Underwater data centers fix the cooling bill, not the power bill

Shanghai's wind-powered underwater data center, which launched in June 2025, began full commercial operations in May 2026 ^[11], eleven months later ^[19], while Microsoft, which ran the original experiment in this field, shut its program down in 2024 and declined to say why ^[8]. The gap between those two facts is the story: submersion is a thermal engineering decision that has been marketed as an environmental one.

The physical case is not weak. The AI buildout is driving demand for data centers and with it concern about energy use, water consumption and carbon ^[1], and developers are looking offshore in the hope of better cooling efficiency with less fresh water and less land ^[2]. Shanghai's US$226 million facility uses seawater as coolant rather than refrigerating fresh water, which its operators say cuts the electricity needed for cooling and leaves the site using at least 30% less electricity than a traditional data center, with offshore wind reducing fossil fuel reliance ^[12]. Ulsan, South Korea, began planning an underwater site in 2025 for more than 100,000 servers on the same 30% claim ^[15]. Keppel started building a four-storey floating data center in Singapore in 2026, due to open in 2028 ^[14], roughly a two-year build ^[20], again on seawater cooling and again to avoid consuming scarce land ^[14]. Japan opened a container data center on a floating platform near Yokohama in 2025, powered by on-platform solar with battery storage, with testing running through March 2027 ^[13].

Note what the mechanism is in every case. The saving comes from deleting the chiller plant, not from reducing what the servers themselves consume, which means the load a grid has to serve barely moves. The researcher whose work the source article summarises, and who studies the environmental and organisational implications of AI infrastructure, is explicit that moving servers into the ocean does not make energy consumption and carbon emissions disappear, and that it adds fresh concerns about marine harm, regulation, and how operators maintain and expand hardware they have sealed underwater ^[3].

The reliability data is the genuinely interesting result. Microsoft's 2018 Orkney Islands deployment held 864 servers connected to shore by cable ^[5], and after two years the company reported failures at about one-eighth the land-based rate ^[6], roughly 87% fewer ^[17], hypothesising that a sealed vessel spares equipment from oxygen, humidity, temperature swings and technicians bumping into it ^[7]. Microsoft still walked away, and outside analyses point to regulatory friction, including environmental permits, plus the desire to swap hardware quickly ^[9]. It went back to land, where sites are bigger, easier to expand and easier to reach ^[10].

That is the tension nobody has resolved. Ulsan's plan is more than a hundred times the server count of the Orkney box ^[18], and Maine's DeepGreen Western Passage wants a submersible AI facility in the Bay of Fundy on tidal power ^[16], which is a marine construction and consenting problem before it is a computing one.

Watch whether Yokohama's test produces published numbers before March 2027 ^[13], and whether any of these projects triggers a formal marine permitting regime rather than a case-by-case negotiation.