China now has commercial underwater data centers, and that is worth taking seriously. It is also worth not losing our minds over it, which is apparently harder than it should be whenever servers, seawater, and artificial intelligence appear in the same sentence.

The basic pitch is attractive. Data centers generate a lot of heat. The ocean is very large and very good at absorbing heat. Put sealed data center modules underwater, let seawater do a big chunk of the cooling work, use less land, and perhaps site computing closer to coastal demand. It sounds obvious in the same way that putting solar panels in deserts sounds obvious, right up until transmission, maintenance, dust, markets, and permitting wander in carrying clipboards.

…

This is not imaginary technology. Microsoft’s Project Natick put a 12-rack, 864-server data center pod off Scotland’s Orkney Islands in 2018 and retrieved it in 2020. Microsoft reported that the servers in the underwater pod had a failure rate one-eighth that of the land-based control group, likely helped by the sealed, dry, nitrogen-filled environment and stable temperatures. That was a genuinely interesting result, not a marketing hallucination.

…

China has moved further than Microsoft did. A Shanghai underwater data center project in the Lingang area has entered commercial operation, reportedly with a 24 MW target scale and seawater cooling linked to offshore wind power. There have also been Chinese underwater data center developments around Hainan. These are not bathtub demos or university tank experiments. They are real infrastructure projects.

…

Underwater data centers solve a real problem, but only one part of a real problem. Cooling is a major issue for data centers, … But data centers are not just cooling problems. They are power interconnection problems, fiber problems, land-use problems, latency problems, tax-base problems, workforce problems, water problems, construction problems, and increasingly political problems. Moving the racks into the ocean does not make the rest of that disappear. It changes the balance of constraints.

…

A server that fails in a conventional data center is annoying. A technician walks across a floor, opens a cabinet, swaps a part, logs the work, and goes for coffee. A server that fails inside a sealed module on the seabed is a different creature. Maybe the module is designed to run without repair until its planned retrieval. Maybe redundancy absorbs the losses. Maybe the entire pod comes back to shore on a vessel. All of those are workable design choices, but none of them are magic. They are maintenance philosophies with cost, availability, and operational consequences.

Then there are the cables. Underwater data centers still need power and data connections unless they are part of the newer floating fantasy genre in which AI chips bob around offshore, powered by waves, talking to the world by satellite, and somehow becoming economically compelling before the ocean has its usual laugh. Subsea cables are mature infrastructure, but mature does not mean trivial. They have routes, landing points, permitting, repair vessels, security concerns, and failure modes. Anyone who thinks marine infrastructure is easy has not spent enough time looking at offshore wind, submarine power cables, or undersea telecom repair logistics.

…

A sealed underwater module is not automatically an ecological disaster. Nor is it automatically benign because someone says “the ocean is big” in a confident voice. The questions are local: seabed disturbance, heat plume, antifouling, noise, electromagnetic effects from cables, construction impacts, retrieval, corrosion protection, and what happens over a multi-decade deployment cycle. The answers will vary by site.

…

There is also a cybersecurity wrinkle that sounds like a movie plot until researchers start testing it. University of Florida researchers and collaborators showed that underwater data center systems could be vulnerable to acoustic attacks because sound travels well in water and can interfere with hard drives and system operations. That does not mean every underwater data center is one pool speaker away from collapse, but it does mean the underwater environment introduces different attack surfaces than a fenced warehouse beside a substation. Different is not necessarily worse, but it is not free.

…

For some places, underwater data centers may make sense. Dense coastal cities with expensive land, constrained cooling water, strong port infrastructure, nearby clean electricity, and a policy desire to develop subsea industrial capability could justify them … But edge cases are not the same as a dominant pathway. Underwater data centers are likely to be a real niche, not a global reset of digital infrastructure. They are cool in both senses of the word, but cooling advantage alone is not enough. Infrastructure is where single-factor stories go to be mugged by reality.

…

Underwater data centers are real. They are interesting. They may even be useful. But if someone tells you they are the obvious future of AI infrastructure, ask them how the failed server gets replaced, where the power cable lands, who owns the marine risk, what the environmental permit says, and what land-based alternative they forgot to compare it with. The sea is very good at cooling things. It is also very good at making easy stories expensive.

…

Web Archive link