EXPLAINER

Why Data Centers Are Moving Into the Ocean

Microsoft sank a data center off Scotland and its servers failed one-eighth as often as on land. China now runs them commercially. How underwater data centers work.

Why put a data center in the ocean?

Three reasons, in Microsoft's own framing: half of us live near the coast, so the customers are already there; seawater is a free, always-on heat sink; and a sealed pod keeps oxygen, humidity, and people away from the hardware. Land data centers spend a large share of their electricity and water on cooling. The ocean offers to do that job for free — and to do it with no freshwater at all. How much water does a data center use? sets out what land cooling costs in water; the ocean version of the same question has a different answer, because the cooling water is the sea itself.

The idea is older than the deployments. Google patented a "water-based data center" in 2008 — wave-powered platforms floating three to seven miles offshore. The barges Google actually built in 2011–2013 were never data centers at all: the company confirmed in 2013 they were showrooms, halted work late that year, and sold them off in 2014. The first real test of servers on the seabed came from Microsoft.

Microsoft's Project Natick: the experiment that proved it

Project Natick ran in two phases. Phase 1, in 2015, was a single rack inside an eight-foot steel capsule named Leona Philpot, after a character from Microsoft's Halo games. It sat 30 feet down off the California coast near San Luis Obispo for 105 days, filled with pressurized nitrogen, and came back up working. That proved a sealed pod could survive real ocean conditions. It could not prove anything about long-term reliability — three months is not years.

Phase 2 did. In June 2018, Microsoft lowered a 40-foot steel pressure vessel — Northern Isles — onto the seabed off Scotland's Orkney Islands, 117 feet down, at the European Marine Energy Centre. Inside were 12 racks, 864 servers, and 27.6 petabytes of storage, sealed in dry nitrogen, powered by local renewables, and connected to shore by a single subsea power and data cable. Nobody touched it for just over two years. It ran Azure workloads the whole time, and even contributed spare capacity to COVID-19 research (Folding@home and World Community Grid).

When the vessel was hauled up in July 2020, crusted with barnacles and anemones, the headline result was reliability. Microsoft reported that the submerged servers failed at one-eighth the rate of an identical land-based control group running the same workloads — six failures in the sea against eight in a 135-server group on land. The company also reported a power usage effectiveness of 1.07, with no freshwater consumed for cooling at all. Microsoft's own explanation was a hypothesis, not a proven cause: the nitrogen atmosphere is less corrosive than oxygen, and the absence of people meant nobody bumped, jostled, or mis-handled the hardware. The vessel, its heat exchangers, and its servers were recycled, and the seabed was restored.

Watch the related Short →

Why Microsoft walked away anyway

In June 2024, Microsoft confirmed Project Natick was over. Noelle Walsh, who runs the company's Cloud Operations and Innovation division, told Data Center Dynamics: "I'm not building subsea data centers anywhere in the world." The project had worked; the company would apply what it learned to land data centers instead — mostly liquid cooling and robotics.

The reasons are structural, not technical. A sealed pod is fixed capacity: you cannot open it to swap in a new generation of GPUs, and in an AI build-out that is disqualifying. A land data center swaps hardware in place; an underwater pod has to be hauled to the surface, opened, resealed, and lowered again. Seabed permits cross jurisdictions in a way a land permit does not. And the economics that looked attractive at 2020 demand levels did not justify the extra cost of building and retrieving steel pressure vessels. Natick was designed to run up to five years without maintenance — but "without maintenance" also means "without upgrades."

China: the idea goes commercial

China did not walk away. Beijing Highlander Digital Technology, through its subsidiary HiCloud, built an underwater data center off Lingshui, Hainan, in the South China Sea. The first commercial capsule — a 1,300-tonne sealed vessel at about 35 metres depth — was deployed in March 2023, and in February 2025 an 18-metre module holding more than 400 servers was added, with the Hainan underwater AI computing center going live the same month. China Daily described it as the world's first commercial underwater data center: the first one run as a paying service rather than a research pilot, with plans for 100 pods in three phases. Highlander reports a power usage effectiveness of about 1.1 for the Hainan site — a company figure that has not been independently audited.

