In the summer of 1858, a copper-core cable crossed the Atlantic seabed, linking London and New York.

The significance was never just transmission speed. It was about power. Whoever laid the undersea cable could take a cut from the flow of information. The British Empire used this global telegraph network to keep colonial intelligence, cotton prices, and news of war in its grip.

The empire’s strength lay not only in its fleet, but also in that cable.

More than 160 years later, the same logic is playing out again in an unexpected form.

In 2026, Chinese large models are quietly eating into the global developer market. The latest data from OpenRouter shows that among the top 10 models on the platform, Chinese models account for 61% of token consumption, with all of the top three coming from China. API requests sent every day by developers in San Francisco, Berlin, and Singapore travel across Pacific undersea fiber cables to Chinese data centers, 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 across borders through tokens.

The Great Migration of AI Models

On February 24, 2026, OpenRouter released weekly data showing that total token consumption among the platform’s top 10 models was about 8.7 trillion. Chinese models accounted for 5.3 trillion tokens, or 61%. MiniMax M2.5 landed at No. 1 with 2.45 trillion tokens, followed by Kimi K2.5 and Zhipu GLM-5. All three of the top models came from China.

Latest data as of February 26

This was no accident. A single fuse set everything off.

At the start of this year, OpenClaw appeared seemingly out of nowhere. The open-source tool lets AI truly begin to “work” by directly controlling computers, executing commands, and running complex workflows in parallel. Within weeks, it had passed 210,000 stars on GitHub.

John, a finance professional, installed OpenClaw immediately and connected it to the Anthropic API, using it to automatically monitor stock-market information and deliver timely trading signals. A few hours later, he stared at his account balance for several seconds: dozens of dollars were gone.

That is the new reality OpenClaw introduced. In the past, chatting with AI consumed only a few thousand tokens per conversation, making the cost almost negligible. Once OpenClaw is connected, AI runs more than a dozen subtasks in the background at the same time, repeatedly calling context and iterating in loops. Token consumption is not linear; it is exponential. The bill accelerates like a car with its hood open, the fuel gauge dropping with no way to stop it.

Developer communities soon began circulating a “clever trick”: using OAuth tokens to plug Anthropic or Google subscription accounts directly into OpenClaw, turning monthly “unlimited” plans into free fuel for AI agents. Many developers adopted this method.

The official countermeasures came quickly.

On February 19, Anthropic updated its terms to explicitly ban the use of Claude subscription credentials in third-party tools such as OpenClaw. Anyone seeking Claude functionality must use the API billing channel. Google went further, carrying out broad bans on subscription accounts that accessed Antigravity and Gemini AI Ultra through OpenClaw.

“The world has suffered under Qin for too long,” John said, before promptly turning to Chinese large models.

On OpenRouter, China’s MiniMax M2.5 scores 80.2% on software engineering tasks, while Claude Opus 4.6 scores 80.8%, a gap that is almost negligible. But the price difference is dramatic: MiniMax charges $0.30 per million input tokens, while Claude charges $5, roughly 17 times more.

John switched over. His workflow kept running, and his bill shrank by an order of magnitude. The same migration is now happening worldwide.

OpenRouter COO Chris Clark put it plainly: Chinese open-source models have captured substantial market share because they are unusually heavily represented in agent workflows run by U.S. developers.

Electricity Goes Global

To understand what token exports really mean, you first have to understand the cost structure of a token.

A token looks weightless. One token is roughly equal to 0.75 English words, and an ordinary conversation with AI may consume only a few thousand tokens. But when tokens pile up by the trillions, the physical reality behind them becomes heavy.

Break down the cost of a token and there are only two core components: computing power and electricity.

Computing power is the amortized depreciation of GPUs. If you buy an NVIDIA H100 for roughly $30,000, its lifespan is converted into depreciation cost for each inference run. Electricity is the fuel that keeps data centers running. A single GPU consumes about 700 watts at full load, and once cooling-system overhead is included, the power bill for a large AI data center can easily exceed hundreds of millions of dollars a year.

Now draw this physical process on a map.

A U.S. developer sends an API request from San Francisco. The data leaves California, travels across Pacific undersea fiber cables, and reaches a data center somewhere in China. GPU clusters begin working. Electricity flows from China’s grid into those chips. The inference is completed, and the result is sent back. The entire process may take only one or two seconds.

The electricity never leaves China’s power grid, but its value is delivered across borders through tokens.

Here lies something ordinary trade cannot match: tokens have no physical form, do not pass through customs, cannot be hit by tariffs, and do not even fall under any current trade-statistics category. China is exporting large volumes of computing power and electricity services, but in official goods-trade data, they are almost invisible.

Tokens have become a derivative of electricity. Token exports are, in essence, electricity exports.

