How Underwater Data Centers Actually Work

Data centres have an unglamorous problem: they run hot. Servers convert electricity into computation and waste heat in roughly equal measure, and a large share of a facility's total power goes not into computing at all but into chillers, fans, and cooling towers fighting to keep the machines from cooking themselves. Meanwhile, sitting just offshore from most of the world's population is an effectively infinite reservoir of cold, freely circulating water. The idea behind underwater data centres is to stop fighting the heat problem and simply move the computers to where the coolant already is. It sounds like a thought experiment, and for a decade that is largely what it was. It is now a commercial reality, though not in the country that pioneered it. Here is how these facilities work, and what has actually happened to them.

 

1. Servers Are Installed in a Sealed Pod

 

The facility begins as a sealed steel pressure vessel, essentially a submarine without a crew, packed with standard server racks. The critical design decision is that no human being will ever enter it once it is closed. That single constraint changes everything about how it is built, because there is no need for walkways, lighting, breathable air, or the space required for a technician to reach a failed component.

The most consequential consequence is atmospheric. Because no one has to breathe inside, the pod can be filled with dry nitrogen instead of air. Oxygen and humidity are the enemies of electronics, driving corrosion and oxidation over years of operation, and removing them creates a chemically inert environment that servers seem to like a great deal. Microsoft's Project Natick vessel held 855 servers in exactly this configuration. The absence of people turns out to matter in a second way too, since nobody bumps a cable, jostles a rack, or introduces dust and vibration while working nearby.

 

2. The Data Center Is Deployed Offshore

 

The completed module is towed or carried out by ship and lowered onto the seabed, generally in relatively shallow coastal water. Project Natick's vessel sat about 117 feet, or roughly 36 metres, down off the Orkney Islands in Scotland. China's commercial deployments off Hainan and Shanghai sit at around 35 metres, depths a recreational diver could reach.

The shallow siting is deliberate and reveals the real logic of the concept. These are not deep-ocean installations; they are coastal ones, placed close to shore because they need power cables, fibre connections, and proximity to users. Roughly half the world's population lives near a coast, so a facility a few kilometres offshore can be physically closer to its users than an inland data centre while consuming no land, generating no neighbourhood noise, and avoiding the local opposition that increasingly greets data centre construction on land. The modules themselves are substantial: the Hainan units weigh around 1,300 to 1,400 tonnes each, roughly the mass of a thousand cars.

 

3. Seawater Provides Natural Cooling

 

Cooling is the entire economic argument. Heat from the servers passes through heat exchangers into the surrounding seawater, which carries it away by natural circulation. Some designs pump seawater through radiators at the back of the server racks. Either way, the ocean does for free what chillers, pumps, and cooling towers do expensively on land, and it does so at a stable temperature that barely varies through the day or the seasons.

The efficiency gains are real and measurable. Data centre performance is usually expressed as power usage effectiveness, or PUE, the ratio of total facility power to the power actually reaching the computing equipment, where 1.0 would be perfect and typical land facilities sit well above that. China's Shanghai underwater facility claims a PUE below 1.15, and the Hainan project has been described as 40 to 60 percent more power efficient than conventional equivalents. Since cooling is where most of the overhead in a data centre lives, eliminating it changes the operating economics substantially.

 

4. Electricity Powers the Facility

 

Power arrives from shore through subsea cables, and the choice of power source has become central to the concept's appeal. Because these facilities sit offshore anyway, they can be connected directly to offshore wind farms, cutting out the grid entirely and pairing a large electrical load with generation that is already in the water beside it.

The Shanghai project in the Lingang Special Area is the clearest example of this logic taken seriously. Reported at 24 megawatts of capacity and drawing over 95 percent of its electricity from nearby offshore wind, it entered full commercial operation in May 2026 after launching the previous June. The pairing is genuinely elegant: offshore wind often generates more power than the nearby grid can absorb, while data centres are enormous, flexible, and location-indifferent consumers of electricity. Putting the two together on the seabed solves a problem for each.

 

5. Data Travels Through Fiber Optics

 

Submarine fibre-optic cables carry data between the servers and users on land. The physics are favourable, since light travels through fibre at enormous speed and the distances involved, a few kilometres offshore, are trivial compared with the continental hops data normally makes.

