Can China win the AI race with cheap power?

Out in the waters near Shanghai’s Lin-Gang Special Area, with its abundance of wind resources, rows of white turbine blades turn slowly and rhythmically. They go on no matter the hour, feeding electricity directly down to an underwater data centre beneath the turbines.

This “offshore wind direct-linked” underwater data centre, powered by offshore wind and using seawater as a natural cooling source, was completed in October 2025 and recently went into operation. The project, which claims that 95% of its electricity comes from green sources, is led by Beijing Highlander Digital Technology Co Ltd, a firm with a history of supplying technology to China’s military. With an investment of 1.6 billion RMB (US$231.9 million), it is one of China’s cutting-edge pilots to expand its computing power footprint as well as explore ways to cut electricity costs.

Unlike traditional data centres, these artificial intelligence (AI) data centres mainly serve AI large model training, inference and other high-intensity data processing tasks. They are core infrastructure in the AI era, also known as computing power centres or intelligent computing centres. In China, this domain has been elevated to the level of a national strategic initiative.

In 2022, China launched the “Eastern Data, Western Computing” project, designed to steer tech firms in the east to shift their computing power needs to the energy-rich, lower-cost western regions, while building multiple computing clusters in central China to form a nationwide computing power network.

This initiative has accelerated the pace for the setting up of computing power centres across the country. According to tender data compiled by International Data Corporation (IDC), between 2022 and 2024, there were 219 projects related to urban intelligent computing centres, with 243 such projects launched in the first half of 2025 alone.

In an interview with Lianhe Zaobao, Charles Yonts, head of Asian Sustainability Research at Macquarie Capital, said that China ranks second only to the US globally in the development of computing power centres. “Looking out over the next five years, we see China growing at a similar rate to the US, albeit from a smaller base.”

An article on the AI competitive landscape published on the US Federal Reserve’s website in June 2025 estimated that the US holds around 74% of global high-end AI computing power, while China accounts for about 15%. But industry insiders believed China’s overwhelming advantage in electricity supply would be a key pillar supporting its computing centre development.

Generating roughly 5,000 words with a large AI model consumes about one kilowatt-hour (kWh) of electricity, while producing images or video requires even more. As a saying from the tech world goes: the ultimate limit of AI is computing power, and the ultimate limit of computing power is electricity.

Yonts noted that the US is struggling to absorb all of the incremental power demand stemming from data centre build-out, while “this is almost a rounding error in China”.

Yonts added: “While the US has been adding an average of 20 gigawatts of new power capacity each year for the past decade, most of that was replacement of retired coal plants with gas and renewables rather than added capacity, so actual generation has been stagnant over that period, at just 0.6% average annual growth.”

The strain that data centres place on electricity growth is very different in China and the US. In China, demand from data centres (including computing power centres) accounts for around 15% of the annual increase in power demand growth, whereas in the US, related demand is equivalent to roughly 400% of the country’s average annual growth in power capacity in recent years.

Electricity is also cheaper in China. In Xinjiang, where wind resources are abundant, power prices can be kept to around 0.3 RMB per kWh, while US Energy Information Administration data show that commercial electricity prices in the US are typically at about US$0.14 per kWh.

Shen Hong, director of the Domestic Development Cooperation Department at Peking University’s Institute of New Structural Economics, said when interviewed that China’s planning at the national level is intended to further improve the efficiency of energy allocation.

He said tasks requiring real-time responsiveness, such as certain computation and inference workloads, are better handled by computing centres in the eastern region, while the west, with cheaper electricity, can handle model training, game rendering and other latency-tolerant work.

Shen also felt that China’s promotion of “power from the west to the east”, along with the construction of hydropower and nuclear plants and the relatively new state of its power infrastructure, gives it a comparative advantage in its power system.

However, besides electricity, the ceiling on a computing centre’s capabilities is still determined by the chips themselves.

David Dong, partner at FutureX Capital, said that as long as the manufacturing process for Chinese-made chips cannot advance further, it would be hard to increase computing density, and data transfer speeds between chips would also become a bottleneck.

“Take communication speed for instance; chips produced with a 7-nanometre process can generally only support 112 Gb/s data interfaces, whereas 224 Gb/s has already become the mainstream standard in the US”.

In his view, constrained by semiconductor technology bottlenecks, the development of China’s data centre computing and communication capabilities would both slow markedly, leaving it “roughly two generations behind on the whole”.

It is understood that tech giants such as ByteDance and Alibaba still rely mainly on high-end Nvidia chips stockpiled before US export bans to train their large models, while new projects spearheaded by local governments have been instructed to use domestically-produced chips from firms such as Huawei as their mainstay.

Yonts said that without access to the most advanced chips, Chinese designers are linking larger numbers of less efficient chips into nodes that can rival the power of Nvidia clusters — Huawei’s CloudMatrix clusters are a leading example.

He pointed out that this approach is less efficient, but as power is not an issue for China, it is a very promising approach that leverages China’s strengths.

As China accelerates the construction of computing power centres, industry figures say the configuration of computing resources often fails to match corporate needs in practice, leaving firms reluctant to rent capacity and contributing to the idling of many data centres.

In a speech in April 2025, former China Telecommunications chief engineer Wei Leping said that intelligent computing centres were “blooming everywhere” in China, but GPU (graphics processing unit) utilisation was highly uneven. “Some are overstretched, while some are underused, with average utilisation under 30%.”

Wei’s remarks highlighted the current mismatch between supply side performance of China’s computing power and actual demand.

Industry insiders noted that many local data centres were still equipped mainly with CPUs (central processing units) geared towards traditional IT and cloud service scenarios, and are ill‑suited to the needs of AI model training and inference. In some places, the chip mix and computing architecture are not ideal, resulting in very narrow application scenarios; even with low power costs, the resources cannot be effectively used.

Commenting on such cases, Dong was blunt: “The fundamental reason for some of these mismatches is that the people building data centres are not the ones who actually use computing power.” He also pointed out that technical standards for data centres vary across regions, making cross‑regional management difficult.

China’s current data centre boom is driven by AI demand and by local governments promoting computing power as a new form of digital infrastructure as traditional infrastructure spending slows. Many projects, beyond those by the three state telecom operators and major tech firms, are led by local governments.

Professor Wen Yonggang, President’s Chair at the College of Computing & Data Science, Nanyang Technological University, said these projects involve land, power and communications, require large investments and long industrial chains, and significantly boost the digital economy. This, he noted, explains why local governments make them a priority.

Compared with the US, where large tech companies mainly invest in and build such facilities based on their own and market needs, Wen felt that China’s top-down planning offers more advantages in concentrating resources and pushing projects forward efficiently. But he cautions that if highly specialised computing power infrastructure is not led by the end-users of that computing power, or if operators are primarily driven by investment and returns, the risk of mismatches between capacity built and capacity used could be easily amplified.

Wen added that similar problems appeared in the early days of cloud computing, when China built a large number of cloud centres across the country — yet in some regions, their effective utilisation rate once fell to below 15%. The development of computing power centres must likewise guard against an over-emphasis on construction at the expense of application, which leads to wasted resources.

This article was first published in Lianhe Zaobao as “AI需求驱动中国全力加速算力建设”.

Related: The true price of AI: Water and energy demands behind Asia’s data centre boom | Has China solved AI governance?