AI Token en el extranjero: electricidad barata y muros infranqueables

China accounts for 61% of the global AI Token call volume, but the money earned may be less than a fraction of what Anthropic makes.

A late night in February 2026, in the Mission District of San Francisco, an Indian-born developer named Arjun stared blankly at the bill on his screen. He was running an automated code review agent workflow using Claude. Over a dozen sub-tasks in parallel, repeatedly calling context; the consumption of AI Tokens was not linear, but exponential. That night, dozens of dollars were burned.

The next day, he switched to MiniMax’s M2.5 model on OpenRouter—the world’s largest AI model aggregation platform. With the same workflow, the bill shrank by an order of magnitude. The code ran the same, with little difference in results.

What Arjun didn’t know was that each of his requests was originating from California, traveling across the Pacific Ocean via submarine cable, arriving at a data center in northwest China. GPU clusters powered on, electricity flowing from the national grid to the chips, inference completed, results returned in under two seconds.

Electricity never left China’s grid, but the value of electricity, through large models, carried by AI Tokens, has gone overseas.

This is not just a developer’s choice; all of Silicon Valley is engaged in “full token-maxxing.” Arjun is just a microcosm of this global AI Token consumption frenzy.

On the other side of the Pacific, the scene is starkly different.

In spring of the same year, on the Gobi Desert in northwest China, several data centers that had been bustling two years earlier had fallen silent. An old-timer, who had been building compute centers for years, told me during a period of FOMO:

“AI Token going overseas isn’t as simple as many media say.”

He wasn’t talking about technology, but digestion. Self-built compute centers by large model companies can handle their own digestion—this is the optimal path; telecom operators are second best, with channels to backstop; the most awkward are those private capital ventures chasing the AI trend—seemingly sunny, but without anyone to back them up.

The cross-border firewall is insurmountable—those abandoned or semi-idle server rooms in northwest China aren’t unbuildable; it’s that after building, they realize: the wall is even higher than the power station.

On one side, Silicon Valley engaged in full token-maxxing; on the other, idle data centers in northwest China. China’s cheapest electricity has produced the world’s cheapest AI Tokens. But how much value can Chinese AI companies actually capture from this?

1. Ten years ago, another group of guests visited the same power stations

Around 2015, managers of power stations in Sichuan, Yunnan, and Xinjiang began hosting a strange group of visitors. They rented abandoned factories, packed them with dense machines, and operated 24/7. The machines produced nothing but kept solving a math problem repeatedly. Occasionally, from endless computations, a Bitcoin would emerge.

That was the 1.0 version of electricity going overseas. Cheap hydropower was hashed by mining rigs into digital assets circulating globally, then cashed out on exchanges for USD. Electricity never crossed borders, but its value, carried by Bitcoin, flowed worldwide. At its peak, China accounted for over 70% of global Bitcoin hash power.

Bitcoin’s value capture path was extremely short—electricity turned into computation, computation into BTC, BTC into USD. No reliance on anyone in between. Bitcoin itself is a terminal product, a form of energy-backed digital gold, which can be cashed out right at the mine. It can go overseas naked.

But a short path means a shallow foundation.

In 2021, a regulatory crackdown scattered miners, and the value capture instantly plummeted to zero. Hash power migrated to Kazakhstan, Texas, Canada. Power plants continued to produce electricity, but that electricity no longer translated into USD.

The logic of electricity didn’t disappear; it was just waiting for a new shell.

After ChatGPT’s emergence, the same power stations, the same factories, some even with the same power contracts, transformed into AI data centers. Mining machines replaced by GPUs, Bitcoin replaced by AI Tokens.

But AI Tokens are not Bitcoin.

Bitcoin is a terminal product; AI Tokens are semi-finished. AI Tokens must go through layers of model, product, and workflow packaging to become something customers are willing to pay for. They cannot go overseas naked. The industrial chain built around this semi-finished product is far more complex—ranging from underlying green electricity and liquid cooling data centers, to AI chips and servers, to upper-layer large model APIs, aggregation platforms, and cross-border compliance, with seven interconnected layers.

The chain has lengthened. Every layer is vulnerable to hijacking.

What remains unchanged are electricity and one unchanging question: how much value can this time leave behind?

2. The journey of a kilowatt-hour, and its broken bridge

To answer this, let’s follow a kilowatt-hour.

Every time Arjun presses Enter, five links pass the baton: electricity → compute power → model training → model inference → AI Token delivered to his screen.

The first three steps are long completed. A kilowatt-hour from a Sichuan hydro plant flows into a data center in Inner Mongolia, powering GPU clusters for months, feeding trillions of data points, training a large model. After training, the model’s “recipe” is fixed. China’s electricity advantage is fully encoded here—lower training costs, more efficient architecture design, engineering optimizations driven by fierce competition among a dozen companies. The energy of a kilowatt-hour is compressed into a few hundred GB file.

