The world's first commercial undersea data center, powered entirely by offshore wind, has gone online in China. The facility uses seawater for natural cooling and bypasses land-use constraints entirely. Microsoft experimented with this years ago but never scaled it. If China can make underwater data centers work at commercial scale, it changes the economics of AI infrastructure.

According to New Atlas, the facility operates off Shanghai's coast in the Lin-hang Special Area, approximately seven months after phase one's completion in October 2025. Offshore wind turbines supply roughly 95% of the electricity needed, with current output reaching 2.3 MW and planned expansion to 24 MW capacity—enough to power approximately 20,000 households.

The cooling system is the real engineering achievement here. Rather than requiring freshwater—a massive constraint for traditional data centers—the facility uses a circulating copper-pipe heat exchange system that transfers thermal energy to the surrounding ocean. Professor Li Zhen from Tsinghua University noted that cooling accounts for "only about one-tenth of total power consumption" for undersea facilities, compared to the 30-40% typical for surface data centers.

The facility houses 192 server racks distributed across four underwater levels, and it reduces land requirements by more than 90% compared to conventional facilities. In dense urban areas where real estate is expensive and power infrastructure is strained, that's a massive advantage. You're not competing with residential or commercial development for space, and you're not stressing the local power grid.

Here's why this matters for AI: training frontier models requires enormous amounts of power and cooling. The current approach—build massive data centers in regions with cheap electricity and hope the grid can handle it—is hitting physical and political limits. Communities are pushing back against data center construction because of the strain on local resources. Underwater facilities could sidestep both problems.

Microsoft's Project Natick proved the concept worked in 2018, deploying a datacenter off the coast of Scotland. But the company never scaled it to commercial deployment. China just did. If the economics work and the reliability holds up, we might be looking at a new model for AI infrastructure—one that doesn't require communities to sacrifice power and water resources to host compute.

The technology is impressive. And for once, it might actually solve a real infrastructure problem.