Power Tokenization Goes Global: China's Electricity Computing Power Set to Tackle the World

Diagram/IC

Insufficient computing power, “Lobster” (OpenClaw) is not popular.

To “raise Lobster,” you need to feed it tokens—computing power; feeding tokens (subwords) consumes electricity.

How much electricity do tokens really consume?

Long Jiang Securities states that electricity costs account for 60%-70% of the operating costs of AI large models, so tokens can be seen as a kind of “electricity derivative.” Assuming that future domestic large models’ annual call volume expands to 1×10⁵ trillion tokens, corresponding to an annual electricity consumption of 87.5 billion kWh, about 0.84% of the total societal electricity use in 2025.

According to Huatai Securities, the global daily token usage at the scale of trillions could have a 10% elasticity on China’s electricity and power supply, significantly influencing green certificate prices, capacity electricity prices, and even electricity prices.

These data outline a highly imaginative new overseas market—“Electricity Tokenization for Global Expansion,” making China’s electricity a global topic.

Influenced by the “Lobster” craze, the power sector has attracted market attention. On March 11, Green Power Generation and Huadian Energy both hit two consecutive limit-ups, and JinkoSolar’s stock surged 16.6% over two days.

Electricity hasn’t gone cross-border yet, but via tokens, it can go overseas

To understand this electricity export, we first need to analyze the physical properties of tokens.

Beneath the digital surface, tokens are not just characters generated out of thin air. They are “digital fuel” condensed after high-performance GPUs consume large amounts of electricity and perform hundreds of billions of calculations.

For a long time, electricity has been the most difficult commodity to trade across borders—UHV transmission can only send power to neighboring countries, liquefied natural gas requires expensive receiving stations, and battery storage costs more than the electricity itself. But the emergence of tokens gives electricity its first lightweight global carrier: electricity stays domestically, while value flows overseas via tokens.

This allows tokens to become a cross-border settlement medium for electricity value.

The world’s largest AI model API aggregation platform, OpenRouter, recently showed that Chinese models are sweeping overseas markets with a “large volume, high capacity” approach.

Data from OpenRouter indicates that from February 9 to 15, Chinese models with 4.12 trillion token calls first surpassed the US models’ 2.94 trillion tokens in the same period; from February 16 to 22, Chinese models’ weekly call volume further surged to 5.16 trillion tokens, while US models’ calls dropped to 2.7 trillion. Among the top five models by platform calls, four are from Chinese companies: MiniMax’s M2.5, KimiK2.5 from Moonlit, GLM5 from Zhipu, and V3.2 from DeepSeek. These four models account for 85.7% of the top five total calls. Considering that most OpenRouter users are overseas developers—47.17% from the US and only 6.01% from China—this more objectively reflects the true global appeal of Chinese AI models.

Foreign users calling Chinese APIs, though electricity isn’t crossing borders, the value is delivered internationally through tokens.

Why can our tokens go global?

This “cross-border trade” system is founded on a core advantage—cost-effectiveness.

Long Jiang Securities’ research shows that, in terms of input prices, MiniMax M2.5 and Zhipu GLM-5 both cost $0.3 per million tokens, while Anthropic’s Claude Opus 4.6 costs $5, which is 16.7 times more expensive.

When developers can get the same or better intelligent responses at less than one-tenth of the cost, the market will naturally choose accordingly.

But “cheapness” itself is a result, not the cause. The real question is: why are our tokens so cheap?

The market might think China’s large model tokens’ competitive edge in going overseas is due to electricity prices, but Huatai Securities’ calculations show that electricity costs currently account for only about 10%. As chip inference efficiency and economics improve, the proportion of electricity costs per token could continue to rise.

Shi Yuxia, senior engineer at the Policy and Economic Research Institute of China Academy of Information and Communications Technology, and a distinguished expert at the Strategic Research Center for Information and Electronic Engineering Development of the Chinese Academy of Engineering, offers a more comprehensive explanation: “Electricity prices are not the core factor enabling China’s tokens to outperform abroad in costs. Our ability to go overseas with tokens results from the combined effects of advanced AI model technology, energy cost advantages, and supply chain strengths.”

