The Data Center Goes to Sea — The Hard Part Is the Ship, Not the Servers

There is more than one way to build a data center in the ocean. One path seals the servers into a capsule and sinks it into the sea; the other floats them on the surface. Neither has stopped at research. Both are already running somewhere.

What this piece follows is the second one — the floating data center, the one that stays on the surface. It is also the path Korea's shipbuilders have jumped into.

A floating data center puts servers on a structure afloat at sea and cools them with seawater. It doesn't go under; it rides the surface like a ship. Think of one of those huge offshore oil rigs, except that what it carries is servers rather than oil. A wide, flat cargo vessel — a barge — with server containers set on the deck.

So why stay on the surface instead of sinking it? What is so hard about that ship? And what stage has commercialization actually reached?

Why the sea in the first place? The answer is heat.

Why the Sea — The Real Bottleneck Is Heat

A data center is a facility that gathers thousands to tens of thousands of servers — computers — and runs them together. Turn on a few computers in a room and the room warms up. Gather tens of thousands in one place and you have a heat mass on the scale of a power plant.

A data center throws off nearly all the electricity you feed it as heat. Much the way an electric blanket turns nearly all its power into warmth. The share that leaves the building — the signal that gets transmitted — is under 1%. So a data center is, in effect, a giant heat generator.

"Nearly all" means the electricity that goes into the servers. A very small share stays behind as the work of computation, but in practice it comes out as heat at close to 100%. That isn't exaggeration; it's physics.

So the electricity a data center draws splits into two branches. One is the electricity the servers spend on computation. The other is what gets spent outside the servers — mostly the cooling that pumps that heat back out, with shares like distribution losses mixed in as well. The first becomes heat; the second is spent again to clear that heat away.

When people say data centers are running short, what usually follows is a story about sites. Land prices, permits, and how much land is left. And the site really is one of the constraints that come with building on land.

But rank the constraints and the site is not the first one. There are plenty of cases where land is available and you still can't build. The bottleneck isn't the site. It's heat.

How much electricity cooling takes varies enormously from facility to facility. By the International Energy Agency's reading, servers take somewhere around 60% of the electricity a data center uses, while cooling runs about 7% in a well-built large facility and past 30% in one that isn't — an in-house server room, say. Which means that for the same computation, where and how you dump the heat can swing the electricity spent on cooling alone by several times over.

Why is heat the bottleneck? AI chips throw off far too much of it. The shelf slot you plug servers into is called a rack. The heat coming out of a single rack has shot up. The average rack doubled in two years, and a leading-edge AI rack pushes one slot past 100 kilowatts. That heat load amounts to dozens of space heaters crammed into a narrow closet.

The trouble is that air can't carry this heat away. Air cooling — fans, essentially — has a limit. Past roughly 15 to 30 kilowatts per rack, air can't keep up. Leading-edge AI racks have blown through that line by several times over.

So attention turns to water. Water carries heat far better than air. Submerge the servers in coolant outright — immersion and liquid cooling — and 100 to 200 kilowatts per rack becomes manageable. But this is a technology for getting heat out of the rack. Once out, the heat still has to be dumped somewhere outside, into air or into water. So the real problem moves: where do you put that heat?

The ocean is the largest reservoir of cold water in the world. Pour heat into it and it barely warms — an effectively infinite heat sink. A heat sink being the reservoir that takes heat off your hands.

To measure how much seawater saves you, you need a yardstick. That yardstick is power usage effectiveness, or PUE: the data center's total electricity divided by the electricity the servers actually used. The less that leaks into cooling, the closer to 1. Lower is better. It is like a household electricity bill: the smaller the cooling share, the thriftier the house.

The average land-based data center runs a PUE of about 1.55. A 1.0 would be the ideal case, not a watt spent on cooling. At 1.55, a little over a third of total electricity leaks into overhead like cooling. Let seawater take that heat and the waste shrinks.

This logic is not new. Why do thermal and nuclear power plants so often stand on a river or a coast? To make electricity? No. Because of the heat they have to throw away. A plant can't convert all of its fuel's heat into electricity, and a considerable share gets discarded as heat. It needs cooling water for that waste heat, so it stands by the water.

Data centers are now in the same position. They are heat sources pouring out large volumes of waste heat, so they get pushed toward the water. The force pushing them there is a siting logic they share with power plants: heat that has to go somewhere.

Just don't mistake the two for the same thing. A data center is a heating element that converts electricity into heat. A power plant is the reverse — a heat engine that converts heat into electricity. The directions are exactly opposite. The one thing they share is the siting logic: dump large volumes of waste heat, and you stand by the water.

