In 2018, Microsoft did something that sounds like a prank: it took a data center — the kind of humming, climate-controlled server hall that normally fills a warehouse — sealed it inside a white steel cylinder the length of a shipping container, and dropped it into the sea.

The cylinder, nicknamed Northern Isles, came to rest 117 feet down on the seabed off the Orkney Islands, north of the Scottish mainland. Inside were 855 servers. Then Microsoft walked away and left them there for two years.

Why sink a computer

The experiment, called Project Natick, was chasing a real problem: data centers run hot, and cooling them is enormously expensive and thirsty, burning power and often fresh water to keep the machines from cooking themselves.

The ocean is a free, infinite heat sink. Sink a sealed server pod and the cold seawater carries the heat away for nothing — no chillers, no cooling towers, no water bill. Better still, roughly half the world’s people live near a coast, so an undersea pod could sit just offshore from a big city and cut the distance data has to travel. Orkney added a bonus: its grid runs on close to 100 percent renewable power from wind, tide and wave, so the pod ran green as well as cool.

But there was a serious catch, and it was the whole reason for the two-year wait. Once you weld servers inside a capsule and drop it to the seafloor, nobody can get in to fix anything. On land, when a server dies, a technician swaps it out. Under 117 feet of ocean, a dead server stays dead. For the idea to work at all, the machines had to survive largely untouched.

The number that surprised everyone

When the cylinder was winched back up in the summer of 2020 — hauled dripping onto a ship, its white hull crusted with barnacles, anemones and a film of sea scum — the servers inside had done far better than anyone expected.

The raw count is more striking than the ratio. Of the 855 servers sealed in the sea, just six had failed over two years. Microsoft had run an identical control group on land over the same period — and of those, eight failed, out of only 135. Do the arithmetic and the sea pod’s failure rate was roughly one-eighth of the land group’s: about 0.7 percent underwater versus nearly 6 percent on the surface. As project lead Ben Cutler put it, the failure rate in the water was one-eighth of what they saw on land.

Sealed, drowned and unmaintained for two years, the underwater machines had thrashed the ones people were free to walk up and service.

Why the ocean was kinder

The reasons turn out to be a quiet indictment of the ordinary server room, and they come down to two things the sea provided by accident.

The first is nitrogen. Before sinking the pod, the team pumped out all the ordinary air and replaced it with dry nitrogen. Earth’s atmosphere is about 21 percent oxygen, and oxygen is corrosive — it slowly attacks the metal and components inside every machine. A pure-nitrogen atmosphere is inert; it simply doesn’t eat the hardware. The second factor is even more humbling: nobody was there. On land, servers are surrounded by technicians who bump cables, jostle boards, and cycle the room’s temperature and humidity every time they come and go. The undersea pod had perfect stillness — steady temperature, steady humidity, and not one human hand disturbing it for two years.

Put simply, the two biggest enemies of a server turned out to be oxygen and people. The sea removed both.

The idea that sank, and surfaced elsewhere

Here’s the twist the success story usually leaves out: Microsoft killed the project anyway. In 2024, the head of the company’s cloud operations confirmed there would be no commercial undersea data centers — the lessons about corrosion, vibration and reliability would be folded into ordinary land facilities instead, and Microsoft’s attention had turned to the AI build-out.

But the idea didn’t drown with it. The result landed at exactly the moment the world’s demand for computing — and the crushing cooling costs that come with it — went vertical, driven by AI. A data center cooled by the sea, powered by offshore wind, and paying no freshwater bill is precisely the combination every AI infrastructure planner is now agonizing over. And where Microsoft stepped back, others stepped in: Chinese firms have begun deploying commercial underwater data centers at real scale off their own coast, putting into production the thing Microsoft proved and shelved.

The barnacled cylinder off Orkney was, in the end, a strange kind of success — an experiment that worked so well its own inventor decided not to build it, and that someone else is now building because it worked. Microsoft asked a simple, slightly absurd question: what happens to a computer if you drop it in the ocean and leave? The answer, it turned out, was that it lasts longer than the one on your desk — because the ocean, unlike an office, never touches it and never lets it rust.