This also benefits from China’s relatively low electricity prices. Its blended power costs are about 40% lower than those in the U.S., a physical cost gap that competitors can easily replicate.

Chinese AI large models also have advantages in algorithms and brutal domestic competition.

DeepSeek V3’s MoE architecture activates only part of its parameters during inference. Independent tests show its inference cost is about 36 times lower than GPT-4o’s. MiniMax M2.5 similarly has 229 billion total parameters but activates only 10 billion.

At the top layer is relentless domestic competition. Alibaba, ByteDance, Baidu, Tencent, Moonshot AI, Zhipu, MiniMax, and more than a dozen other companies are crowding the same track, pushing prices below any reasonable profit range. Losing money to win attention has become standard practice in the industry.

Look closely, and this resembles China’s manufacturing push overseas: using supply-chain advantages and fierce internal competition to drive token prices sharply lower.

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 groups of unusual visitors.

These people rented abandoned factories, packed them with rows of machines, and kept them running 24 hours a day. The machines produced nothing. They simply kept solving a mathematical problem, and occasionally, from that endless calculation, they would produce a bitcoin.

This was the first generation of electricity exports: converting cheap hydropower and wind power into globally circulating digital assets through the hash calculations of mining rigs, then cashing them out for dollars on exchanges.

The electricity crossed no border, but its value flowed into global markets through bitcoin.

In those years, China’s computing power once accounted for more than 70% of global Bitcoin mining power. China’s hydropower and coal power took part, through this circuitous route, in a redistribution of global capital.

In 2021, all of it came to an abrupt halt. Regulators cracked down, miners scattered, and computing power migrated to Kazakhstan, Texas, and Canada.

But the logic itself never disappeared. It was simply waiting for a new shell. Then ChatGPT emerged, large models began competing fiercely, and former Bitcoin mining farms were transformed into AI data centers. Mining rigs became computing-power GPUs, bitcoin became tokens, and the one constant was electricity.

Bitcoin exports and token exports share the same underlying logic, but tokens have greater commercial value today.

Mining with mining rigs is pure mathematical calculation. The bitcoin produced is a financial asset whose value comes from scarcity and market consensus, with no relation to “what was computed.” Computing power itself is not productive; it is more like a byproduct of a trust mechanism.

Large-model inference is different. GPUs consume electricity and produce real cognitive services: code, analysis, translation, and ideas. The value of tokens comes directly from their utility to users. This is a deeper form of embedding. Once a developer’s workflow depends on a particular model, the cost of switching rises over time.

There is, of course, another key difference: Bitcoin mining was pushed out of China, while token exports are being actively chosen by global developers.

The Token War

The undersea 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, and they face heavy resistance.

Data sovereignty is the first wall. When a U.S. developer’s API request is processed through a Chinese data center, the data physically passes through China. For individual developers and small applications, this is not a problem. But in scenarios involving sensitive corporate data, financial information, or government compliance, it is a hard constraint. This is also why Chinese models have the highest penetration in developer tools and personal applications, while remaining almost absent from core enterprise systems.

Chip restrictions are the second wall. China’s AI development faces export controls on NVIDIA’s high-end GPUs. MoE architectures and algorithmic optimization can only partly offset that disadvantage. The ceiling still exists.

But today’s resistance is only the prologue. A larger battlefield is taking shape.

Tokens and AI models have become a new dimension of strategic competition between China and the U.S., no less important than semiconductors and the internet in the 20th century, and perhaps closer to an older analogy: the space race.

In 1957, the Soviet Union launched Sputnik 1, shocking the United States and prompting it to launch the Apollo program, pouring in resources equivalent to hundreds of billions of dollars today to ensure it would not lose the space race.

The logic of the AI race is strikingly similar, but its intensity will far exceed the space race. Space, after all, is physical space, and ordinary people do not feel it directly. AI penetrates the capillaries of the economy. Behind every line of code, every contract, and every government decision-making system, a large model from some country may be running. Whichever country’s models become the default infrastructure choice for global developers will gain structural influence over the global digital economy, largely out of sight.

This is exactly what makes China’s token exports truly unsettling for Washington.

When a developer’s codebase, agent workflows, and product logic are all built around the API of a Chinese model, migration costs rise exponentially over time. At that point, even if U.S. lawmakers impose restrictions, developers will resist with their feet, just as programmers today cannot give up GitHub.

Today’s token exports may only be the opening chapter of this long contest. Chinese large models have not claimed they will overturn anything. They are simply delivering services, at lower prices, into the hands of every developer in the world with an API key.

This time, the cable is being laid by engineering teams writing code in Hangzhou, Beijing, and Shanghai, and by GPU clusters running day and night in a province somewhere in southern China.

There is no countdown clock for this contest. It runs 24 hours a day, measured in tokens, with every developer’s terminal as the battlefield.