Power plants stand by the water — beside it. Floating data centers take one step further and stand on it. The same siting logic, taken to its extreme.

And heat is a necessary condition for going to sea, not the only force. Without the heat problem, the ocean adds nothing but cost. But once heat is in play, other forces push from behind: shortages of land, of transmission lines, of water, plus the advantage of sourcing your own power. Those come later.

Unfamiliar Engineering, but Not a New Problem

None of this means the server side is fully solved. Whether AI servers run reliably in marine conditions is still being verified right now. But that is confirming a known problem, while the truly unfamiliar engineering belongs to the ship.

Look at a floating data center as a data center and the core seems to be the servers. Look at it as a huge floating structure whose payload happens to be servers and the picture changes. A structure that floats and holds out for decades — that is an entirely different discipline.

That discipline already has a half-century-old cousin: the FPSO. A vast installation that floats at sea to pump, store, and offload crude oil, holding out for decades without ever putting into port. It survives waves and typhoons, survives the corrosion of saltwater, and runs remotely with no one aboard.

The problems a floating data center has to solve are exactly that list. Mooring, to hold it in place so waves don't drag it off station. A hull that survives typhoons and saltwater. Safety for an offshore facility with no permanent crew. And classification society approval, certifying that the vessel meets the rules. None of it has anything to do with servers, and all of it is offshore engineering's long-standing homework.

Which is why this ship — at least the large newbuilds in the tens of megawatts — gets built by shipbuilders, not by cloud companies. In Korea, Samsung Heavy Industries is out front. Samsung Heavy Industries has designed a 50-megawatt-class floating data center on a barge-type hull. HD Korea Shipbuilding & Offshore Engineering has entered as well, and Hanwha Ocean is at the review stage. Companies that have built ships for half a century have turned toward data centers.

One misreading rides along with this FPSO analogy. Let's head it off up front. An FPSO carries oil; a floating data center carries servers. What they carry could not be more different. What they share is only the engineering — a huge unmanned structure that floats at sea and holds out against extreme conditions. This is not a claim that both of them pump oil.

Taking One Ship Apart, Piece by Piece

It starts with the hull. A wide, flat barge-type hull sets the data center on the water. That single move erases one of the big bottlenecks on land: no site is needed. You skip the years of buying land, clearing permits, and pulling in transmission lines.

Now that it sits on the water, it has to be held against drift. That's mooring. Mooring means fixing a floating structure in place with anchors, chains, and lines so that waves, wind, and current can't push it off. It's the same as tying a ship to a pier with rope, or staking a tent to the ground. Mooring is exactly where the FPSO engineering we just saw carries straight over.

Standing in place, it now has to shed heat. That's cooling. Pumps draw seawater up, it takes the heat off the servers, and it goes back to the sea. The largest heat sink in the world, kept alongside and put to use.

Electricity is needed too. Samsung Heavy Industries' design takes power along two paths. One is a subsea cable pulling electricity from shore. The other is generating electricity right on the vessel from LNG. LNG is liquefied natural gas, and a fuel cell turns it into electricity. Generate your own and you don't have to wait on the grid onshore.

Last comes the module. Servers go on in container units, so when one ages or fails you swap the whole thing out. Why that matters is what the next section returns to — and it is the core of this piece.

Threaded into one line, the sequence runs like this. Site, transmission, and heat bottlenecks → the hull erases the site → mooring holds the ship in place → seawater takes the heat → onboard generation bypasses transmission → the module opens the door to replacement. The result is a compute platform you can stand up right in front of coastal demand and keep repairing.

So what floating sells is not the coldest water. It's the most practical deployment: no site, no transmission line, standing up quickly in front of the demand.

The industry also says floating sidesteps even the regulatory bottleneck facing land-based data centers. Several US states are indeed putting the brakes on large data centers. But "offshore, therefore outside regulation" is not yet a verified claim. Not needing a site and being exempt from regulation are two different things. If anything, regulation of offshore facilities is closer to a gray zone.

Why Float It Instead of Sinking It — A Deliberate Cooling Trade-Off

On cooling alone, floating on the surface is not the best answer.

The best way to cool is something else: seal the servers in a capsule and sink it into cold seawater outright. That is the subsea capsule. Picture a sealed, submarine-shaped tube set down under the sea.

The subsea capsule is the star pupil of cooling. With the servers submerged in the middle of cold seawater, there is no better place to dump heat. Microsoft's experiment, Project Natick, is the leading case. It lowered a capsule holding 864 servers to a depth of about 36 meters off Scotland and left it there for two years. The inside of the capsule was filled with nitrogen to strip out oxygen and moisture, and Microsoft reported that doing so cut component failures sharply.

