An article by Hannah Hawthorne-Briscoe, Chief Operating Officer at SubSea Cloud, around the five myths of underwater data centres.
In the summer of 2020, a crane on a barge off the Orkney coast lifted a white steel cylinder the
length of a shipping container out of the North Sea, and everyone present stopped to look at it.
It had gone down two years earlier, clean and bare. It came up wearing the sea. Algae, barnacles
and sea anemones covered the hull so thickly that the engineers had to hose it down before they
could read the markings. Inside, sealed away from all of it in dry nitrogen, sat hundreds of
ordinary servers, the same machines you would find in any data centre in Slough or Ashburn,
and the reason the barge was there at all was to find out how many of them had died.
The exercise had a control group. Microsoft had run identical servers, from the same batch, in a
conventional facility on land, where technicians could walk between the racks and check on
them daily. The underwater machines had been unreachable for two years. No engineer had laid
a hand on them. If anything had gone wrong, nothing could have been done about it.
The underwater servers failed at roughly one-eighth the rate of the ones on land.
That result has been public for years, and I have come to think it is the most ignored number in
data centre engineering. Not disputed. Ignored, which is different and more interesting. When I
describe it to people in this industry, the reaction is rarely disagreement. It is a kind of polite
stall, followed within a sentence or two by a question about maintenance.
The stall is the thing worth examining, because it is where every objection to underwater
computing begins.
We are custodial animals. Watch how anyone treats something valuable and fragile: they keep it
close, they check on it, they build a room around it with a door they can walk through. The entire
architecture of the data centre industry expresses that instinct in concrete. Raised floors, cold
aisles, badge readers, an engineer within a few minutes’ walk of every rack at all times. The
unexamined premise underneath it is that access equals safety, that a machine you can reach is a
machine you can protect.
The Orkney number suggests the premise runs the other way, and once you accept that, the
standard objections start to look less like technical problems and more like symptoms.
Take the first one, the one that arrives within a sentence of the stall: you cannot maintain it. This
is true, and the truth in it matters. Nobody swims down with a screwdriver. What the model does
instead is redundancy inside the pod, continuous remote telemetry that notices degradation long
before a human walking past would, and a scheduled cycle of roughly four years when the pod is
recovered, its hardware renewed and the unit returned to the seabed. But notice what the
objection assumes: that the ability to intervene is a form of protection. In Orkney it was the
intervention that was absent, along with the dust, the humidity swings, the thermal cycling and
the human beings, and the machines were better for all of it. Maintenance access is not only a
safeguard. It is also an opening, and openings have costs that nobody puts on the balance sheet.
Then there is the objection that sounds like worldliness: if this worked, Microsoft would have
commercialised it. Natick was research and it produced research, which is what it was for.
Microsoft has never published a full account of its commercial reasoning and I am not going to
invent one on its behalf. What I would observe is that a software company declining to become a
marine infrastructure operator tells you something about that company’s shape rather than about
the physics. The gap has since been filled by others. China has a commercial underwater facility
running off Shanghai. Named deployment projects are underway elsewhere, including in our
own industry.
The environmental objections take a different form, and they deserve the most careful handling,
because the sea is worth protecting and the burden of proof belongs with the people putting
things in it. The charge that a pod heats the ocean runs into a piece of physics that reframes it:
computing produces the same heat wherever it runs. On land, that heat is removed by chillers and
fans and, in most large facilities, by evaporating drinking-quality water, all of which consumes
energy on top of the computing and releases the heat regardless. On the seabed, moving water
carries the same heat away for nothing. Our own published worst case, which assumes the water
around a one-megawatt pod sits perfectly still for an hour and absorbs every joule, comes to
roughly four tenths of a degree across a small patch of seabed, inside the range coastal water
varies by between morning and afternoon on its own. Still water does not exist in the sea, so
working sites do better than that.
The related worry, that warm seas cannot do the cooling at all, dissolves against the temperature
gap. Server processors run at around 80°C. The Gulf, the warmest sea on Earth, has never
reliably recorded water above roughly 38°C, and the pods are validated well beyond that. The
margin holds in August in Doha, which settles the question for every colder ocean there is.
And the last one, that pods damage marine life, has an answer sitting in plain sight on the barge
at Orkney. Hard structures on the seabed become habitat; the effect is well documented around
wrecks, piers and platforms. What responsible deployment adds is care about where: bare sand or
soft sediment, away from seagrass, reef and protected habitat, with stainless hulls, a
biodegradable coolant chosen so that even the failure case is benign and which has never leaked
in this technology’s operating history, and environmental assessment supported wherever a
country requires it.
None of which makes this settled technology. The operating history is short. The regulatory
pathway for seabed computing is still being built, country by country, and anyone who tells you
otherwise is selling something. Pods will not replace gigawatt campuses and are not trying to;
they are a coastal, fast-deploying, water-free layer for the workloads that suit one.
But go back to the barge for a moment, and to that cylinder streaming seawater in the Orkney
sun, so thickly colonised that the engineers had to wash it before they could read it. The instinct
in the room was that the sea was what the machines had survived. The data said something closer
to the opposite. The sea had been keeping them: cold, steady, uninterrupted, and entirely uninterested in walking between the racks.

Leave a Reply