A $226 million data center is now operating on the seabed off Shanghai. The servers are sealed inside pressure-resistant modules sitting about 35 meters below the surface, and the electricity comes directly from an offshore wind farm positioned above them. It entered full commercial operation in May 2026, after a construction phase that started in June 2025.
The facility is run by HiCloud Technology, a relatively young Chinese private company founded in 2020, in partnership with state-backed operators including China Telecom. Its current capacity is 24 megawatts, which is modest: a large land-based data center can run into the hundreds of megawatts. HiCloud has publicly stated plans to eventually build a 500-megawatt undersea deployment. That is an ambition, not a commitment, and the gap between 24 megawatts and 500 is enormous.
The core claim here is about energy efficiency. Cooling is the biggest non-computing cost in any data center, typically accounting for a quarter to 40 percent of total electricity use. The Shanghai facility uses the ocean water around it as a passive heat sink: hot air from the servers heats a refrigerant, which rises to an upper chamber, exchanges heat with seawater, and falls back down again, all without a pump or compressor running. Chinese media reports put the facility's overall energy efficiency ratio at below 1.15, meaning only 15 cents of every dollar of electricity goes to overhead rather than to computing. The global average for large data centers is closer to 1.5. That gap is significant.
This is not the first time someone has tried to build underwater. Microsoft ran a decade-long research program called Project Natick, placing server capsules off California in 2015 and off the Orkney Islands in Scotland in 2018. The results were genuinely impressive: the underwater servers failed at one-eighth the rate of land-based equivalents, with only 6 servers failing out of 855 submerged for over two years. Yet Microsoft walked away from the program in 2024 without commercializing it. The reason was practical: AI workloads require constant hardware upgrades, and you cannot swap out a GPU cluster from 35 meters underwater the way you can in a server room. Maintenance, servicing, and rapid scaling are all harder when your infrastructure is on the ocean floor.
China is moving forward anyway, for reasons that have everything to do with national energy strategy. The country's data center power consumption has been growing at roughly 38 percent annually for the past five years and is projected to maintain a 19 percent growth rate through 2030. Land in Shanghai is expensive and scarce. Renewable energy from coastal wind farms is abundant but physically distant from population centers. Placing the computing infrastructure directly next to the power source, and using the ocean as a free cooling system, sidesteps both problems at once.
The broader context matters for anyone running operations that depend on computing costs. Data center electricity demand globally could approach 1,050 terawatt-hours by 2026, which would put it on par with Japan's total national consumption. Energy costs, water costs, and land costs are already pushing data center pricing up in most Western markets. Power purchase agreement prices rose 35 percent in 2024 alone, driven largely by the surge in AI infrastructure spending.
The maintenance problem is real and unsolved. Replacing or upgrading servers at depth requires specialist marine operations. The facility in Shanghai houses around 2,000 servers right now. Scaling that to 500 megawatts would mean tens of thousands of servers that someone has to service, upgrade, and eventually replace, all below the surface. Environmental questions about heat discharge into coastal ecosystems are also open. Microsoft's Scotland capsule came up covered in algae with no measurable disruption to marine life, but a facility orders of magnitude larger is a different calculation.
For most business operators, the direct relevance here is not the underwater part. It is the energy constraint underneath it. The cost of running AI tools is tied to electricity and cooling costs, and those costs are rising structurally. Any solution that materially cuts the overhead, whether underwater, underground, or at the poles, will eventually show up in the pricing of cloud services.