The second site is bigger and wind-powered. A 24 MW underwater data center off Shanghai's Lingang district, powered by offshore wind, reached full commercial operation in May 2026. The contrast with Microsoft is the point: the same technology, the same ocean, a different decision about whether the economics and the permits add up.

Watch the related Short →

How the cooling works — and its limits

A submerged data center does not pump seawater through its servers. Inside the vessel, fans move air across heat exchangers, which pass the heat to a liquid that is pumped to exchangers on the outside of the hull, where the ocean carries it away. There are no chillers, no cooling towers, and no freshwater consumed. That is why Microsoft could report a PUE of 1.07 and zero freshwater use for the project.

The limits are physical. Heat has to go somewhere, and at scale it goes into the surrounding water. Microsoft reported no detectable temperature impact a few metres downstream of its single module — but one module is not a fleet, and the thermal effect of dozens of pods on one site has not been measured in public. Marine life also grows on everything: the Natick vessel came up coated in barnacles and anemones, which is a maintenance problem for heat exchangers and a retrieval problem for anything left down too long. How liquid cooling moves heat off a chip explains the land version of the same heat-transfer problem.

The problem nobody solved: maintenance

Everything about an underwater data center is a trade for reliability: you seal the pod, fill it with nitrogen, and walk away. The price is that you cannot fix anything. A failed drive, a dead fan, a faulty power supply — on land, a technician swaps it in minutes. Underwater, the part waits until the whole vessel is retrieved, which is a marine operation with a ship, a crane, and a weather window. That is why the design goal is years without a visit, and why the failure-rate result mattered so much: the concept only works if very little fails.

It is also why the GPU upgrade cycle is the real obstacle. AI hardware turns over every year or two. A pod sealed for five years would be running hardware two generations old by the time it surfaces. For steady workloads that do not need the newest chips — storage, inference, content delivery — that may be acceptable. For training, it is not.

Can a speaker crash an underwater data center?

Yes — in a lab, and in a lake. Hard disk drives are vulnerable to sound: at the drive's resonant frequency, acoustic waves vibrate the read/write head and the spinning platters until the drive cannot read or write. Researchers at the University of Michigan showed this in air in 2018 ("Blue Note"), using ordinary speakers. Researchers at the University of Florida then took it underwater: in "Deep Note" (2023) they showed acoustic interference could damage the availability of hard-disk storage in submerged enclosures, and in "AquaSonic" (IEEE Symposium on Security and Privacy, 2024) they showed, with a realistic testbed and open-water lake experiments, that a commercial underwater speaker could cut a fault-tolerant RAID 5 storage system's throughput by 17% to 100%, knock a distributed filesystem node offline after 2.4 minutes of sustained sound, and degrade throughput at a distance of 6.35 metres.

Three things keep this in proportion. First, it is a demonstrated vulnerability, not a recorded attack — no attack on a real commercial underwater data center has been reported. Second, it affects spinning hard disk drives, not solid-state drives, which have no moving parts to vibrate. Third, the researchers' own work includes defenses: software that detects acoustic disruption and reallocates resources before the system crashes. The point is not that underwater data centers are doomed. It is that water carries sound further than air, and a sealed pod cannot be physically guarded the way a fenced building can.

Watch the related Short →

Submerged is not floating

"Data centers in the water" covers three different things, and they should not be confused.

Submerged pods sit on the seabed: Microsoft's Natick, China's Hainan and Lingang sites, and Subsea Cloud's Jules Verne pilot pod off Port Angeles, Washington (a 20-foot container, about 9 metres down, pressure-equalized with dielectric fluid instead of nitrogen; its commercial status is unconfirmed).