But one large constraint comes attached to the elegance. Seal it and sink it, and when something breaks, no one can get to it. This isn't us speculating from the outside. Operators actually running subsea data centers admit it too. The operator of the subsea data center at Lingang in Shanghai, China says that "replacing subsea components is much harder than on land."

Here one property of data centers becomes decisive. A data center is a facility where servers are swapped out and added to endlessly, on a cycle of a few years. New chips come out and get slotted in; demand grows and capacity gets added. For a facility like that, "can't be fixed" is fatal.

So floating makes the opposite choice. It gives up the best cooling on purpose and stays on the surface. Stay on the surface and you can pull a boat alongside to fix it, and you can swap out whole modules. It hands over a little cooling efficiency and buys repairability.

The reason the commercial bets aiming at scale go floating is not that the cooling is better. It's that it can be fixed. What sets scale is not cooling efficiency but operability.

One more thing to make clear. That does not mean the subsea capsule has been pushed into a small niche. The Lingang subsea data center just mentioned has gone past experiment and is in formal operation at 24 megawatts. Subsea is scaling up too, repair problem and all. So the difference between the two paths is not "who can't reach scale" but "how you buy scale."

Set the two paths side by side and it looks like this.

The Price of Staying on the Surface — Floating's Own Weak Points, and Its Hedges

That very choice — staying on the surface to buy repairability — brings a new set of weak points along with it. The constraints of no repairs and a nitrogen seal belong to the sunken capsule, not to floating. Floating's burdens lie elsewhere.

The first is isolation and latency. Latency is the time data takes to travel between a user and a server. The farther away, the larger it gets. It's the delay that makes a voice arrive late on a video call, the "ping" in an online game. Isolation is that a facility at sea gets hard to reach when the communications line goes down.

How does floating handle this? It moors on the coast or in a harbor rather than in open water. Close to land, latency and connectivity are easier to secure. For communications, subsea cable is laid in multiple strands so that one cut doesn't take the facility down.

But this does not mean the problem is already solved. It is only a design choice aiming at an offset. Samsung Heavy Industries itself lists latency and maintenance access as concerns that remain unresolved. And in large-scale AI training, what really trips you up is not the distance to users. It's the ultra-low-latency interconnect that ties chips and racks to each other. So, positioned honestly: large-scale AI training, where tens of thousands of chips have to be bound together at ultra-low latency, may not yet be floating's job. Workloads less sensitive to latency come first. Floating, too, is a card for particular uses rather than a wholesale replacement for land.

The second is mooring and offshore safety. It has to hold out against typhoons and waves, endure the corrosion of saltwater, and run safely with no one aboard. This is not a problem that has to be solved from scratch. The half-century of FPSO engineering we saw earlier carries straight over. Which is why a good share of floating's weak points sit on the shipbuilders' home ground.

The third is thermal discharge. The power plant lens from earlier shows its other face here. Cool with seawater and that heat does not disappear. It becomes warmed water, returns to the sea, and raises the temperature around it. That is thermal discharge — thermal pollution. It's like emptying your bathwater straight into a mountain stream.

This is structurally identical to what power plants have lived with for half a century. Plant discharge is generally regulated to keep the surrounding temperature rise within 7 degrees Celsius. There is also research finding that a rise of that size cuts the growth and reproductive functions of marine organisms by about half. Independent academics warn, too, that as ocean data centers grow, thermal pollution needs watching with care. We won't insist it's emission-free, and we won't sneer that it therefore shouldn't be done. If you're a heat source, this is the physics you inherit.

So what if you used the heat instead of throwing it away? On land, that path already exists: hand the heat a data center gives off to a city's heating network. In Odense, Denmark, waste heat from Meta's data center feeds district heating and warms roughly 10,000 homes. The European Union is pushing it through regulation as well. The Energy Efficiency Directive requires data centers above one megawatt either to put their waste heat to use or to demonstrate, through a cost-benefit analysis, why they can't. But conditions come attached. The heat a data center puts out is lukewarm, generally 30 to 45 degrees Celsius, so getting it into a network that runs at 60 to 70 means lifting it once with a heat pump. And there is one more condition, harder than that one. The heating network has to be within a few kilometers, and there has to be demand that will take the heat all year round.

Open water is the position furthest from those conditions. One data center in Norway stands on land, on the coast, and still gave up on district heating because too few people live nearby; it routed its warmed cooling water to a lobster farm instead. A moored vessel has even less to hand it to. And in fact the facilities that have run on and under the water so far have returned their waste heat straight to the water. The 2026 announcement of the floating project from Japan's Mitsui O.S.K. Lines and Hitachi contains no waste heat reuse either. Going to the easiest place in the world to dump heat also means going as far as possible from anywhere that could use it.

