_________

Published on Mar. 30, 2026

A software engineer in San Francisco sends a prompt to a Chinese AI model. The data races through a fibre-optic cable under the Pacific to a data centre in western China and returns with the answer in less than two seconds.

Every token, the basic unit of computation in large language models (LLMs), consumes energy. Electricity costs typically account for between 46 and 60 per cent of the operating costs of AI data centres, according to the International Data Corporation (IDC).

In this cross-border use of LLMs, the energy generated in China never physically leaves the country. Yet its value has become exportable without pipelines or high-voltage transmission lines. "Tokens are the new oil of the AI era," wrote the Chinese business portal Shidai Caijing (in Chinese).

The sums involved are no longer trivial. MarketsandMarkets has estimated the total global inference market at US$106 billion for 2025. Nobody has yet precisely measured China's share of token exports within that figure, but it is certainly already worth several billion dollars and growing explosively from what is currently a relatively small base.

As Chinese chatbots and LLMs gain popularity, an ever-larger share of these revenue flows is crossing China's borders. In mid-February “Chinese AI models had for the first time overtaken those from the United States in token usage", the Chinese tech portal Huxiu reported.

In the week of 9 to 15 February, Chinese models recorded 4.12 trillion token calls, compared with 2.94 trillion for American AI models. In the weeks that followed, the gap widened further.

At that point, four of the five most-used models worldwide on OpenRouter were Chinese, including MiniMax M2.5, Kimi K2.5 from Moonshot AI, Zhipu GLM-5 and DeepSeek V3.2.

OpenRouter does not capture tokens that enterprises route directly to AI providers via APIs. But the figures reveal a clear trend.

"Every time a developer anywhere in the world calls a Chinese model, they are effectively purchasing a token service produced by China's electricity, computing infrastructure and algorithmic efficiency," writes Shidai Caijing.

That quote neatly captures all three of the main reasons why Chinese tokens are cheaper than American or other alternatives. Energy costs for data centres are lower in China. The infrastructure, from Chinese-manufactured GPUs to server cooling systems, is cheaper thanks to fully domestic supply chains. And models such as DeepSeek and Kimi are known for their computational efficiency, which saves electricity.

For end users, the computing costs in China's data centres matter because they are significantly lower than those abroad, and Chinese AI companies can deploy this as a competitive advantage.

In February, processing one million input tokens cost roughly US$0.30 at MiniMax M2.5, according to the company. DeepSeek V3 was cheaper still at US$0.14. Anthropic's Claude Opus 4.6 was listed at US$5.00 per million input tokens, according to the company's published rates, roughly 17 times the price of MiniMax and 35 times that of DeepSeek.

Prices change constantly, so this can only be a snapshot. But the price differential is significant, and it is worth examining the three structural pillars behind it in some detail.
Model efficiency stems from the fact that nearly all Chinese LLM developers use Mixture-of-Experts (MoE) architectures. Rather than activating the entire neural network for every query, MoE models activate only the specific processing pathways required for a given task.

DeepSeek V3 uses this approach and achieves inference costs roughly 30 times lower than those of GPT-4o. MiniMax M2.5 has 229 billion parameters in total but activates only 10 billion at any given time. The result is drastically lower computing costs per token.

The second pillar of the new token economy is China's cheap energy. China is the world's largest electricity producer. By the end of 2025, the country's total installed generation capacity had reached 3.89 billion kilowatts, slightly more than three times the equivalent capacity in the United States.

Industrial electricity prices in China vary by region but can be as low as 0.15 to 0.28 yuan per kilowatt-hour for green power from wind and solar sources in the country's western provinces (roughly US$0.02 to US$0.04). In Europe, industrial electricity prices are typically at least five times higher than China's green power rates; in the United States, they are two to three times higher.

Electricity price data varies widely depending on whether industrial or other rates are being compared, which discounts are factored in and which time period is under discussion. What matters here is the order of magnitude, not a precise dissection of energy costs.

The price gap between Chinese and non-Chinese electricity that remains after all possible caveats is of considerable significance in the international token economy.

"The export of tokens has quietly become China's most efficient high-value-added energy trade. While raw electricity from China sells for roughly 0.5 yuan (US$0.07) per kilowatt-hour when exported, converting the same electricity into AI computing power and selling it as tokens yields an estimated 22-fold increase in value," wrote the China Daily, citing an industry study.

