Author: Black Lobster, Deep Tide TechFlow
Original Title: Token Exports: Selling China's Electricity to the World
In the summer of 1858, a copper-core cable was laid across the Atlantic seabed, connecting London and New York.
The significance of this event was never about transmission speed but about power dynamics. Whoever laid the submarine cables could control the flow of information. The British Empire, with its global telegraph network, held intelligence from colonies, cotton prices, and war news in its grasp.
The empire's strength was not just its fleet but also that cable.
Over 160 years later, this logic is replaying in an unexpected way.
In 2026, Chinese large models are quietly dominating the global developer market. The latest data from OpenRouter shows that Chinese models account for 61% of the token consumption among the top ten models on the platform, with the top three all from China. API requests from developers in San Francisco, Berlin, and Singapore are crossing the Pacific submarine cables to data centers in China, where computing power is consumed, electricity flows, and results are sent back.
The electricity never leaves China's power grid, but its value is delivered cross-border through tokens.
The Great Migration of AI Models
On February 24, 2026, OpenRouter released weekly data: the total token consumption of the top ten models was about 8.7 trillion, with Chinese models accounting for 5.3 trillion, or 61%. MiniMax M2.5 topped the list with 2.45 trillion tokens, followed by Kimi K2.5 and Zhipu GLM-5, all from China.
Latest data as of February 26
This is no coincidence; a fuse has been lit.
Earlier this year, OpenClaw emerged, an open-source tool that allows AI to truly work, directly controlling computers, executing commands, and handling complex workflows in parallel. Its GitHub stars surpassed 210,000 within weeks.
John, a finance professional, installed OpenClaw immediately, connected it to the Anthropic API, and began automatically monitoring stock information and generating trading signals. A few hours later, he stared at his account balance in disbelief: dozens of dollars, gone.
This is the new reality brought by OpenClaw. In the past, chatting with AI consumed a few thousand tokens per conversation, with negligible costs. With OpenClaw, AI runs dozens of subtasks in the background, repeatedly calling context and iterating cycles. Token consumption isn't linear; it's exponential. The bill accelerates like a car with the hood open, the fuel gauge dropping, unstoppable.
A "clever trick" soon spread among developer communities: using OAuth tokens to directly connect Anthropic or Google subscription accounts to OpenClaw, turning the "unlimited" monthly quota into free fuel for AI agents. This became a common method for many developers.
Official countermeasures followed.
On February 19, Anthropic updated its terms, explicitly prohibiting the use of Claude subscription credentials for third-party tools like OpenClaw. To access Claude functionality, users must use the API billing channel. Google went further, widely banning subscription accounts that accessed Antigravity and Gemini AI Ultra via OpenClaw.
"The world has suffered long under tyranny." John immediately switched to domestic large models.
On OpenRouter, the domestic model MiniMax M2.5 scores 80.2% on software engineering tasks, while Claude Opus 4.6 scores 80.8%, a negligible difference. But the prices are worlds apart: the former charges $0.3 per million input tokens, while the latter charges $5, a difference of about 17 times.
John switched over. His workflow continued to run, and his bill shrank by an order of magnitude. This migration is happening globally.
OpenRouter's COO, Chris Clark, put it bluntly: Chinese open-source models are capturing a significant market share because they account for an unusually high proportion of the agent workflows run by U.S. developers.
Electricity Exports
To understand the essence of token exports, one must first understand the cost structure of a token.
It seems light—one token is roughly equivalent to 0.75 English words. An ordinary conversation with AI consumes only a few thousand tokens. But when these tokens stack up in trillions, the underlying physical reality becomes heavy.
Breaking down the cost of a token, there are only two core components: computing power and electricity.
Computing power is the depreciation of GPUs. Buying an NVIDIA H100 costs about $30,000, and its lifespan, converted to each inference, is the depreciation cost. Electricity is the fuel for data centers to operate continuously. A GPU consumes about 700 watts at full load, and with cooling systems, the electricity bill for a large AI data center can easily exceed hundreds of millions of dollars annually.
Now, map this physical process.
A U.S. developer in San Francisco sends an API request. The request travels from California via Pacific submarine cables to a data center in China. GPU clusters start working, electricity flows from China's grid to those chips, inference is completed, and results are sent back. The entire process takes maybe a second or two.
Electricity never leaves China's power grid, but its value, through tokens, completes cross-border delivery.
Here's something miraculous that ordinary trade cannot achieve: tokens have no physical form, do not need to pass through customs, are not subject to tariffs, and are almost invisible in any current trade statistics. China exports a vast amount of computing power and electricity services, but in official commodity trade data, it is nearly invisible.
Tokens have become derivatives of electricity. Token exports are essentially electricity exports.
This is also thanks to China's relatively low electricity prices, which are about 40% lower than those in the U.S. This is a physical cost difference that competitors cannot easily replicate.
