Underwater Data Centres Are No Longer Science Fiction — But Are They Actually the Answer?

Microsoft proved the idea works, then walked away from it. China has one running commercially right now. A small town in Maine is fighting to stop one being built in its bay. Here’s the full, honest picture of where this technology actually stands.

The idea sounds inherently impractical: sink your servers into the ocean, leave them there for years, and retrieve them when the hardware needs upgrading. And yet, one of the world’s most rigorous technology companies ran exactly this experiment for over two years off the coast of Scotland — and the underwater servers outperformed the ones on dry land. Then they shut the project down anyway. Now China has built one commercially. And a small fishing community in Maine is in the middle of a council fight trying to stop one being planned for their bay. What’s actually going on with underwater data centres, and why now?

DW Planet A’s independent explainer on the underwater data centre trend — useful background context, not specific to the projects covered in this article.

The Problem They’re Solving

Data centres already consume roughly 485 TWh of electricity globally per year — and AI is accelerating that number faster than any previous technology shift. The cooling problem is at the heart of it: up to 40% of a conventional data centre’s energy consumption goes not to computation but to keeping the servers from overheating. That’s a staggering proportion of resource expenditure that produces no useful output — it just manages waste heat.

Seawater solves this directly. The ocean is a vast, free, naturally stable heat sink. A data centre submerged in seawater can offload its waste heat into the surrounding water passively, without the elaborate cooling infrastructure that land-based facilities require. The secondary problem is land. A single large AI data centre can occupy hundreds of acres, consume enormous volumes of freshwater for cooling towers, and face fierce local opposition anywhere near populated areas. An underwater facility has essentially no surface footprint.

Microsoft Proved It Worked — Then Walked Away

The most revealing data point in this entire technology’s history is what happened when Microsoft retrieved its Project Natick capsule from the seabed off Scotland’s Orkney Islands in July 2020. The capsule had been submerged since June 2018 — 25 months and eight days, fully unattended — containing 855 servers. When Microsoft’s team examined the results, only six of those 855 servers had failed. Running simultaneously as a comparison, 135 equivalent servers in a land-based facility lost eight.

In percentage terms: 0.7% failure rate underwater versus 5.9% on land. The submerged servers were roughly eight times more reliable. Microsoft attributed this to two factors — the constant temperature stability of seawater eliminating the thermal cycling that stresses components on land, and the use of inert nitrogen gas inside the capsule instead of oxygen, which dramatically reduces corrosion.

So why did they stop? Walsh was direct: “I’m not building subsea data centres anywhere in the world.” The reason wasn’t technical failure — it was the fundamental economics of modern AI infrastructure. The critical problem is that the capsule is sealed. Once submerged, you cannot easily swap in new GPUs, add storage, or upgrade the hardware. For a company running AI workloads where the compute landscape changes every 12 to 18 months, a facility you seal and leave for five years is structurally incompatible with how the industry now operates. The technology worked. The operational model didn’t fit the current demand.

China Went Ahead Anyway

Shanghai Hailanyun Technology — known as HiCloud, a division of Chinese company Highlander — took a different view. In May 2026, China’s Shanghai Lingang underwater data centre entered full commercial operation, positioned 10 kilometres offshore in the Lingang Special Area, sitting at 35 metres below the surface. The facility had been officially launched in June 2025 and construction completed in October 2025 — full commercial operations began in May 2026 following successful initial trials. At 24 megawatts of planned capacity, this isn’t an experiment — it’s real operational infrastructure, powered directly by an adjacent offshore wind farm via subsea cable, eliminating the need for grid routing entirely. HiCloud has already signed agreements for a 500-megawatt expansion.

The numbers are specific and verified by Chinese government sources: cooling electricity consumption reduced by 22.8%, freshwater use eliminated entirely, land footprint reduced by more than 90% compared to an equivalent land-based facility. The facility is being used for big-data annotation and developing domestic large language models — directly serving AI workloads. HiCloud’s parent company was itself inspired by Microsoft’s Project Natick, which is a reasonable summary of the dynamic: a Western company proved the concept, applied the learnings elsewhere, and walked away from commercialisation. A Chinese company built the commercial version.

The Eastport Problem — Why “No Surface Footprint” Isn’t the Whole Story

In June 2026, a proposal surfaced to build an underwater AI data centre in Passamaquoddy Bay, off Eastport, Maine — a small coastal city whose economy depends heavily on lobster fishing, scallops, and clams. Developer Louis Wolfson described the project as tidal-powered, environmentally minimal, and economically beneficial. Eastport residents described it differently.

“The temperature of the pods themselves will heat up the water, and that will kill the lobsters, and the scallops, and the clams, and everything here that Eastport depends on,” said resident Suellen Hendrix at a council meeting. The Eastport City Council rejected a moratorium on the project, but made clear it remains at a very early stage and would require extensive environmental and regulatory review before any construction could begin. The concern isn’t abstract: thermal discharge from submerged infrastructure can alter oxygen levels, pH, and sediment composition in surrounding water. HiCloud is monitoring water conditions at its earlier 2023 Hainan installation — but that monitoring is still ongoing, not concluded.

The Eastport case makes something concrete that the engineering arguments tend to gloss over: “no surface footprint” is not the same as “no environmental impact.” The ocean is not an empty background — it is a working ecosystem. The fishing communities, marine biologists, and coastal planners who will live with the consequences of underwater data centres are not in the same conversation as the infrastructure investors and AI companies who want to build them.

Where This Actually Stands

Frequently Asked Questions

Clayton Samuel (Mr Wangdoo), QFA

Founder & editor, Wangdoo.com. Covers the tech worth knowing about before everyone else is talking about it.