China sank a $226M data center under the sea. AI's energy problem didn't get smaller.

China's first commercial underwater data center sits 35 meters below the surface in the Lingang Special Area off Shanghai, a sealed cylinder of servers drawing power from an onshore wind farm and shedding heat into the surrounding seawater. The mechanism is straightforward. The scale, against the problem the facility is meant to address, is the actual question.

Operator HiCloud reports the facility cost about $226 million to build, runs at an initial 24 megawatts, and clears a PUE (power usage effectiveness, the ratio of total facility energy to the energy reaching the servers) below 1.15, where typical land-based data centers hover near 1.5. Chinese state media say the design reduces power consumption "close to 23%" and cuts water and land use by 100% and more than 90%, respectively, compared with conventional builds. The cooling works by passive heat exchange: seawater carries heat away from the servers without the chillers, cooling towers, and evaporation that land facilities rely on. Those figures originated with HiCloud and Chinese government releases; they have not been independently measured at load.

The macro problem this facility is meant to address is the data-center industry's place in the global energy stack. A United Nations University report estimated that data centers consumed roughly 448 terawatt-hours of electricity in 2025, ranking them 11th among countries, behind France and ahead of Saudi Arabia. The same research projects AI's water demand could equal the household needs of 1.3 billion people by 2030. The global AI market, by the same UN University's June estimate, is set to grow from $189 billion in 2023 to about $5 trillion by 2033.

At 24 megawatts, the underwater facility is one tile in that demand curve. The Shanghai facility, by Data Center Dynamics' reporting, is the first commercial-scale deployment of a concept that has, until now, existed mostly as engineering exercises. A 24 MW plant serves a few thousand households' worth of compute; the global data-center fleet is being asked to absorb the training and inference workloads of a technology that has crossed the threshold of mass consumer adoption.

Eric Masanet, who studies data-center energy use at the UC Santa Barbara Bren School, told WIRED that the empirical evidence is too thin to declare underwater data centers a sustainability solution. The Chinese figures depend on assumptions about load profile, ambient seawater temperature, and the energy cost of the cooling-loop pumps that move heat from the servers to the surrounding water. None of those have been audited by a body outside Chinese state channels.

The closest analogue is Microsoft's Project Natick, a 117-foot pod off Scotland that ran from 2018 to 2020. Microsoft calls it a feasibility study and says it has no commercial underwater data centers in the water. A prior HiCloud commercial module, deployed off Lingshui in Hainan in late 2022, has run for more than two years; the company's project general manager told People's Daily it delivers a 40–60% energy-efficiency gain over comparable land-based servers, a wider range than the Shanghai claim.

What would actually resolve whether submersion is a real option? Third-party measurement at sustained load rather than at launch-day peak, when servers often run well below capacity. Lifecycle analysis that includes the marine environment the facility sits in: what does the heat plume do to local water ecology over a decade, and what does the steel pressure vessel cost to manufacture and retire. Comparable load profiles to a land-based reference of the same IT capacity, run side by side over months. Until those land, the public record is what HiCloud and Chinese state media put in it, and the global demand curve keeps climbing.

The Hainan module has run for more than two years; Shanghai is the next step. The question worth carrying from one announcement to the next is whether the engineering survives independent measurement at sustained load, with a marine lifecycle audit, and at a scale that matches the demand curve AI is putting on the grid. Submersion is a mechanism. It is not, by itself, a solution.