Floating data centers sit on the surface, on a barge or a converted ship, moored in a harbour or river. Nautilus Data Technologies has operated one since 2021 — Stockton 1, a 7 MW facility on a barge in the Port of Stockton, California, cooled by river water in a closed loop. In March 2026, Mitsui O.S.K. Lines, Hitachi, and Hitachi Systems signed an agreement to develop floating data centers converted from second-hand vessels, with operations targeted for 2027 or later. And a Japanese consortium (NYK, NTT Facilities, Eurus Energy, MUFG Bank) announced in October 2026 a port-quay model: computing tied up alongside a dock, using existing port infrastructure — our news brief on the NYK concept sets out why the dock matters, and what the concept does not yet prove. Floating is easier to service than submerged — you walk onto a barge — but it gives up the sealed, stable environment that made Natick reliable.

Seawater-cooled on land is a third category: a normal land building that draws seawater for cooling. It is not underwater at all.

From shore to seafloor, in order

  1. Build and seal on land. The vessel, racks, and cooling loop are assembled and tested in a factory, filled with dry nitrogen, and sealed. Nothing inside is meant to be touched again until retrieval.
  2. Tow to site and lower. The pod is loaded onto a barge, towed to the site, and lowered onto the seabed with a crane — a marine operation that needs a weather window.
  3. Connect power and fiber. A subsea cable brings power in from shore and carries data back. Without that cable the pod is an expensive anchor. Getting power to the site is its own project — how a data center connects to the grid explains what the shore side of that cable has to do.
  4. Run lights-out. The pod operates unmanned for years, monitored remotely. There are no visits, no repairs, and no upgrades until retrieval.
  5. Retrieve to service. A ship hauls the vessel up, the hardware is swapped or recycled, and the pod is resealed and lowered again — or retired. The seabed is restored.

Frequently asked questions

Does the ocean actually cool better than a land data center?

For the cooling itself, yes: seawater is a large, stable heat sink, and Natick's hull heat exchangers needed no chillers and no freshwater. Microsoft reported a PUE of 1.07 for the project. But PUE only measures the energy a facility uses while running — total facility energy divided by the energy the IT equipment uses, of which cooling is the largest part. It says nothing about the energy cost of building a steel pressure vessel, towing it to sea, and hauling it back up.

Why did Microsoft quit if it worked?

Because the pod cannot be upgraded. A sealed vessel is fixed capacity for the whole deployment, and AI hardware turns over too fast for that. Seabed permits and the cost of marine operations were the other two reasons Microsoft cited. The company kept the lessons — nitrogen atmospheres, no-touch operation, liquid cooling — and applied them on land.

Are underwater data centers actually running today?

Yes, in China: the Hainan site has run as a commercial service since 2023, and the wind-powered Lingang site reached full operation in May 2026. In the US, the only operating waterborne data center is Nautilus's floating barge in Stockton, California — floating, not submerged. Microsoft has none in the water.

Do underwater data centers harm the ocean?

The honest answer is that we know very little at scale. Microsoft reported no detectable temperature impact a few metres downstream of its one module, recycled the vessel, and restored the seabed. But that is one module for two years. Nobody has published environmental monitoring for a multi-pod commercial site, and the Hainan and Lingang operators have not published theirs. Heat, noise, and the physical footprint of dozens of pods are open questions, not settled ones.

Do I need to worry about the sound attack?

Only if you run spinning hard disks underwater. The AquaSonic research is real and peer-reviewed, but it is a laboratory and lake demonstration, not a recorded attack, and it targets HDDs. Solid-state storage is not affected the same way, and the researchers published detection and mitigation methods alongside the attack.

Where are underwater data centers on the map?

Our map of US data centers covers land sites. Underwater sites are few enough to list: Natick (retired, Orkney), Hainan and Lingang (China), the Subsea Cloud pilot (Port Angeles), and Nautilus's floating barge (Stockton). None of them is a hyperscale campus — yet.

Sources and further reading

Glossary — plain definitions for words used here.