Specifically, our AI large model companies’ innovative architecture enhances technical capabilities, reducing the computing power needed per token; energy cost advantages lower electricity expenses per unit of compute; supply chain advantages spread out infrastructure investments—these three layers of benefits reinforce each other, culminating in an exceptional cost-performance ratio for tokens.

When Chinese models leverage the combined advantages of “technology + energy + supply chain” to “go global,” could “electricity tokenization for overseas expansion” become a new industry direction?

Shi Yuxia states that, for now, it is not a standalone industry sector. Instead, it represents an advancement and upgrade of existing industry chains toward higher value.

In other words, token export isn’t creating a new track out of thin air but is elevating China’s existing computing power, power, and AI industries within the global value chain.

However, this doesn’t mean there are no new possibilities. Shi Yuxia adds, “In the future, some fields may give rise to new models and formats of tokenization for overseas expansion.”

The end of computing power is electricity—where do we win?

The limit of computing power is electricity—and we have won two battles on this chain.

First, we win by “saving electricity.”

Most Chinese models adopt a mixture of experts (MoE) architecture, avoiding “full deployment.” A model with hundreds of billions of parameters only activates a small subset of “expert networks” for simple questions. This “on-demand activation” is itself a form of fine-tuned electricity scheduling.

Second, we win by “affordable use.”

More critically, the electricity powering these GPUs is cheap.

By the end of 2025, China’s total installed power capacity will reach 3.89 billion kW, with total electricity consumption surpassing 10 trillion kWh, both ranking first globally. The large scale dilutes the unit cost.

Currently, industrial electricity prices in China have been stable at around 0.6 RMB per kWh, making it one of the lowest-cost regions worldwide. When this electricity cost advantage reflects in the operational costs of large models, Chinese large models naturally gain a “power price premium.”

Even more importantly, the electricity is not only cheap but also stable.

China boasts the world’s most powerful grid system, with UHV transmission technology leading globally, enabling “west-to-east” and “north-to-south” energy transfer: wind power from Xinjiang can be directly sent to Shanghai, Sichuan hydropower can supply Beijing, and cross-regional energy dispatching capabilities far surpass those of other countries.

Three challenges amid the boom

The “raising lobsters” craze has rapidly propagated through the “compute-electricity synergy” industry chain. On March 9, ShaoNeng shares hit two consecutive limit-ups; Yinxing Energy and GCL New Energy hit the limit and set new highs.

On the other hand, opportunities come with challenges.

Shi Yuxia summarizes three pressures behind the token export boom: First, the explosive growth in computing power demand raises higher requirements for compute-electricity coordination. Second, intensified industry competition is squeezing profit margins. The large model sector itself is a red ocean, with ongoing price wars. “This compresses corporate profits, affecting their R&D investments.” Third, high-end computing power remains limited by technological levels, as China’s advanced process technology still faces bottlenecks.

More fundamentally, brand recognition needs to advance.

Shi Yuxia emphasizes: “Chinese companies’ brand recognition still needs to be elevated to higher-end levels. We must move beyond the perception that cost advantage is the only strength in the international market.”

Cost-effectiveness is just a stepping stone. Moving from “high cost-performance” to “high-end trust” is a crucial hurdle for China’s AI industry to go global.

Electricity tokenization for overseas expansion also addresses another urgent issue—renewable energy absorption.

China leads the world in wind and solar installations, but “curtailment” has long been a hidden pain—during peak generation, excess power can’t be used or transmitted, leading to wasted green energy.

It’s not due to subsidies or profit sacrifices—it’s because we’ve turned “unused electricity” into “affordable tokens.”

Researcher Tao Ye from Beijing New Energy Zero Carbon Institute

Editor Wang Jinyu

Proofreader Mu Xiangtong