Additionally, Chinese AI models have algorithmic and "internal competition" advantages.
DeepSeek V3's MoE architecture activates only part of the parameters during inference, with independent tests showing its inference cost is about 36 times lower than GPT-4o. MiniMax M2.5 similarly activates only 10B of its 229B total parameters.
At the top layer is internal competition. Alibaba, ByteDance, Baidu, Tencent, Moon Dark Side, Zhipu, MiniMax... over a dozen companies are competing on the same track, driving prices below reasonable profit margins. Losing money to gain traction is already the industry norm.
Looking closely, this mirrors China's manufacturing exports: leveraging supply chain advantages and internal competition to drastically lower token prices.
From Bitcoin to Tokens
Before tokens, there was another form of electricity export.
Around 2015, power station managers in Sichuan, Yunnan, and Xinjiang began receiving strange visitors.
These people rented abandoned factories, filled them with dense rows of machines, and kept them running 24/7. The machines produced nothing but continuously solved a mathematical problem, occasionally mining a Bitcoin from this endless calculation.
This was the first-generation form of electricity exports: converting cheap hydropower and wind power, via mining machines' hash computations, into globally circulating digital assets, which were then cashed out into dollars on exchanges.
Electricity never crossed any border, but its value, embodied in Bitcoin, flowed to the global market.
During those years, China's computing power once accounted for over 70% of global Bitcoin mining power. China's hydropower and coal power, in this roundabout way, participated in a global redistribution of capital.
In 2021, this came to an abrupt halt. Regulatory crackdowns followed, miners scattered, and computing power migrated to Kazakhstan, Texas, and Canada.
But the logic itself never disappeared; it was just waiting for a new shell. Until ChatGPT emerged, large models began competing, and former Bitcoin mining farms transformed into AI data centers. Mining machines became computing GPUs, and the Bitcoin produced became tokens. The only constant was electricity.
Bitcoin exports and token exports are structurally similar in underlying logic, but tokens hold more commercial value today.
Bitcoin mining is pure mathematical computation, producing Bitcoin as a financial asset. Its value comes from scarcity and market consensus, unrelated to "what was computed." The computing power itself is unproductive, more like a byproduct of a trust mechanism.
Large model inference is different. GPUs consume electricity to produce real cognitive services: code, analysis, translation, creativity. The value of tokens comes directly from their utility to users. This is a deeper embedding; once a developer's workflow relies on a model, the cost of switching increases over time.
Of course, there is another key difference: Bitcoin mining was driven out of China, while token exports are actively chosen by global developers.
Token War
The submarine cable laid in 1858 represented the British Empire's sovereignty over the information highway. Whoever owned the infrastructure could define the rules of the game.
Token exports are also an undeclared war, full of resistance.
Data sovereignty is the first barrier. An API request from a U.S. developer processed by a Chinese data center means data physically flows through China. For individual developers and small applications, this isn't an issue. But for enterprise-sensitive data, financial information, and government compliance scenarios, this is a hard stop. This is why Chinese models have the highest penetration in development tools and personal applications but almost no presence in core enterprise systems.
The chip ban is the second barrier. China's AI development faces export controls on NVIDIA's high-end GPUs. MoE architecture and algorithmic optimization can only partially offset this disadvantage; the ceiling still exists.
But the current resistance is just the prologue; a larger battlefield is forming.
Tokens and AI models have become a new strategic dimension in the U.S.-China rivalry, comparable to semiconductors and the internet in the 20th century, or even closer to an older analogy: the space race.
In 1957, the Soviet Union launched Sputnik 1, shocking the U.S., which then launched the Apollo program, investing resources equivalent to hundreds of billions of dollars today to ensure it did not fall behind in the space race.
The logic of the AI race is strikingly similar but will be far more intense than the space race. Space is physical, something ordinary people don't feel. AI渗透的是经济的毛细血管,渗透到每一行代码、每一份合同、每一个政府决策系统的背后,都可能运行着某个国家的大模型. Whose model becomes the default infrastructure choice for global developers gains structural influence over the global digital economy.
This is what truly worries Washington about China's token exports.
When a developer's codebase, agent workflows, and product logic are built around a Chinese model's API, migration costs rise exponentially over time. Even if the U.S. legislates restrictions, developers will resist with their feet, just as no programmer can abandon GitHub today.
Today's token exports may be just the opening chapter of this long game. Chinese large models claim no intent to颠覆什么; they simply deliver services to every global developer with an API key at a lower price.
This time, those laying the cables are the engineering teams writing code in Hangzhou, Beijing, and Shanghai, and the GPU clusters running day and night in a southern province.
This rivalry has no countdown; it happens 24 hours a day, measured in tokens, on every developer's terminal.
Twitter:https://twitter.com/BitpushNewsCN
Bitpush TG Discussion Group:https://t.me/BitPushCommunity
Bitpush TG Subscription: https://t.me/bitpush
