Underwater Data Centers Emerge as Viable Option for AI Infrastructure
Microsoft's Project Natick experiment aimed to build an underwater data center off the coast of Scotland in 2018. The project involved dropping a steel cylinder, nicknamed Northern Isles, into the North Sea and leaving it for two years before recovering it.
The cylinder contained 855 servers across 12 racks, held 27.6 petabytes of storage, and drew a total of 240 kW. To prevent damage from corrosion and human error, the team replaced ordinary air with dry nitrogen and sealed the tube before lowering it into the ocean.
When the cylinder was recovered in 2020, it showed surprisingly low failure rates for its servers. Only six of the 855 servers failed compared to eight out of 135 servers on land, a significant improvement over traditional data centers.
The success of Project Natick has sparked interest in building underwater data centers, with Chinese firm HiCloud launching a commercial version off Shanghai's Lingang district in May 2026. The facility, which cost around $226 million, uses offshore wind power and seawater cooling to minimize energy consumption and water usage.
While Microsoft walked away from the project due to logistical challenges and focus on land-based operations, the engineering case for underwater data centers is now closed. These facilities can potentially provide a more efficient way of powering AI models with minimal environmental impact.