InDepth:BeneaththeSea,ChinaTeststheFutureofAIPower

Ten kilometers off the coast of Shanghai, a 32-meter (105-foot) steel colossus rises from the waters of the East China Sea.

Beneath the waves, adjacent to more than 50 offshore wind turbines, a submerged cylindrical cabin houses a high-density computing hub.

Built by Shenzhen HiCloud Data Center Technology Co. Ltd., this facility is purported to be the world’s first underwater data center powered directly by offshore wind. Naturally cooled by the ocean and consuming almost no freshwater, its first phase launched commercially in February with an ultra-efficient power usage effectiveness (PUE) rating of 1.15.

The futuristic facility is more than an engineering marvel; it is a testing ground for China’s accelerating rollout of its “computing-power synergy” strategy. As the global boom in artificial intelligence (AI) drives explosive demand for electricity, Chinese local governments and tech giants are scrambling to find stable, green and cost-effective energy solutions.

The mandate comes straight from the top. In March, China’s cabinet officially enshrined the concept of computing-power synergy in its annual government work report, calling for ultra-large AI clusters to be developed in lockstep with power infrastructure to forge a new “intelligent economy.”

Nowhere is this more evident than in Shanghai’s Lingang Special Area, a national economic zone driving the city’s tech ambitions. Power consumption by the local grid surged nearly 30% in the first five months of 2026, driven overwhelmingly by data centers.

By the end of the year, Lingang’s intelligent computing capacity is expected to double from last year’s 40 quintillion floating-point operations per second.

To cope, local power grids have rapidly doubled the number of high-voltage substations over the past year, while the government dishes out millions of yuan in subsidies to computing projects.

Underwater servers

Operating a data center underwater capitalizes on natural cooling to slash energy use while saving valuable land. A person familiar with the HiCloud project said that the facility serves as a supplement to onshore computing, providing clients like China Telecom Corp. Ltd. with millisecond-level low latency.

The current design — a surface platform for power equipment and a submerged 1,950-ton cabin for servers — overcomes a major hurdle from an earlier 2020 pilot in Hainan province. That fully submerged model required costly engineering ships for routine maintenance, ballooning operational expenses.

The Lingang redesign drastically cuts these maintenance costs and ensures top-tier reliability by drawing power from three distinct sources: the adjacent offshore wind farm of Shenergy Group Co. Ltd., the onshore grid, and backup diesel generators.

Yet regulatory bottlenecks remain. Despite being physically wired to the offshore wind farm, HiCloud’s underwater data center still purchases its green electricity through the state grid at standard commercial rates of about 0.75 yuan ($0.11) per kilowatt-hour, more than double the direct cost of the wind farm’s power supply.

Industry insiders said implementing true point-to-point direct-purchase policies could drastically cut costs for the project’s second phase, which is set to break ground later this year.

HiCloud’s underwater data center (UDC), which draws power from a neighboring offshore wind farm, got up and running in February off the coast of Shanghai in the East China Sea. Photo: HiCloud

Sending data west

While some servers are running underwater, others are having their workloads virtually transported across the country. In late May, China Telecom dispatched computing tasks from Lingang to a green computing center over 4,000 kilometers (2,485 miles) away in Karamay in the Xinjiang Uyghur autonomous region.

The test, part of a national strategy dubbed “East Data, West Computing,” aimed to relieve strain on the eastern power grid during peak hours. By shifting demand to the west — where land and renewable energy are abundant — the Shanghai facility dropped its local power load by 75%. Crucially, the system seamlessly transferred tasks across heterogeneous hardware, shifting workloads from Huawei Technology Co. Ltd. processors in the east to Nvidia Corp. chips in the west.

China Telecom’s infrastructure and power chief in Lingang, Yuan Xiaoyang, said that similar tests had successfully shifted loads to Hubei and Fujian provinces in 2025. However, Yuan acknowledged that these cross-country dispatches remain largely experimental.

Making them routine will require a market-driven pricing mechanism that offers clients cheaper rates in exchange for allowing their computing tasks to be relocated. Currently, the structural shortage of high-end computing chips in China and the vastly varying energy demands of different AI models make widespread, automated implementation a challenge.

AI energy storage

Back on land, data centers are also deploying large-scale battery storage and AI models to shave costs.

SenseTime Group Inc. has integrated an 18-megawatt energy storage station, backed by an investment fund affiliated to battery giant Contemporary Amperex Technology Co. Ltd., into its Lingang computing hub. Other local players, such as cloud computing service provider Shanghai Yovole Networks Inc., have similarly deployed Tesla Inc.’s Megapack systems.

Shanghai experiences steep peak-to-valley electricity price gaps, with peak rates nearly 4.6 times higher than off-peak hours. SenseTime’s facility uses a proprietary AI energy model to predict its own computing loads. When the system anticipates a surge in processing demand, it preemptively releases battery-stored electricity or adjusts ambient cooling systems. This ensures the facility’s power draw from the state grid stays just below the threshold that triggers expensive peak-capacity surcharges, which can run around 38 yuan per kilowatt.

Head of SenseTime’s Lingang operations, Zhang Xu, said that this hardware-software synergy has shaved annualized electricity costs by 7%, dropping the rate below the regional industry average.

As Shanghai transitions to a continuous spot electricity market with prices fluctuating every 15 minutes, Zhang anticipates this AI-driven approach will become an essential tool.

The ultimate goal, he said, is to precisely calculate the energy consumed and the exact cost required to produce every single AI-generated token, setting a new benchmark for production efficiency in the AI era.

Contact editor Han Wei (weihan@caixin.com)