What Stage Has Commercialization Reached

So how far has floating actually come?

There is a skeptical read on Samsung Heavy Industries' progress: "all they got was approval in principle (AIP) from a classification society — isn't it still just paper?" Half of that is right. But the question is mixing up "floating" with "Samsung Heavy Industries." The stage is set not by the technical approach but by who is building.

Take the industry as a whole first, and floating has already moved past concept. Nautilus, in the US, ran a barge-type data center moored in a harbor, for real, for more than three years. Keppel, in Singapore, secured government approval, investment, and customers for a floating data center park and broke ground. That is not paper. It is actually running, and actually being built.

That said, the ones that have actually run on the water so far are not shipbuilders. Nautilus is a data company, and Keppel is a data center developer and operator. The shipbuilders' stage is the next one — the branch that builds large 50-megawatt-class newbuilds from scratch.

And Korea's three shipbuilders, Samsung Heavy Industries in particular, are still a step short of that. Here it matters not to blur four stages. Approval in principle (AIP) from a classification society is an early gate saying "this design satisfies the rules in principle." In building terms, it is the drawings clearing review — not groundbreaking, and not completion. A joint development program (JDP) is a promise to develop together, not an order contract. And an order, and then operation, come after that again.

Samsung Heavy Industries is through the first two of those. The classification society ABS announced basic design approval in May 2026 (it was unveiled at an exhibition in April). It then signed joint development programs with Supermicro and a Greek shipowner. But there is still no actual order, no confirmed customer, and no installation site. In short, this is the classification approval and joint development stage, not a contract won.

Draw the stages as a ladder and it turns honest. At the top is operation: Nautilus for floating, Lingang for subsea. Below that is construction — Keppel. Below that again is design and approval — Samsung Heavy Industries. Operating cases exist on both the subsea side and the floating side. This is not a paradox of architecture. The stage simply differs by who is building, not by which approach.

One more thing. "Floating scales better than subsea" is also design logic, not a track record. In practice the side further along on scale right now is, if anything, subsea. The Lingang subsea data center is running at 24 megawatts. "On the surface you can fix it, so it's better placed to grow" is logic about what comes next, not a result already proven. It is design logic about how you buy scale — the same proposition we saw earlier, that scale opens only where things can be fixed.

And we won't leave you with only the optimistic case. Floating is not a new idea either. Google filed a floating data center patent back in 2009. It even built a barge, then folded the effort in 2014. Regulatory and cost complexity was the reason. Nautilus, too, after three years of running, turned the business over to a sale in 2026. That is less a shutdown in failure than a pivot, after three years of real operation, to licensing the cooling technology. Even so, whether it's a complete withdrawal like Google's scrapping or a strategic retreat like Nautilus's, precedents for stepping back plainly exist.

What about the economics? Take one brokerage's estimate. A land-based data center takes more than three years to get through permitting and grid connection. Floating, it says, can be seen at about two years. The point is that deployment speed is the advantage. But this is an analyst's estimate, not a settled figure.

So the honest picture right now is this. At the industry level, floating is already a reality that is operating and breaking ground. But Korea's three shipbuilders are still at the design-and-approval stage, and there are no orders. The data rejects both the overstatement that it is already a commercial standard and the understatement that it is only paper.

The Ideal and the Reality Come Out of the Same Choice

Compress what a floating data center is into one line and it is this: a floating offshore structure whose payload is servers. The genuinely hard part is not the computing but the ship. Which is why shipbuilders build it, not cloud companies.

And that ship's weak points all hang on a single choice, to stay on the surface. Sink it and the cooling is the best there is, but you can't fix it, so you can't reach scale. Stay on the surface and the cooling is a compromise, but you can fix it, so scale opens up. That one choice brings floating's strength and its new weak points at the same time. The ideal and the reality come out of the same choice.

With the engineering humans have spent half a century developing for structures at sea, we hand over to the ocean, whole, the heat and the power that AI throws off. The elegance of that craft — and the honesty of admitting its commercialization is still scattered across different stages at different companies. Where the floating data center stands right now is somewhere between the two. Not a future that replaces land entirely, but a frontier opening up where the conditions happen to overlap.

What AI pours out is not only data. It is heat, and electricity, and water. That bill has to be paid physically somewhere, and where and how it gets paid is not settled yet. The ship floating on the ocean is one of the answers now being tested for where to hand the bill.