Token goes overseas, selling Chinese electricity to the world

Author: Black Lobster, Deep Tide TechFlow

In the summer of 1858, a copper-core cable crossed the Atlantic Ocean floor, connecting London and New York.

The significance of this event was never about transmission speed, but about power structures; whoever laid the submarine cable could siphon off information flow. The British Empire relied on this global telegraph network to control intelligence from its colonies, cotton prices, and news of wars.

The strength of the empire was not only in its fleet but also in that cable.

More than 160 years later, this logic is playing out in an unexpected way.

By 2026, China's large models are quietly consuming 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 coming from China. Developers in San Francisco, Berlin, and Singapore are sending API requests daily, which traverse the Pacific submarine cable to reach data centers in China, where computing power is consumed, electricity flows, and results are sent back.

Electricity has never left the Chinese power grid, but its value has been delivered across borders 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 on the platform was approximately 8.7 trillion, with Chinese models alone accounting for 5.3 trillion, or 61%. MiniMax M2.5 topped the list with 2.45 trillion tokens, followed closely by Kimi K2.5 and Zhipu GLM-5, all three from China.

Latest data from February 26

This is not a coincidence; a fuse has ignited everything.

Earlier this year, OpenClaw emerged, an open-source tool that truly allows AI to "work," capable of directly controlling computers, executing commands, and completing complex workflows in parallel. Within weeks, it surpassed 210,000 stars on GitHub.

Financial practitioner John installed OpenClaw immediately and connected it to the Anthropic API, starting to automatically monitor stock market information and report trading signals in real-time. 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, a negligible cost. After integrating OpenClaw, AI runs dozens of sub-tasks in the background, repeatedly calling context and iterating in cycles; token consumption is not linear but exponential. The bill accelerates like a car with its hood open, the fuel gauge dropping, unable to stop.

A "clever trick" quickly spread in the developer community: using OAuth tokens to directly connect Anthropic or Google subscription accounts to OpenClaw, turning monthly "unlimited" quotas into free fuel for AI agents, a method adopted by many developers.

The official countermeasures soon followed.

On February 19, Anthropic updated its agreement, explicitly prohibiting the use of Claude subscription credentials for third-party tools like OpenClaw. To access Claude's features, one must go through the API billing channel. Google even broadly banned subscription accounts accessing Antigravity and Gemini AI Ultra through OpenClaw.

"People have suffered under Qin for too long," John then turned to domestic large models.

On OpenRouter, the domestic large model MiniMax M2.5 scored 80.2% on software engineering tasks, while Claude Opus 4.6 scored 80.8%, a negligible difference. But the price difference is staggering: the former charges $0.3 per million tokens at the input end, while the latter charges $5, a difference of about 17 times.

John switched over, the workflow continued to operate, and the bill shrank by an order of magnitude; this migration is happening globally.

OpenRouter's COO Chris Clark stated directly, the reason Chinese open-source models can capture a large market share is that they occupy an unusually high proportion in the proxy workflows run by American developers.

Electricity Going Abroad

To understand the essence of tokens going abroad, one must first clarify the cost structure of a token.

It seems light; one token is roughly equivalent to 0.75 English words. A typical conversation with AI consumes only a few thousand tokens. But when these tokens stack up in trillions, the physical reality behind them becomes heavy.

Breaking down the cost of tokens, there are only two core components: computing power and electricity.

Computing power is the depreciation of GPUs; when you buy an NVIDIA H100 for about $30,000, its lifespan translates into a depreciation cost for each inference. Electricity is the fuel for the continuous operation of data centers; each GPU consumes about 700 watts at full load, and with cooling system costs, the electricity bill for a large AI data center can easily exceed hundreds of millions of dollars annually.

Now, let's map this physical process.

An American developer sends an API request from San Francisco. The data travels from California, via the Pacific submarine cable, to a data center in China, where the GPU cluster begins to work. Electricity flows from China's power grid to those chips, the inference is completed, and the results are sent back. The entire process may take only one or two seconds.

Electricity has never left China's power grid, but the value of electricity has been delivered across borders through tokens.

There is a magical aspect here that ordinary trade cannot reach: tokens have no physical form, do not need to pass through customs, are not subject to tariffs, and are not even included in any current trade statistics. China has exported a large amount of computing and electricity services, but in official commodity trade data, it is almost invisible.

Tokens have become derivatives of electricity; the essence of tokens going abroad is electricity going abroad.

This is also aided by China's relatively low electricity prices, which are about 40% lower than those in the United States, representing a physical cost difference that competitors can easily replicate.

Additionally, China's AI large models have algorithmic and "involution" advantages.

