a few months. Meta, Amazon, Tesla, and Adobe have all moved to restrict employee access to the most expensive US AI models. JPMorgan Chase blocked its Hong Kong staff from accessing Anthropic’s models entirely, citing contractual obligations under the export control framework.
The migration is visible in platform data. On Vercel’s developer infrastructure, DeepSeek’s share of gateway tokens rose steadily from May to June. Z.ai’s GLM-5.2, released on June 13 under an MIT license at roughly one-sixth the cost of GPT-5.5, posted the fastest adoption of any model Vercel has tracked in 2026: in its first full week, daily token volume grew approximately 27 times, and the number of customers using it grew roughly 80 times. GLM-5.2 now ranks fifth on OpenRouter by usage, having entered the top five in under a month. On LaunchLemonade, it is also in the top five.
Performance Closing the Gap With US Frontier Models
The adoption is not driven solely by price. GLM-5.2 landed within a percentage point of Anthropic’s Opus 4.8 on one closely watched agentic benchmark, at approximately one-fifth the cost. Harpreet Arora, head of agentic infrastructure at Vercel, put it plainly: “Price is doing the work here.” But several researchers have noted that GLM-5.2 can perform on par with top US labs on some cybersecurity benchmarks, a capability that has drawn attention from both enterprise buyers and national security analysts.
Kyle Chan of the Brookings Institution estimates that Chinese models remain six to nine months behind the top US rivals in overall capability. But he argues they are “particularly attractive to American companies now as AI costs skyrocket”, a framing that captures the core dynamic: the capability gap is narrow enough that cost becomes the deciding factor in most commercial use cases.
The six most popular AI models on OpenRouter are now Chinese. Chinese models accounted for nearly half of all open-source AI downloads from Hugging Face between February 2025 and 2026.
The Geopolitical Dimension of Developer Dependency
Yacine Jernite of Hugging Face has articulated the structural risk that this trend creates: “There is a real risk that users get stuck having to choose between performant but expensive US proprietary models, or using Chinese models as the only feasible alternative.” That binary is already taking shape in markets outside the US and Europe, where developers in Southeast Asia, Latin America, and Africa are building on DeepSeek and GLM-5.2 not as a political statement but as the most economical choice available.
The US administration is aware of the dynamic. At the end of June, OpenAI limited the rollout of new models at the government’s request. The administration is also looking at regulating the most powerful AI models and halting the rapid adoption of overseas alternatives. But as Kyle Siler-Evans of the RAND Corporation noted, “The scenario to worry about is China has good-enough models at a quarter of the price. I think that is likely the future we’re headed toward.”
The median Ramp customer spends just $11 per employee on AI, according to Ara Kharazian of Ramp, who argued that OpenAI and Anthropic will have “ample time to respond.” Whether that response comes in the form of price cuts, capability leaps, or regulatory protection will define the next phase of the US-China AI competition.