Chongqing - A developer in San Francisco sends an Application Programming Interface (API) request. Seconds later, the response comes back from a data center in China, where GPUs processed the task using electricity from the Chinese grid.
In simple terms, while the request is sent from the U.S., the actual computing work—and the electricity needed to power it—takes place in China, before the result is quickly transmitted back, showing how even digital tasks depend on physical energy across borders.
The power never physically left China. But the value created by that power was packaged in tokens, the unit AI companies use to measure and bill model usage. That is the business logic increasingly taking shape behind the global rise of Chinese AI models.
A conceptual illustration of AI infrastructure, showing how electricity, data centers, and global connectivity are increasingly linked in the business of cross-border AI services. (Graphic/Zheng Ran)
How tokens let China monetize AI power at home
In February 2026, Chinese models surpassed U.S. models for the first time in weekly token usage on OpenRouter, a major AI model API aggregation platform, with U.S. developers accounting for 47.17% of users versus China's 6.01%.
Chinese models logged 4.12 trillion tokens in one week, ahead of the U.S.'s 2.94 trillion, with four Chinese models- MiniMax M2.5, Kimi K2.5, Zhipu GLM-5, and DeepSeek V3.2- occupying four of the top five spots and collectively accounting for 85.7% of the top-five total.
The momentum continued into spring. Chinese large language models processed 12.96 trillion tokens on OpenRouter in the week of March 30 to April 5, up 31.48% from the previous week and marking the fifth straight week that Chinese models outperformed U.S. rivals on the platform.
The key concept is the token. A token is the basic unit AI models use to process text. It is also the unit customers pay for. More tokens usually mean more usage, more computation, and more revenue.
But tokens are also tied to something more tangible: electricity. Changjiang Securities said power costs account for 60% to 70% of large-model operating costs, describing tokens as a kind of "power derivative." In other words, every token reflects not just software capability but also energy consumption, translated into a digital product.
That framing helps explain why Chinese AI's overseas expansion is increasingly being discussed not only as a software story, but as an energy and infrastructure story.
Traditionally, electricity has been hard to export. Transmission is limited by geography. Shipping stored electricity is expensive and inefficient. Tokens change that. They allow China to keep electricity at home while exporting the economic value generated from that electricity through AI inference services. Guotai Haitong Securities said this dynamic could turn overseas demand for computing power into domestic demand for electricity, especially in lower-cost regions in western China.
Price is part of what is driving adoption. According to Changjiang Securities, MiniMax M2.5 and Zhipu GLM-5 charge $0.30 per million input tokens on OpenRouter, compared with $5 for Anthropic's Claude Opus 4.6. That makes the Chinese models roughly one-sixteenth of the price.
China's edge: low-cost power fuels high-value AI exports
Shi Yuxia, a senior engineer at the China Academy of Information and Communications Technology, said China's token export capability comes from three overlapping strengths: better large-model technology, lower energy costs, and supply-chain advantages.
First, there is engineering efficiency. Chinese AI companies widely use Mixture of Experts, or MoE, an architecture that activates only part of a model for simpler tasks instead of using the full system every time. That reduces the amount of computing needed per token.
UCloud Vice President Liu Jie said that Chinese large models benefit from lower training costs due to engineering capabilities and algorithmic optimization. If those models are then deployed in Chinese data centers with relatively lower electricity prices, he said, total costs fall sharply. Zhipu CEO Zhang Peng said large cloud-based foundation models would become even more efficient through economies of scale and highly optimized inference.
Second is the power infrastructure. By the end of 2025, China had 3.89 billion kilowatts of installed generation capacity, with wind and solar contributing 1.84 billion kilowatts, or 47.3% of the total. Its ultra-high-voltage transmission network channels electricity from resource-rich western provinces to eastern demand centers, including major computing hubs.
For the AI industry, that amounts to a national-scale energy platform capable of supporting power-intensive inference and data center growth.
A conceptual illustration of renewable power, transmission infrastructure, and data centers. (Graphic/Zheng Ran)
Third, there is industrial coordination. Fudan University economist Li Zhiqing said China has built reinforcing advantages across AI chips, servers, computing infrastructure, cross-border networks, edge computing, and settlement systems.
One unnamed industry source added that chips still make up the majority of token costs today, while electricity is only part of the equation. In that view, the more strategically important shift would be exporting AI services based on a fully domestic stack of Chinese models and Chinese chips.
That broader ecosystem is also being reinforced by a rapidly expanding domestic user base. As of December 2025, China had 602 million generative AI users, up 141.7% from a year earlier. The surge is pushing AI from novelty to everyday use and into deeper applications such as office collaboration and industrial design, while broader adoption is also generating the data feedback needed for continued model improvement.
Projections from the China Academy of Information and Communications Technology suggest that, in a high-growth scenario, China's computing centers could consume more than 700 billion kilowatt-hours by 2030, or 5.3% of total electricity demand.
The scale-up is already visible on the ground. He Zhaodong, deputy general manager of the network department at China Mobile Zhejiang, said communications-network cabinets used around 2 kilowatts each a decade ago. Today, an average rack in a computing center uses more than 10 kilowatts, with some reaching 100 kilowatts. A small computing center, he said, can now consume as much electricity as a small city.
Li said token exports represent a value-chain upgrade for China's established power, computing, and AI industries rather than the creation of a wholly new sector. By converting low-cost domestic electricity into computing capacity and delivering it to overseas users through large-model APIs, China is effectively exporting AI services with high commercial and broader economic value.