In spring 2018, a sealed steel cylinder packed with 864 servers went over the side of a barge off Scotland’s Orkney Islands and settled on the seafloor at 117 feet. It stayed there for roughly two years, running without a single technician stepping inside, and when Microsoft hauled it back into daylight the company reported a result that was hard to argue with on the spreadsheet: the servers underwater had failed at one-eighth the rate of an identical rack of machines kept on land. The numbers and the framing come from a Microsoft Source feature by John Roach published on September 14, 2020, which documented the retrieval and the findings from what the company called Phase II of Project Natick.
The comparison matters more than the novelty. Plenty of hardware has been sunk in the ocean for science. Very little of it has been a working datacenter with a control group sitting in a building, running the same workloads, being counted the same way.
A cylinder, twelve racks, and a grid made of wind
The vessel had a name, the Northern Isles, and a shape borrowed from submarine engineering: a sealed steel tube, roughly the proportions of a shipping container, built with France’s Naval Group and its Naval Energies subsidiary, which brought experience in submarine construction and heat exchange to the design. Inside were 12 racks holding the 864 servers, plus associated infrastructure. The atmosphere inside was not air. Microsoft filled the cylinder with dry nitrogen, sealed it, and left it that way.
The site was the European Marine Energy Centre in Orkney, a test bed for tidal and wave power in some of the most energetic water in Europe. Green Marine, an Orkney firm, supported the deployment and later the retrieval. Heat left the servers the way it leaves a submarine, through pipes and a heat exchanger into cold seawater, which is a great deal more obliging as a coolant than air and requires no freshwater at all. The Microsoft Source feature notes that point directly: a datacenter that draws its cooling from the ocean is not drawing it from a municipal water supply.
Power came from the Orkney grid, which the feature describes as supplied 100 percent by renewables, a mix of wind, solar, and experimental marine energy technologies being tested on the islands. That combination, seafloor cooling plus a wind-fed grid, was part of the point. So was the isolation. Once the cylinder was sealed and lowered, nobody opened it. Anything that broke stayed broken for two years.
This was not the first attempt. Phase I, in 2015, put a smaller vessel in the Pacific for 105 days and established that the basic idea worked at all: that you could sink a datacenter, keep it dry, keep it cool, and keep it talking to the network. Phase II was the harder question, whether the approach was practical at something closer to useful scale, in genuinely rough water, for a length of time that would show up in a reliability statistic.
What came up off the seafloor
The retrieval happened in the summer of 2020. A ship pulled the cylinder up, and the first thing anyone saw was the biology: a coating of algae, barnacles, and sea anemones accumulated over two years on the seafloor. Crews power washed the hull, and researchers took their samples before the servers came out for analysis.
Then came the number. “Our failure rate in the water is one-eighth of what we see on land,” Ben Cutler, who leads Project Natick within Microsoft’s Special Projects research group, told Microsoft Source. The feature puts the same finding another way: the underwater servers were eight times more reliable than the land based replica Microsoft had run alongside them as a control.
The Microsoft team was careful about causation, and the coverage reflects that. The leading hypotheses in the feature are mundane rather than exotic. Nitrogen is far less corrosive to electronics than oxygen, and the sealed cylinder had no oxygen in it. Nobody was inside, which means nobody bumped a cable, jostled a drive sled, pulled the wrong component, or walked through with a cart. And the temperature was steady, because the ocean at 117 feet does not have afternoons. Each of those is a plausible contributor. Untangling which one carries the most weight is exactly the kind of question a controlled comparison is supposed to open up rather than close.
Worth saying plainly: published accounts of how many individual servers actually failed do not agree with each other, and the figure is not the durable finding here. The ratio is. One-eighth, against a land twin, over roughly two years.
What a sealed box suggests, and what it does not
The interesting implication is not that datacenters belong in the sea. It is that a lights-out, sealed, nitrogen-filled enclosure is a variable worth testing on dry land too. If the absence of oxygen and the absence of people account for much of the eight-fold gap, those are conditions a terrestrial facility could approximate without getting wet.
For the ocean itself, the logic Microsoft laid out is geographic. A large share of the world’s population lives near a coast. A cylinder that can be manufactured, shipped, and dropped offshore is a way to put compute close to those users, on a short network path, cooled by seawater and powered where possible by offshore or island renewables. Cutler frames Natick in the feature as a proven building block rather than a product with a rollout date, and that distinction is worth preserving. Nothing in the September 2020 account promises an underwater Azure region.
What it does report is narrower and more durable: 864 servers, 12 racks, 117 feet down off Orkney, two years sealed in nitrogen, failing at one-eighth the rate of their twin on shore.
