Author: Black Lobster , Deep Tide TechFlow
In the summer of 1858, a copper-core cable crossed the Atlantic Ocean, connecting London and New York.
The significance of this event never lay in the speed of transmission, but in the power structure. Whoever laid the undersea cable could profit from the flow of information. The British Empire, through this global telegraph network, held the intelligence of its colonies, the price of cotton, and news of war.
The empire's strength lies not only in its fleet, but also in that cable.
More than 160 years later, this logic is being repeated in an unexpected way.
In 2026, Chinese large-scale models are quietly devouring the global developer market. Latest data from OpenRouter shows that Chinese models account for 61% of the token consumption of the platform's top ten models, with the top three all originating from China. API requests sent daily by developers in San Francisco, Berlin, and Singapore are traveling across the Pacific Ocean via undersea fiber optic cables to data centers in China, where computing power is consumed, electricity flows, and results are transmitted back.
Electricity never leaves the Chinese 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: the total token consumption of the platform's top ten models was approximately 8.7 trillion, with Chinese models accounting for 5.3 trillion, or 61%. MiniMax M2.5 debuted at the top with 2.45 trillion tokens, followed by Kimi K2.5 and Zhipu GLM-5. The top three were all from China.
Latest data as of February 26
This was no accident; a single spark ignited everything.
At the beginning of this year, OpenClaw emerged, an open-source tool that truly enables AI to "do work." It can directly control computers, execute commands, and complete complex workflows in parallel. Within weeks, its GitHub stars exceeded 210,000.
John, a finance professional, immediately installed OpenClaw and connected to the Anthropic API to automatically monitor stock market information and provide timely trading signals. A few hours later, he stared at his account balance for a few seconds and was stunned: tens of dollars gone.
This is the new reality brought about by OpenClaw. In the past, chatting with AI involved thousands of tokens per conversation, with negligible costs. After integrating OpenClaw, the AI runs more than a dozen subtasks in the background simultaneously, repeatedly calling the context and iterating in loops. The token consumption is not linear, but exponential. The bill is like a car with its hood open accelerating, the fuel gauge dropping, and it just won't stop.
A clever trick quickly emerged in the developer community: use OAuth tokens to directly connect Anthropic or Google subscription accounts to OpenClaw, turning the monthly "unlimited" credits into free fuel for the AI Agent. This is a method adopted by many developers.
Official countermeasures followed immediately.
On February 19th, Anthropic updated its agreement, explicitly prohibiting the use of Claude subscription credentials for third-party tools such as OpenClaw. Access to Claude features must now be through the API billing channel. Google has also significantly reduced the number of subscription accounts that accessed Antigravity and Gemini AI Ultra via OpenClaw.
"The people have suffered under Qin for too long," Jhon then turned to domestically made large-scale models.
On OpenRouter, the domestically developed large-scale model MiniMax M2.5 scored 80.2% on the software engineering task, while Claude Opus 4.6 scored 80.8%, a negligible difference. However, their prices are vastly different: the former costs $0.3 per million tokens of input, while the latter costs $5, a difference of approximately 17 times.
John switched over, the workflow continued, and the bill shrank by an order of magnitude. This migration is happening simultaneously around the world.
Chris Clark, COO of OpenRouter, put it bluntly: the reason why Chinese open-source models have been able to capture a large market share is because they account for an unusually high percentage of the proxy workflows run by US developers.
Power going overseas
To understand the essence of token going global, one must first understand the cost structure of a token.
It seems lightweight; one token is roughly equivalent to 0.75 English words, and a typical conversation with AI consumes only a few thousand tokens. But when these tokens are stacked in trillions, the underlying physical reality becomes incredibly complex.
Breaking down the cost of a token, there are only two core components: computing power and electricity.
Computing power is the depreciation and amortization of GPUs. If you buy an Nvidia H100 for about $30,000, its lifespan, calculated over each inference iteration, is the depreciation cost. Electricity is the fuel that keeps data centers running. Each GPU consumes about 700 watts at full load. Adding the cost of the cooling system, the electricity bill for a large AI data center can easily exceed hundreds of millions of dollars per year.
Now, draw this physical process on a map.
An American developer in San Francisco sends an API request. The data travels from California, via a submarine fiber optic cable across the Pacific Ocean, to a data center in China. The GPU cluster starts working, electricity flows from China's power grid to the chips, inference is completed, and the results are sent back. The whole process may only take a second or two.
Electricity has never left China's power grid, but the value of electricity has been delivered across borders through tokens.
Herein lies a remarkable aspect unattainable by ordinary trade: tokens have no physical form, do not pass through customs, are not subject to tariffs, and are not even included in any current trade statistics. China exports a large amount of computing power and electricity services, but they are virtually invisible in official commodity trade data.
Tokens have become a derivative of electricity, and the overseas expansion of tokens is essentially the overseas expansion of electricity.
This is also thanks to China's relatively low electricity prices, which are about 40% lower than those in the United States. This is a cost difference at the physical level that competitors can easily replicate.
