Microsoft put a data center on the ocean floor. Then it pulled the plug.

In 2018, Microsoft lowered a twelve-rack, 864-server pod onto the seafloor off Scotland's Orkney Islands, bolted an underwater cable to the shore, and left it there. Two years later the company pulled the pod back up and reported that the servers had failed at roughly one-eighth the rate of comparable land-based equipment, as summarized in a JapanToday analysis by a University of South Florida researcher who studies the environmental implications of emerging technology. In 2024 Microsoft ended Project Natick, the research program that built the pod, and has not publicly explained why. The category question the pilot was meant to answer is now being tested by other operators with very different answers.

The lower failure rate came from a sealed nitrogen atmosphere, stable temperature, no humidity, and almost no human handling. A land-based data center is opened, swapped, and dusted constantly. Every intervention that lowers failure rate also slows the upgrade cycle, and AI training fleets turn over every 18 to 36 months. Microsoft has never reconciled that trade in public, and the category inherits it.

A subsea data center is not a faster data center. It is a slower, steadier one, and that profile sits inside a wider search for siting options: the AI buildout is bumping into land, water, and permitting limits faster than the industry is solving them, and underwater is one of four non-traditional branches, alongside space launch, modular nuclear, and edge micro-data centers, now in active testing. Of those four, underwater has the most real-world operational data and the fewest active commercial operators.

The newest test sits off the Chinese coast. In May 2026, the Shanghai Lingang undersea datacentre, built by HiCloud Technology and state-owned China Communications Construction, went live more than six miles offshore and about ten meters below the surface, with a 24-megawatt capacity and power drawn from a nearby offshore wind farm. The Guardian reported the launch, and the offshore wind trade press framed it as the world's first wind-powered underwater facility. Chinese government statements claim the project cuts power consumption by more than one-fifth compared with land-based equivalents, a figure that should be treated as vendor-adjacent until independent measurements appear. Even if the savings hold, they come from seawater cooling, not from the servers themselves needing less energy.

The third data point is a memorandum, not a deployment. On March 27, 2025, NYK Line, NTT Facilities, Eurus Energy, MUFG Bank, and the City of Yokohama signed an MoU for a 25-meter by 80-meter floating data center off Osanbashi Pier, according to NYK's release. The unit is sized as a disaster-countermeasure and berthing facility, runs on solar plus battery storage, and is explicitly framed as a demonstration. The five signatories include a shipping line, a telecom infrastructure arm, a wind developer, a megabank, and a city government, an unusual mix for a single infrastructure demo. The unit is sized to clear Japan's permitting process before anyone files a full subsea build.

Seawater carries heat away more efficiently than air, which lets operators shrink or eliminate chillers and reclaim fresh water that a land-based plant would burn through in evaporative cooling. Total electricity demand is unchanged, and the carbon footprint follows the grid mix behind the cable, not the salinity of the cooling medium. Marine thermal discharge, acoustic effects on marine life, and the long permitting cycle for a structure that has to be retrieved for hardware swaps are new problem sets, not solved ones.

Microsoft has not published a follow-up study since ending Project Natick. China's power-reduction figure for Shanghai Lingang is sourced to government statements, not to a peer-reviewed audit. Japan's floating unit is sized to test permitting, not capacity. Until at least one of the three projects publishes independent operational results, the category is a portfolio of pilots, and the gap between cooling efficiency and total energy cost is the line the next eighteen months will be drawn on.