A programmer in San Francisco types a prompt and hits "enter". In milliseconds, the request travels through a fiber-optic cable beneath the Pacific Ocean to a data center thousands of miles away in western China. There, rows of GPUs – fed by abundant green energy – crunch the query and beam the answer back to California almost instantly. What actually crossed the ocean were not physical goods but “tokens” – the tiny digital units that measure how artificial intelligence models read and generate text. China is using its massive reserves of cheap electricity and computing power to process tokens. To a great extent, this invisible, frictionless trade marks a new form of “token export” and a historic pivot for the world's second-largest economy.
In a watershed moment in mid-February, Chinese AI models processed 4.12 trillion tokens on the global aggregator platform OpenRouter, surpassing U.S. models' 2.94 trillion for the first time. The following week, Chinese developers – including MiniMax and DeepSeek – claimed four of the platform's top five spots, heavily driven by U.S. users who make up nearly half of the platform's base. Exporting tokens has quietly become China's most efficient value-added energy trade. While raw power in China sells for roughly CNY0.5 per kilowatt-hour if exported, converting that same electricity into AI computing power and selling it as tokens yields an estimated 22-fold increase in value. Industry insiders point out that the low cost is only part of the equation. Shi Yuxia, Senior Engineer at the China Academy of Information and Communications Technology, said: “Electricity prices are not the core factor allowing Chinese tokens to crush foreign competitors on cost. It is the combination of energy cost, improved Al technical capabilities, and supply chain dominance”.
Technologically, almost all leading Chinese Al large models are open-sourced. Chinese developers have also adopted lightweight Mixture of Experts (MoE) architectures. Instead of activating the entire artificial “brain” for every simple question, MoE models wake up only the specific neural pathways needed for a given query. This drastically cuts the computing power – and the electricity – needed to produce a single token.
To maximize this, the Chinese government has included computing-electricity synergy as a national priority in this year's Government Work Report. This infrastructure push connects China's world-leading ultra-high-voltage grid with its tech sector, allowing high-latency AI training to soak up cheap, abundant green energy in the remote west, while real-time inference runs closer to eastern tech hubs. The resulting cost-to-performance ratio has made Chinese models the quiet engines behind many Western software platforms, driving a 421% global surge in Chinese token consumption over the past year. But geopolitical considerations may still limit China's token exports, the China Daily reports.