While there is no ultra-high-voltage transmission line between China and, say, India, Chinese electricity can now nonetheless be "consumed abroad" thanks to the global AI boom, the paper wrote.

China's government has recognised the direct link between cheap energy supply and the country's economic competitiveness in the AI era. In the 2026 government work report, "electricity-computing synergy" was elevated to a national priority for the first time.

To strengthen the synergies between China's already massive expansion of wind, solar and other renewable energy and the scaling of AI across the country's industry, an increasing share of green power is now being channelled directly to AI data centres.
New data centres are being built next to solar and wind farms in the sun- and wind-rich desert provinces of western China. The green power produced at these sites is exceptionally cheap and, in these projects, is not fed into the grid at all. Instead, it powers the AI servers directly on site.

In addition, growing volumes of wind and solar power from Inner Mongolia and the provinces of Gansu and Ningxia are being transmitted via ultra-high-voltage lines to other parts of the country where data centre demand exceeds that of the deserts and semi-arid regions in the west.

China's government calls this programme "East Data, West Compute". By February 2026, 84 direct green power connection projects had been approved nationwide, with a combined renewable capacity of 32.59 million kilowatts, according to data from the National Energy Administration in Beijing.

The new clusters in the west of the country, where data centres and transformers for UHV lines stand alongside vast photovoltaic arrays, are now referred to in Chinese trade media as the "Western Data Valley", a nod to Silicon Valley.

New data centre nodes designated as "national" hubs must source at least 80 per cent of their electricity from renewable sources, under a regulation issued by the central government in July 2024 entitled "Green and Low-Carbon Development of Data Centres". The logic behind it is to create a flywheel effect.

It works like this. Growing token demand drives the construction of more data centres, which absorb more renewable energy, which lowers electricity costs, which makes Chinese tokens cheaper, which in turn enables faster scaling of AI across industry and society.

As a side effect, this also makes Chinese tokens attractive internationally. From the perspective of China's central planners, however, that is secondary.

It is possible that politicians in Washington and Brussels will soon discover the token economy as a target for tariffs or other trade barriers, with the ostensible aim of shielding their domestic AI providers from Chinese competition. That could reduce token exports in the future, but it would not diminish the cost advantage of Chinese tokens.

Beijing will, this much can be predicted, continue building this new ecosystem of green energy and AI data factories. Even if legitimate or politically instrumentalised concerns about cyber and data security, along with geopolitical tensions, were to constrain the nascent token export economy, China's industrial policy will continue to maximise the synergies described above.

In that scenario, the productivity gains from cheap, green-powered AI computing would in future be available only to the Chinese economy, no longer to the United States or Europe. From Beijing's perspective, that would be a tolerable outcome.

The third pillar of China's "tokenomics" is the growing autonomy of the entire Chinese supply chain for data centres and the infrastructure they require. From transformers and liquid cooling systems to servers and AI chips, virtually everything can now be built in China. The order books of transformer manufacturer Baobian Electric, for example, are already full through to 2027.

Where US chip export controls bite, denying China access to the most powerful Nvidia GPUs, bottlenecks remain. Computing capacity cannot be scaled up quite as fast as the Chinese government and the country's companies would like.

Yet even here, Chinese semiconductor manufacturers such as Huawei, Cambricon and Hygon are steadily closing the gap. Their chips are not quite as powerful as Nvidia’s, but more of them can simply be clustered together to achieve comparable computing performance. Space for larger data centres is something China has in abundance.

In return, chips and other high-tech components labelled "made in China" tend to be cheaper than comparable imports, owing in part to lower energy costs and the economies of scale available in Chinese manufacturing. The amortisation of GPU costs is, in turn, another important factor that determines token prices.

"Electricity prices alone are not the decisive factor that allows Chinese tokens to undercut foreign competitors on cost," Shi Yuxia, an expert at the China Academy of Information and Communications Technology, was quoted as saying in Chinese media. "It is the interplay of energy cost advantages, improved AI competence and supply chain dominance."

Each of these three stars in the firmament of China's AI infrastructure shines brightly enough on its own. Together, however, they form a constellation that is difficult to replicate.