The last two steps are happening right now. Every time Arjun presses Enter, a data center must start inference: load the recipe, consume compute and electricity, produce a batch of AI Tokens on-site, and send back results. Each AI Token’s birth requires real-time electricity.

The question is: which data center?

2.1 Two completely different paths for AI Token going overseas

Arjun is now following the first— inference completed in China, AI Tokens delivered cross-border via API. Requests cross the Pacific to Guizhou, then return to San Francisco. The 0.38 RMB/kWh green electricity directly lowers the marginal cost of each AI Token. The low price Arjun enjoys is essentially China’s hydropower paying the bill, with the electricity advantage fully realized.

The second—moving the recipe overseas, doing inference in Virginia or Singapore. Companies like Zhipu, DeepSeek deploy on Microsoft Azure; MiniMax on Amazon AWS. They use local electricity and GPUs, and each AI Token’s cost bears no relation to Chinese electricity prices.

2.2 Coca-Cola never exports bottled water

The second model is like Coca-Cola—never shipping bottled water from U.S. factories worldwide; it exports the formula, which is then made locally with local water and bottling lines.

In Alibaba Cloud’s Baolian backend, there’s a dropdown menu: deployment scope—“Mainland China,” “International,” “Global.” Choosing “International,” inference runs in Virginia, compliance disappears, and electricity prices become American. That dropdown is the switch from the first to the second mode.

But Coca-Cola’s global success relies on its formula being uncopyable. Large models are different—DeepSeek, Tongyi Qianwen are open source, weights files downloadable by anyone. The formula is public, but the bottling lines are others’, and the electricity is others’. The only remaining anchor for value capture is: speed of releasing the formula.

A report from Galaxy Securities states that the global large model iteration cycle has shrunk from half a year to a few months. Once chip bans slow down iteration, the window of freshness shortens further.

Arjun doesn’t care which country the formula comes from. He only cares about which is cheapest this week, and whether it will still be usable next week.

The overseas journey of a kilowatt-hour has been cut off between training and inference. Not just the cost chain, but also the value capture.

The first mode preserves the electricity advantage but loses the market—enterprise clients won’t accept data passing through China, and latency is a hard constraint. The second opens the market but loses the cost advantage—you’re no longer selling cheap electricity, but the formula itself.

Time is not on China’s side either. Deloitte predicts that the global AI compute focus is shifting from training to inference— inference will account for two-thirds of AI compute by 2023, possibly over 80% in the future. Training costs are a one-time expense; inference bills are daily. As inference becomes more important, the part of electricity prices can influence shrinks.

This is the fundamental difference between AI Token going overseas and photovoltaic going overseas. PV’s cost advantages from silicon material to modules to shipping are fully transmitted without breaks. AI Token’s cost chain cracks open between training and inference—if the power dividend can’t pass through, what truly crosses borders is the engineering capability embedded in that hundreds-of-GB formula.

Two modes, two dilemmas. How far can the route of selling AI Tokens cheaply with electricity go?

3. Cheap electricity, insurmountable walls

Not far. Three walls are narrowing this route.

First is the physical wall.

Arjun’s requests cross the Pacific, with 150 to 300 ms round-trip latency. It’s imperceptible in a single exchange. But an agent workflow involves dozens of continuous calls; machines don’t wait, and delays accumulate to seconds, causing the workflow to stall. It’s not a political issue, not a systemic issue—it’s limited by the speed of light.

Either stand in the light, or the light stands there.

Second is the institutional wall.

When US companies procure AI services, the technical leaders must answer five questions: Does the data go into China? Where are logs stored? Are inputs and outputs used for training? Is it compliant with local laws? Who is responsible if something goes wrong?

If they can’t answer these five questions, the procurement process stalls. It’s not that the model isn’t useful; it’s that compliance departments won’t sign off.

Morgan Stanley’s chief economist Xing Ziqiang cited a sharper precedent: Huawei’s 5G equipment also had technical and pricing advantages, but after 2018, it was still kicked out of Western telecom networks. 5G base stations, like AI Tokens, involve data passing through whose equipment and servers. His words:

“Don’t overhype the power advantage of AI Token going overseas, and ignore geopolitical and security considerations.”

Technical advantage can’t fix trust deficits.

Third is the political wall.

Chip bans block training; model review blocks deployment. The most unpredictable variables.

These three walls narrow the scope: relying on cheap electricity to sell AI Tokens overseas can only reach those developers who are insensitive to compliance, tolerant of latency, and highly sensitive to price. The formula export route, with its electricity advantage, can’t pass through.

Two modes, two dead ends. Where do Chinese AI companies stand in the global value chain?

4. Champions on the small stage, spectators in the value chain

On February 24, 2026, OpenRouter data revealed: the top ten models’ total AI Token consumption was 8.7 trillion, with Chinese models accounting for 5.3 trillion, or 61%.

