AI is running out of room on land. The ocean might be the next stop.

AI's appetite for power and water is running into freshwater limits, grid queues, and urban land. The industry's first real test of putting servers underwater, Microsoft's Project Natick off Scotland, recorded failure rates roughly eight times lower than comparable land-based systems during a two-year trial. Microsoft ended the program in 2024 because the sealed pods were too hard to expand and repair. The current wave of ocean and floating data center projects is the next attempt to solve the same constraint. So far, the constraint just moves.

Microsoft deployed an 864-server pod 117 feet below the surface of the North Atlantic near Orkney in 2018 and ran it for two years in a nitrogen-filled, sealed environment that eliminated human contact and held temperature and humidity constant. The reported failure rate was about one-eighth that of land-based peers. By 2024, the company had stopped the program, and the most-cited reason was operational: surfacing the pods to swap a failed drive or upgrade a generation of chips was expensive enough to break the model. The data center of the future, the company concluded, wasn't going to be at the bottom of the North Sea.

The current wave isn't only underwater. It's a mix of underwater pods, floating barges, and shore-based facilities cooled by seawater. In Shanghai, a wind-powered underwater data center launched in 2025 and began commercial operations in 2026; the operator reports it uses seawater cooling and consumes at least 30% less electricity than a comparable land-based site, per The Conversation's summary of the Chinese rollout, though the figure is reported as "reportedly" in the syndicated coverage and hasn't been independently verified. Singapore's Keppel is planning a four-story floating data center scheduled to open in 2028. South Korea has a plan in Ulsan for an underwater facility that could hold more than 100,000 servers, the largest in the catalog so far. Japan is testing container-sized data centers on floating solar-powered platforms. Maine and Portugal have early-stage proposals for tidal- and seawater-cooled sites.

The capital is following the bet. In May, the Oregon-based startup Panthalassa raised $140 million to build wave-powered floating data centers, with a round led by Peter Thiel according to GeekWire. A separate Ars Technica piece puts the broader pool of "floating AI data center" bets at $200 million, a number worth reconciling against the $140 million Panthalassa figure. Singularity Hub, summarizing the demand side, notes that technology companies are planning to spend roughly $750 billion on data centers this year. Grid power is the binding constraint.

The pitch for the ocean is straightforward. Seawater is cold, dense, and abundant; it can absorb the heat a server farm produces without evaporating the drinking water a conventional cooling tower would consume. The Conversation calls the approach a "useful tool" for cutting freshwater demand, but stops short of calling it a solution. The ocean doesn't generate electricity on its own, so every floating or underwater facility still depends on a cable to shore, a generator on deck, or, in Panthalassa's case, a wave-energy system that hasn't been built at this scale before.

The remaining constraints are operational and ecological. The Natick lesson is the operational one: a sealed pod that has to be raised, opened, and resealed to swap a failed drive is a maintenance problem, not just an engineering curiosity. A floating barge running 24/7 in a tide zone introduces the same headache in a different form. The ecological constraint is seawater discharge. Existing Chinese and Portuguese sites have reported heated effluent raising local water temperatures by less than 1°C, but researchers warn that the same plume could affect dissolved oxygen, pH, and the small marine ecosystems that surround a coastal data center at scale.

The buildout is going to keep coming. The constraint on land is too real for the industry to wait. The question is whether the ocean, in shifting the constraint from freshwater and grid capacity to repair windows and marine ecology, has actually moved the problem, or just put it somewhere harder to see.