The latency argument for underwater data centres is therefore about geography rather than the water. A facility positioned just off a dense coastal city can be closer to its users than an inland site hundreds of kilometres away, and for latency-sensitive applications that proximity matters. There is an obvious dependency here too: the pod is entirely reliant on its cable connections, both power and data, and subsea cables are vulnerable to anchors, fishing gear, and deliberate interference, which is a genuine strategic concern as these facilities scale.

 

6. Systems Operate Remotely

 

Everything is run from shore. Engineers monitor performance, power consumption, temperature, and hardware health continuously through the data link, and the servers are managed exactly as they would be in any remote facility. Nobody goes inside.

This demands a design philosophy of redundancy rather than repair. On land, a failed drive or server is swapped out within hours. In a sealed pod on the seabed, a failed component is simply dead for the remainder of the deployment, so the system must carry enough spare capacity to absorb failures gracefully over years. It is closer to how satellites or deep-sea instruments are engineered than to conventional data centre operations, and it is why the reliability results discussed below mattered so much to the concept's credibility.

 

7. Maintenance Happens After Retrieval

 

Rather than continuous servicing, the entire pod is raised to the surface after several years for maintenance, hardware refresh, or redeployment. Project Natick's vessel spent two years on the seabed before being recovered, power-washed, and examined.

Whether this is an advantage or a limitation is the concept's central unresolved question. It works beautifully if hardware lasts and workloads are stable. It works far less well in an industry where server generations turn over rapidly and where the current wave of artificial intelligence demand has made compute density and rapid hardware upgrades paramount. A five-year refresh cycle looks very different in a world where the hardware people want changes every eighteen months, and the inability to physically access a machine that has failed or become obsolete is a real operational cost, not merely an inconvenience.

 

8. Efficiency and Sustainability Improve

 

The claimed benefits are lower cooling energy, reduced emissions, no land consumption, no local noise or water withdrawal, and better hardware reliability inside a sealed, stable, inert environment. On the evidence available, most of these hold up.

The story of what has actually happened, however, is more complicated than the technology alone would suggest, and it is worth telling plainly. Microsoft's Project Natick, which ran from 2015 and produced the landmark results, was quietly shelved in 2024 despite its success, with the company concluding that the approach did not fit the demands of modern cloud and artificial intelligence workloads, though it said it would apply the lessons elsewhere. Analysts have also cautioned that the celebrated reliability comparison was not a perfectly controlled experiment, since the submerged and land-based systems differed in more than just their location. Meanwhile China moved decisively in the opposite direction. Modules were deployed off Hainan from 2022, commercial operation followed, and in 2025 the Lingshui facility was launched as the world's first commercial underwater data centre, with plans reported for as many as a hundred modules at the site, followed by the wind-powered Shanghai project reaching full operation in 2026.

Two open questions remain. The first is thermal discharge. The ocean is an enormous heat sink and a handful of modules will not measurably warm it, but hundreds of megawatts of continuous heat rejection into shallow coastal water is a different proposition, and local ecosystems in exactly those depths are sensitive. Credible environmental impact assessment, rather than efficiency statistics alone, will determine whether this scales acceptably. The second is security, since seabed infrastructure is difficult to protect and increasingly recognised as strategically exposed. The engineering has been proven. Whether underwater data centres become a significant part of global computing infrastructure now depends on economics, environmental scrutiny, and the shape of demand in an industry being reorganised around artificial intelligence.

 

Did You Know?

 

Microsoft's Project Natick demonstrated that an underwater data centre could operate unattended for over two years with a far lower hardware failure rate than a comparable facility on land. The specific numbers were striking: of 855 servers submerged off Orkney, only six failed, against eight failures among 135 servers in a parallel land-based test, a rate of roughly 0.7 percent versus 5.9 percent, or about one-eighth the failure rate. The team attributed the difference to the inert nitrogen atmosphere and to the simple absence of human beings handling the equipment. The most quietly interesting implication was never about the ocean at all: if sealing servers away from oxygen and people is what makes them last, that lesson can be applied to data centres built firmly on dry land.

 

Note: This article reflects the state of underwater data centre technology as of mid-2026, drawing on sources including Microsoft Research, Data Center Frontier, MERICS, and reporting on Chinese commercial deployments. Performance figures such as PUE and efficiency gains are largely operator-reported and have not always been independently verified.