“China surpasses the US for the first time”—a headline that spread across Chinese internet.

OpenRouter’s COO Chris Clark described what he saw in a podcast: “Chinese open-source models have an unusually high share in US enterprise agent workflows.” A complex coding task costs Claude $50 to $100; DeepSeek V3.2 about $0.50. A hundredfold difference— for startups running dozens of agents, it’s a matter of life and death.

But behind the 61%, there are two truths.

First: this is just a small stage.

OpenRouter’s statistics on AI Token consumption only account for about 3% of the global total. The big stage is elsewhere, with a huge disparity.

By April 2026, Anthropic’s annualized revenue surpassed $30 billion. Just 15 months earlier, it was only $1 billion—30 times growth. Claude Code, a coding tool, reached $1 billion in six months. Over a thousand companies paid over a million dollars annually, with 80% of revenue from enterprise clients. OpenAI’s annual revenue was $25 billion.

On China’s side: MiniMax burned $500 million in nine months, with revenue of $79 million. Caijing magazine was even colder—some Chinese models’ API gross margins might be negative, losing money on every request.

China produced the most AI Tokens, but is almost invisible in the value chain.

Second: not all AI Tokens are equally valuable.

The same AI Token, used for casual chat, is worth $0.01 per million; for coding, $200 per million; for legal review, $1,000 per million. A difference of a hundred thousand times. Industry estimates show that less than 5% of consumption creates over 80% of business value.

In the same week this article was written, Anthropic and Blackstone, Goldman Sachs established a joint venture worth $1.5 billion, deploying engineers directly into portfolio companies. They also launched 10 financial agents—pitchbook generation, KYC review, credit memos, month-end closing, financial audits. Jamie Dimon and Dario Amodei appeared together. A KYC agent consumes dozens of dollars of AI Tokens per run, saving thousands in compliance labor costs.

This is the real picture of high-value AI Tokens.

Goldman Sachs’s Marc Nachmann said: “Having a model alone won’t change your workflow. You need people who can integrate technology with actual business.”

This sentence sharply cuts through the US-China value gap. Chinese companies compete on who has cheaper AI Tokens; Anthropic competes on how to embed AI Tokens into Goldman’s every business line. The former sells raw materials; the latter sells solutions. Chinese models are tools in the global developer toolbox, not chief architects—they are pieceworkers.

It’s very much like China’s photovoltaic industry in 2008—leading in shipment volume worldwide, but the most profitable segment was the least profitable component. Pricing power, brand premium, high-end market share—all in others’ hands. That year, China’s PV industry still had over a decade before truly dominating globally.

But the 2008 PV story didn’t stop at “volume without profit.” The Chinese companies that eventually won globally weren’t just relying on cheap silicon; they controlled the entire stack from silicon to modules to power stations.

The hope for AI Token going overseas isn’t in electricity prices—it’s in whether it can become the infrastructure embedded in enterprise workflows.

Some are trying.

XunCe Technology is using AI Tokens in finance and energy verticals, deeply integrating billing with customer data and business processes. Token billing share rose from 5% to 20-30%, turning profitable in mid-2025. They’re not selling cheap AI Tokens but “re-running your business with AI Tokens.” The logic is similar to Anthropic’s deployment for Goldman—just in a Chinese vertical industry, with $1.5 billion and Jamie Dimon’s backing.

The gap is huge, but the direction is right.

5. Electricity remains unchanged, bills are changing

In May 2026, Doubao launched a paid version at 68 RMB/month. With 345 million monthly active users, it started charging.

In the same week, Tencent Cloud AI services increased prices by 5%, and Zhipu GLM-5’s price rose 50% over the previous generation.

Someone analyzed Doubao’s inference costs—58% hardware depreciation, 29% electricity. Each additional user means more GPU cluster consumption. ByteDance already has internal voices: “No clear path to commercialization; the inference cost per DAU puts pressure on profits.”

The days of dozens of companies losing money to attract users are over.

Price hikes reveal a more fundamental issue than the three walls: electricity advantage reduces AI Token production costs, but AI Token’s value has never depended on production costs; it depends on what it’s used for. A kilowatt-hour’s worth of AI Tokens, whether for casual chat or enterprise decision-making, differs by hundreds of thousands of times.

China has the cheapest electricity in the world, trained the cheapest AI Tokens, and holds 61% of the share on OpenRouter, but Anthropic’s annual revenue may surpass the total of all Chinese large model companies combined.

The water plants and data centers in Guizhou are still turning. Ten years ago, the same electricity powered mining farms, which dispersed. Now, the same electricity powers AI Token factories. The factories are still there, but the AI Tokens produced are increasing, while the captured value diminishes.

Arjun again saved dozens of dollars tonight. But every dollar he saved is precisely the part of value that Chinese AI companies failed to capture.