Via newatlas.com
Project Natick proved servers fail less underwater, but Microsoft is betting on land-based AI clusters instead
Microsoft has pulled the plug on Project Natick, its decade-long experiment in sinking data centers to the ocean floor. The company confirmed in June 2024 that it has no active subsea data center projects and won’t pursue commercial underwater deployments, despite the initiative producing some genuinely impressive results.
What Project Natick actually proved #
Project Natick launched in 2013 with a simple premise: what if you could cool servers with the ocean itself, skip the massive HVAC systems, and seal everything in a nitrogen-filled tube that humans never need to enter?
During Phase 2, which ran from 2018 to 2020 off Scotland’s Orkney Islands, Microsoft submerged a pod containing 855 servers. Over more than two years of operation with minimal maintenance, only six of those servers failed. That translates to a failure rate of roughly 0.7%, compared to about 5.9% for equivalent land-based servers.
The project also achieved a power usage effectiveness (PUE) of 1.07 using seawater cooling. Most conventional data centers operate somewhere around 1.3 to 1.6.
Why Microsoft walked away #
Microsoft’s current infrastructure priorities center on massive land-based AI clusters designed to handle the workloads generated by models like those powering its Copilot products and its partnership with OpenAI. The company’s Fairwater facility in Wisconsin represents this new direction: enormous, purpose-built complexes optimized for the specific demands of training and running large language models.
Microsoft has indicated it plans to apply lessons from Project Natick to its terrestrial operations. Liquid immersion cooling, which shares DNA with the seawater cooling approach used underwater, is one technology the company is exploring for land-based facilities.
The ocean isn’t empty yet #
While Microsoft exits the subsea data center space, Chinese operators have been particularly active, with facilities reaching up to 24 MW scale. These aren’t research experiments. They’re operational infrastructure serving real workloads.
What this means for the infrastructure race #
For investors watching the data center buildout, capital is flowing toward proven, scalable land-based infrastructure in the near term. Companies supplying cooling technology, power infrastructure, and construction services for conventional facilities stand to benefit most from the current cycle. Ocean-based alternatives remain a longer-duration bet, with Chinese operators currently holding the most operational experience in that space. Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our