In addition, China's large-scale AI models also have advantages in algorithms and "involution".
DeepSeek V3's MoE architecture allows only a subset of parameters to be activated during inference. Independent tests show that its inference cost is about 36 times lower than GPT-4o. The MiniMax M2.5 also has 229B total parameters, but only 10B are activated.
The top level is involution, with more than a dozen companies such as Alibaba, ByteDance, Baidu, Tencent, Lunar Dark Side, Zhipu, and MiniMax trampling each other on the same track. Prices have long since fallen below the reasonable profit range, and losing money to gain publicity has become the norm in the industry.
Upon closer inspection, this is similar to the "Made in China" initiative going global, leveraging supply chain advantages and industry competition to drive down token prices.
From Bitcoin to Tokens
Before the token, there was another instance of electricity going overseas.
Around 2015, power plant managers in Sichuan, Yunnan, and Xinjiang began to receive a group of strange guests.
These people rented abandoned factory buildings, filled them with densely packed machines, and kept them running 24 hours a day. The machines didn't produce anything; they just kept doing a math problem, and occasionally, from this endless math problem, they would calculate a Bitcoin.
This is the first generation of electricity export: cheap hydropower and wind power are converted into globally circulating digital assets through hash calculations by mining machines, and then converted into US dollars on exchanges.
Electricity doesn't cross any borders, but its value, represented by Bitcoin, flows to global markets.
During those years, China's computing power once accounted for more than 70% of the global Bitcoin mining computing power. China's hydropower and coal-fired power, in this roundabout way, participated in a global redistribution of capital.
In 2021, all of this came to an abrupt end. Regulatory hammer fell, miners scattered, and computing power migrated to Kazakhstan, Texas in the United States, and Canada.
But this logic itself never disappeared; it was just waiting for a new shell. Until ChatGPT emerged, large-scale models competed fiercely, and the former Bitcoin mining farms were transformed into AI data centers. Mining machines became computing GPUs, and the Bitcoins produced became tokens. The only thing that remained unchanged was electricity.
The overseas expansion of Bitcoin and the overseas expansion of tokens are isomorphic in their underlying logic, but tokens currently have greater commercial value.
Mining is a purely mathematical calculation, and the resulting Bitcoin is a financial asset. Its value comes from scarcity and market consensus, and has nothing to do with "what is being calculated." Computing power itself is not productive; it is more like a byproduct of a trust mechanism.
Large-scale model inference is different. GPUs consume power but produce real cognitive services: code, analysis, translation, and creativity. The value of a token comes directly from its utility to the user. This is a deeper embedding, and once a developer's workflow depends on a particular model, the cost of switching will increase over time.
Of course, there is another key difference: Bitcoin mining was driven out of China, while the overseas expansion of tokens was a choice made actively by global developers.
Token Wars
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.
Going global with tokens is also a war without a declaration of war, and it faces numerous obstacles.
Data sovereignty is the first barrier. An API request from an American developer is processed through a Chinese data center, meaning the data physically flows through China. This isn't a problem for individual developers and small applications, but it's a major vulnerability when it involves sensitive corporate data, financial information, or government compliance. This is why the Chinese model has the highest penetration rate in development tools and personal applications, but is almost non-existent in core enterprise systems.
The chip ban is the second barrier. China's AI development faces export controls on Nvidia's high-end GPUs. The MoE architecture and algorithm optimization can only partially offset this disadvantage, and the ceiling still exists.
But the current obstacles are just the beginning; a larger battlefield is taking shape.
Tokens and AI models have become a new dimension of strategic competition between China and the United States, no less significant than the semiconductor and internet competitions of the 20th century, and even closer to an older metaphor: the space race.
In 1957, the Soviet Union launched Sputnik 1, shocking the United States. The US then launched the Apollo program, investing resources equivalent to hundreds of billions of dollars today, determined not to lose in the space race.
The logic behind the AI race is strikingly similar, but its intensity will far surpass that of the space race. Space is, after all, a physical space, imperceptible to the average person; AI, however, permeates the capillaries of the economy. Behind every line of code, every contract, and every government decision-making system, there may be a large-scale model running for a particular country. Whoever's model becomes the default infrastructure option for global developers gains, implicitly, a structural influence on the global digital economy.
This is precisely what makes Washington truly uneasy about the overseas expansion of Chinese tokens.
When a developer's codebase, agent workflow, and product logic are all built around an API based on a specific Chinese model, the migration costs will increase exponentially over time. At that point, even if the US legislates restrictions, developers will resist it outright, just as no programmer can abandon GitHub today.
Today's token going global may just be the beginning of a long game. The Chinese giants haven't claimed to be disrupting anything; they're simply delivering services to every developer worldwide with an API Key at a lower price.
This time, the cables were laid by teams of engineers writing code in Hangzhou, Beijing, and Shanghai, and a GPU cluster operating day and night in a southern province.
This battle has no countdown; it takes place 24 hours a day, the unit is tokens, and the battlefield is every developer's terminal.
