Artificial intelligence is thirsty. According to the International Energy Agency, data center electricity consumption is set to double — from around 485 TWh in 2025 to nearly 950 TWh in 2030 — and a single 100 MW AI data center can consume between 1.5 and 3 million m³ of water per year for cooling. What if, instead of spending freshwater and electricity to cool servers, we put them where it's already cold: underwater or in the Arctic?
The idea: take the computer to the cold
Heat is the great enemy of a data center. Traditionally, enormous amounts of electricity and freshwater are spent on air conditioning. Nature, however, offers free cooling: the deep sea and polar regions are permanently cold. The proposal is simple — submerge the servers in the ocean, or install them in the frozen north, and let seawater or Arctic air do the work.
Project Natick: Microsoft underwater
Between 2018 and 2020, Microsoft sank a sealed capsule with 864 servers off the Orkney Islands (Scotland), passively cooled by seawater. The results were notable: an energy efficiency index (PUE) of 1.07, compared to the industry average of 1.67 — about 36% more efficient — and a failure rate eight times lower than on land-based servers (no oxygen or humidity to corrode, and no human hands to interfere). Nevertheless, Microsoft ended the project in 2024, concluding that large-scale operation would be difficult to commercialize and maintain.
China advances where Microsoft stopped
Where Microsoft retreated, China accelerated. Off the coast of Sanya, on the tropical island of Hainan, a commercial underwater data center is being built, with about a hundred capsules to be sunk to host large-scale AI computing, using seawater as a cooling system. It's the first truly commercial bet on this technology.
The frozen north: free cooling on land
You don't have to go underwater. Nordic countries like Iceland, Norway, and Sweden have become magnets for data centers precisely because they combine cold air year-round (natural cooling, almost without air conditioning) with cheap renewable energy (hydroelectric and geothermal). Result: much less electricity and freshwater spent on cooling.
The flip side
It's not a magic solution. Submerging servers raises questions: the heat returned to the ocean can affect local marine ecosystems; maintenance and repair underwater are complex; and the massive placement of infrastructure in the ocean has environmental impacts to be evaluated. Taking everything to the Arctic also concentrates geopolitical and environmental risk in a fragile region.
Conclusion
Underwater and cold-climate data centers are a promising and real direction — not science fiction. They drastically reduce the use of freshwater and energy for cooling, and Project Natick's numbers prove the concept. But they don't replace the essential: more efficient computing, clean energy, and reuse of residual heat. They're an ingenious part of the answer to AI's footprint — not the whole answer.
Sources: Microsoft (Project Natick); IEA Energy and AI 2025; reports on the Hainan, China underwater project.