The MoE architecture of DeepSeek V3 activates only part of the parameters during inference, and independent tests show its inference cost is about 36 times lower than that of GPT-4o. MiniMax M2.5 also has 229 billion total parameters but activates only 10 billion.

At the top level is involution, with companies like Alibaba, ByteDance, Baidu, Tencent, Dark Moon, Zhipu, MiniMax... more than a dozen companies stepping on each other in the same race, driving prices well below reasonable profit margins, making losses while gaining attention a norm in the industry.

Looking closely, this is similar to China's manufacturing going abroad, leveraging supply chain advantages and industry involution to drive token prices down sharply.

From Bitcoin to Tokens

Before tokens, there was another form of electricity going abroad.

Around 2015, power plant managers in Sichuan, Yunnan, and Xinjiang began to welcome a strange group of guests.

These people rented abandoned factories, crammed them with machines, and kept them running 24 hours a day. The machines produced nothing but continuously solved a mathematical problem, occasionally calculating a Bitcoin from this endless mathematical puzzle.

This was the first generation of electricity going abroad: converting cheap hydropower and wind power into globally circulating digital assets through the hashing operations of mining machines, then cashing them out as dollars on exchanges.

Electricity did not cross any borders, but its value flowed into the global market via Bitcoin.

During those years, China's computing power accounted for over 70% of the global Bitcoin mining power. China's hydropower and coal power participated in a global capital redistribution in this roundabout way.

In 2021, all of this came to a halt. Regulatory crackdowns forced miners to scatter, and computing power migrated to Kazakhstan, Texas in the USA, and Canada.

But this logic itself has never disappeared; it was just waiting for a new shell until ChatGPT emerged, and the large models began to compete. Former Bitcoin mining sites transformed into AI data centers, mining machines became computing GPUs, and the Bitcoins produced turned into tokens, with only electricity remaining unchanged.

The logic of Bitcoin going abroad and tokens going abroad is structurally similar, but tokens have more commercial value today.

Mining machines operate purely on mathematical calculations, and the Bitcoins produced are financial assets whose value comes from scarcity and market consensus, unrelated to "what was computed." Computing power itself lacks productivity and is more like a byproduct of a trust mechanism.

Large model inference is different. GPUs consume electricity, and the output is 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 certain model, the cost of switching will accumulate over time.

Of course, there is another key difference: Bitcoin mining was expelled by China, while tokens going abroad are actively chosen by global developers.

The Token War

The submarine cable laid in 1858 represented the British Empire's sovereignty over the information superhighway; whoever owns the infrastructure can define the rules of the game.

Tokens going abroad is also a war without declaration, facing numerous obstacles.

Data sovereignty is the first wall. An API request from an American developer is processed through a Chinese data center, with data physically flowing through China. For individual developers and small applications, this is not an issue, but in scenarios involving sensitive corporate data, financial information, and government compliance, this is a hard injury. This is also why the penetration rate of Chinese models is highest in developer tools and personal applications, but they have almost no presence in core enterprise systems.

The chip ban is the second wall. China's AI development faces export controls on high-end GPUs from NVIDIA; the MoE architecture and algorithm optimizations can only partially offset this disadvantage, and the ceiling still exists.

But the obstacles in front of us are just the prologue; a larger battlefield is taking shape.

Tokens and AI models have become a new dimension of strategic competition between China and the United States, comparable to semiconductors and the internet in the 20th century, and even closer to an older metaphor: the space race.

In 1957, when the Soviet Union launched Sputnik 1, the United States was shocked and immediately initiated the Apollo program, pouring resources equivalent to today's hundreds of billions of dollars to ensure victory in the space race.

The logic of AI competition is strikingly similar, but the intensity will far exceed that of the space race. After all, space is a physical realm that ordinary people cannot perceive, while AI permeates the capillaries of the economy; behind every line of code, every contract, and every government decision-making system, there may be a large model from a certain country running. Whichever model becomes the default option for global developers' infrastructure will gain structural influence over the global digital economy.

This is precisely what makes China's tokens going abroad genuinely unsettling for Washington.

When a developer's codebase, agent workflow, and product logic are built around the API of a certain Chinese model, the cost of migration will rise exponentially over time. By then, even if the U.S. legislates restrictions, developers will resist with their feet, just as no programmer today can abandon GitHub.

Today's tokens going abroad may just be the prologue to this long game. Chinese large models do not claim to overthrow anything; they simply deliver services to every developer with an API key at a lower price.

This time, the ones laying the cables are the engineering teams coding in Hangzhou, Beijing, and Shanghai, along with the GPU clusters operating day and night in a southern province.

This competition has no countdown; it is ongoing 24 hours a day, measured in tokens, with the battlefield being every developer's terminal.