From seabed to Gobi: how is China powering the AI age?

Behind it is a years-long push by Chinese planners to treat power grids and computing infrastructure not as separate problems, but as a single system to be designed together.

Hi from Beijing. Artificial intelligence (AI) is consuming electricity faster than the world can generate it. China thinks it has found part of the answer -- 10 meters below the East China Sea. This week’s story starts with a speedboat.

From central Shanghai, drive southeast, cross the Donghai Bridge, and board a speedboat at Yangshan Port. Forty minutes out, a steel platform rises more than 20 meters above the East China Sea. An elevator descends 10 meters below the waterline, where cylindrical steel vaults stand upright on the seabed, thousands of servers humming in the dark -- processing the queries, the transactions, the AI requests of daily life.

Here is the Lingang underwater data center.

It entered full commercial operation in May, with data modules weighing 1,950 tonnes -- roughly the equivalent of 1,300 passenger cars. Electricity runs directly from the turbines of an offshore wind farm to the servers below, without touching the onshore grid.

It is the first facility in the world powered directly by offshore wind in this way. Behind it is a years-long push by Chinese planners to treat power grids and computing infrastructure not as separate problems, but as a single system to be designed together.

FEEDING DATA CENTER’S HUNGER

The problem is not unique to China.

Global data centers consumed an estimated 485 terawatt-hours of electricity in 2025 -- roughly equivalent to France’s entire annual consumption. The figure is expected to double by 2030, accounting for around 3 percent of global electricity demand, according to the International Energy Agency (IEA).

To solve the energy challenges at hand, tech giants are searching for new alternatives with sky-high investments. According to the IEA, the world’s largest technology firms are driving a surge in data center investment, with their combined capital expenditure set to climb another 75 percent in 2026.

Much of that investment is flowing toward nuclear power. Constellation Energy is investing about 1.6 billion U.S. dollars to restart part of the Three Mile Island nuclear plant in Pennsylvania. Under a 20-year power purchase agreement signed with Microsoft in September 2024, the unit will supply electricity to the tech company’s AI data centers, with the reactor now targeted to come online as early as 2027.Amazon expanded its nuclear partnership with Talen Energy in June 2025, signing a 17-year power purchase agreement for up to 1.92 gigawatts from Pennsylvania’s Susquehanna plant.

In January 2026, Meta announced agreements with three nuclear power providers to secure up to 6.6 gigawatts of energy over the next 20 years, with the capacity expected to come online progressively by 2035.

Nuclear offers reliable, weather-independent power -- what wind and solar cannot provide without massive storage, while delivery is years away, as most contracted capacity will not reach the grid until the 2030s.

In Europe, Norway runs its data centers on hydropower, and Iceland on a mix of hydropower and geothermal energy, with cold climates providing free cooling for most of the year. But the clean-power capacity of these countries is far too small in scale to meet global AI demand.

SEA AS A SOLUTION

China’s approach starts from a different premise: building where the clean energy already is, and coordinating energy supply and computing capacity.

Lingang’s offshore data center project offers the latest example of this idea.

“Sinking the data center into the sea is not a gimmick,” said Chen Xi, general manager of the research and development center of HiCloud, the project’s developer. “It is the inevitable choice to break through the constraints on AI development in eastern China.”

China’s eastern coastal provinces, the main arena of its digital economy, face a twin constraint on data-center expansion. Dense populations and concentrated industry have left coastal land scarce and expensive, with little room for the large physical footprint of conventional facilities. At the same time, the region’s abundant offshore wind capacity remains underused, hampered by limits on how much the grid can absorb, leaving much of that clean power’s value untapped.

“The most intuitive advantage is land savings,” said Chen. At equivalent computing capacity, the subsea data center occupies only about one-tenth the footprint of a conventional onshore facility.

For the underwater facility, the bigger prize is the offshore wind it can soak up. As China’s renewable capacity has been added at scale, the grid has struggled to absorb it all, putting growing pressure on system balancing. Since 2025, the curtailment rate for the country’s wind and solar power has risen above 5 percent.

