--- title: "Why Nearly Half of American AI Tokens Now Run on Chinese Open Models" date: 2026-09-03 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-chinese-models-us-tokens language: en --- # Why Nearly Half of American AI Tokens Now Run on Chinese Open Models > On OpenRouter, the share of tokens US companies spend on Chinese AI models has sat above 30% every week since February 2026 and peaked at 46%. Engineers are not switching for ideology — they are switching because the models are good and roughly fifty times cheaper. The most reliable measure of whether a technology is actually being used is not a benchmark. It is a bill. In 2026, a growing share of the world's AI bills are being paid to run models built in China — including by American companies. ## Key takeaways - **Usage, not downloads:** the share of tokens US companies consumed on Chinese AI models via OpenRouter has been **above 30% every week since 8 February 2026**, reaching as high as **46%**. The trailing 12-month average was **11%**; in the first half of 2025 it was **4.5%**. - **Open source is winning the default slot:** on OpenRouter, the open-source share of tokens processed jumped from **34% in January 2026 to 65% in June 2026**, with **more than 500 organisations** switching from proprietary to open models. - **The price gap is enormous:** DeepSeek V4 Flash has been listed at **$0.09 per million input tokens and $0.18 per million output tokens**, against roughly **$5 and $30** for the leading US frontier model — about **55x on input and 166x on output**. - **The performance gap has nearly closed:** Stanford's *2026 AI Index* put the gap between top Chinese and American models at **2.7%** on Code Arena as of March 2026, with the two countries repeatedly trading the lead. - **Named adopters:** Coinbase, DoorDash's co-founder, and the startup Lindy have all publicly described moving production work to Chinese models. ## The switch, in the customers' own words The pattern in public statements is remarkably consistent: not "Chinese models are better," but "Chinese models are good enough, and the arithmetic is undeniable." - **Coinbase** CEO Brian Armstrong wrote in June that the exchange was experimenting with defaulting to open-weight models such as **GLM 5.2 and Kimi 2.7** through its internal LLM gateway, while still letting engineers pick the right model for each task — with AI spending reported to have fallen by roughly half. - **Lindy**, a San Francisco startup, moved **100% of its production traffic** from Anthropic's Claude to **DeepSeek V4**. Its CEO said the switch saved millions of dollars and *improved* performance on many core use cases. - **Andy Fang**, co-founder of DoorDash, said the company now routes "lower-level work" to **Kimi K2.6** and reserves US frontier models for the hardest tasks — a combination he said vastly outperformed the previous all-US setup at lower cost, as reported by the Financial Times. What these have in common is that none of them are research experiments. They are production systems with uptime requirements. ## Why the gap closed Three things happened at once. **Capability.** Chinese labs closed the practical gap in exactly the workloads enterprises care about — coding and agentic behaviour. Zhipu's **GLM-5.2** was reported to have taken first place overall in the Code Arena blind leaderboard with over a million participating users. MiniMax M3, Qwen3-Max-Thinking and DeepSeek V4 each carved out differentiated strengths in long context, code generation and large-parameter inference. **Cost.** Efficiency became a strategy rather than a constraint. A Chinese AI executive quoted by Chinese media framed it directly: great models can be built "not by stacking computing power, but by relying on efficiency and innovation" — an argument shaped by necessity under export controls and vindicated in the market. **Licensing.** Chinese models have largely shipped under permissive licences that allow modification and commercial use. For companies outside the United States, that removes a specific fear: dependence on a supplier that can be switched off by an export rule. ## The generational shift The numbers are stark when placed in sequence. Chinese open-source models accounted for **41% of global large-model downloads** over the past year on one major international open-source platform, and a joint MIT–platform report found that Chinese-developed open models **overtook the United States in global downloads for the first time**, taking first place worldwide. Alibaba's **Qwen** passed **1 billion cumulative Hugging Face downloads** in January 2026, and by March accounted for **over half of global open-source model downloads** — a reversal from roughly four years earlier, when US models held about 60% of that share. The shift from 4.5% to 46% in about eighteen months is not a marketing story. It is a migration. *Usage figures are as reported by OpenRouter and cited by CNBC in July 2026; pricing as listed on OpenRouter in June 2026; adoption examples from company statements reported in the Financial Times and Chinese media.* --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chinese-models-us-tokens Free to quote with attribution and a link to the original.