Lingang tackles this at the source. About 500 meters away, a 200-megawatt offshore wind farm of more than 50 turbines generates over 500 million kilowatt-hours a year, feeding the undersea data center directly through a fiber-optic composite cable.

“Conventional onshore data centers draw power over long-distance lines, which brings heavy losses and high costs,” Chen said. “By siting wind generation and computing capacity close together, we supply more than 95 percent of our electricity directly from green power.”

The model both eases wind curtailment and cuts energy use. While onshore data centers typically record a Power Usage Effectiveness ratio of between 1.4 and 1.6 -- meaning they consume 40 to 60 percent more electricity than their computing equipment alone requires, mostly for cooling -- Lingang’s is below 1.15. At full scale, it is projected to save 61 million kilowatt-hours a year, according to Chen.

The deep water does the rest. At this depth, the seabed maintains an average temperature of 15 degrees Celsius year-round, offering a natural cooling system. A conventional onshore facility of the same scale would consume around 40,000 tonnes of fresh water a year -- enough for a single household for roughly a century.

PICKING UP WHERE MICROSOFT LEFT OFF

The idea of an underwater data center itself is not new. Microsoft launched Project Natick in 2015, sinking a capsule off the coast of California, but shut the experiment down in 2024 without moving to commercial deployment.

The firm cited the concept’s limitations for modern cloud and AI demands and its difficulties in commercialization, as sealed underwater pods make it difficult to upgrade hardware or expand capacity over time.

Analysts say the same challenges that ended Project Natick have not disappeared. “Replacing, upgrading or repairing servers requires specialized marine operations,” said Roy Chua, founder of AvidThink, a technology research firm.

For the Lingang project, the modules are sealed against corrosion using marine-grade anti-fouling coatings, and ballasted with internal water-filled structural tubes to resist typhoon conditions. Failure rates so far are one-eighth of those in conventional facilities, and the design includes redundant backups and external components that are replaceable underwater, according to HiCloud.

HiCloud is already planning the next steps. The company is developing single-module capacity of five to seven megawatts, with clusters planned across the Yangtze River Delta, replicating the offshore wind and underwater computing model along China’s eastern seaboard.

Beyond wind, the company envisions pairing future underwater facilities with wave energy, tidal power, and offshore solar, the firm’s general manager, Su Yang, said.

PAIRING POWER WITH COMPUTING

As China accelerates the push to shift energy-intensive data processing to resource-rich regions, the same logic plays out 2,000 kilometers inland in Zhongwei, a city on the edge of the Tengger Desert in northwest China’s Ningxia Hui Autonomous Region.

Earlier in May, a large-scale local project to supply renewable energy directly to data centers started operating. The project has a solar capacity of 500 megawatts -- with a further 1.5 gigawatts of wind power planned by the end of the year, said Jin Liang, deputy general manager of Datang Zhongwei New Energy Co. Ltd.

At full capacity, the wind and solar farm will generate 4.3 billion kilowatt-hours annually, enough to power the adjacent cloud cluster’s 2.29 billion kilowatt-hour annual demand.

Zhongwei is one node in a larger national pattern. In 2026, coordinating computing infrastructure with energy supply was written into China’s government work report. More regions, including Inner Mongolia, Ningxia and Guizhou, are building integrated projects on the same principle: generate the power where the wind and sun are strongest, consume it on-site, and plan the data centers around the energy rather than the other way around.

Building data centers is only the first step.

Wei Yiming, a researcher on carbon neutrality at the Beijing Institute of Technology, said that while computing infrastructure is expanding quickly, the power systems needed to support its stable operation and green energy consumption have yet to be fully developed.

The shortage of versatile, high-end talent -- people who grasp computing, power grids, energy management, and market rules together -- is another bottleneck holding back the large-scale rollout of coordinated power-computing operations, he said, adding that more needs to be done to close the gap.