Infrastructure
They're Putting Data Centers in Space. We're Putting Them in the Sea.
A billion-dollar AI data center in Kenya just stalled. Not over money. The grid simply could not power it. That is not a Kenyan problem. It is the wall the entire AI industry is about to hit, and the escape route almost nobody is looking at.
In May 2024, Microsoft and the Abu Dhabi firm G42 announced a one-billion-dollar data center for Kenya's Rift Valley, geothermal-powered, running Azure for a new East Africa cloud region. It was the kind of announcement that makes a continent feel it is finally getting a seat at the AI table.
Two years later the project is in limbo. The reason is brutally simple. Kenya's entire national grid runs at roughly 3,000 megawatts. The data center's long-term target was a full gigawatt, about a third of everything the country generates. President William Ruto put it plainly at a state event: to keep the facility running, Kenya would have to "switch off half the country."
No one ran out of money. No one ran out of ambition. They ran out of electrons. And that is the part worth sitting with, because Kenya is not the exception. Kenya is the early warning.
The wall is real, and it is made of three things
The story the headlines miss is that AI's growth is no longer limited by chips, capital, or land. It is limited by power, and increasingly by water, and the constraint is hardening every month.
The numbers are not subtle. In the United States, new high-capacity grid connections in the big data center hubs now face four to seven year waits. The PJM interconnection queue alone holds over 2,600 gigawatts of pending requests, with average waits past five years. Even the boring hardware has become a bottleneck: the lead time for a large power transformer now stretches to around 210 weeks. That is four years to buy a part.
Then there is water, the constraint that gets the least attention and causes the most local anger. A single medium-sized data center can drink about 110 million gallons a year for cooling. Multiply that across a global build-out and the figure climbs toward a trillion liters a year by 2030. In a lot of the places these centers want to go, that water is already spoken for.
Power you cannot get. Water you cannot spare. Land you have to fight for. That is the wall. And the AI boom is driving everyone straight into it at speed.
Name the villain honestly
The villain here is not a company or a person. It is structural. It is grid-and-land dependence itself. The entire industry built its growth on the assumption that you can plug a building into the grid, draw on the local water, and scale. AI broke that assumption, and Kenya is the body it left behind.
That is the honest frame. Not "data centers are evil," which is a moral panic and false. The truth is colder and more useful: the model of computing we inherited has hit a physical ceiling, and the projects dying on that ceiling are real, recent, and expensive.
So the industry is fighting the wall harder
Faced with a power wall, the richest companies are doing two things, and both are revealing.
The first is nuclear. Amazon has locked up 960 megawatts for a Pennsylvania campus. Microsoft revived a twenty-year deal to restart a reactor at Three Mile Island for 837 megawatts. Oracle has floated a gigawatt-scale data center fed by three small modular reactors. The pitch is "AI-ready baseload," and on paper it is clean and it is firm.
But look closer and the strategy quietly stops scaling. A small modular reactor is not a product you ship to a hundred countries. It is fuel you have to secure, waste you have to bury, and a license that turns the site into a national-security asset. Survey after survey shows people do not want a reactor in their backyard, which is exactly why most countries are not building new ones. So nuclear does not scale to the world. It scales to a few wealthy firms in a few permissive jurisdictions. For Kenya, or Nigeria, or most of the planet, "just add a reactor" is not an answer.
The second answer is even further away. Several teams are now seriously pitching data centers in orbit, beaming the heat into space. Everyone is looking up.
Google, to its credit, found a smarter terrestrial path. It signed a 115-megawatt deal with Fervo Energy for enhanced geothermal, hot rock as firm clean power, built on a pilot that worked. It is genuinely good engineering. But it is also geology-dependent and, for now, largely a Google story. It does not generalize to every coastline that needs compute.
So the field looks up at space and digs down for reactors. And almost nobody turns the other way.
Almost nobody is looking at the sea
Which is strange, because the sea answers the exact three constraints that land cannot.
Cooling is free and effectively infinite, because the water is right there and it is cold and it stays cold. Land use drops to near zero, because the hall sits offshore. And paired with its own dedicated clean power, the whole thing comes off the grid entirely. No five-year interconnection queue. No fight over the town's water. No reactor anyone has to approve.
This is not a thought experiment. Microsoft already proved the physics. Project Natick sank a sealed data center off the Orkney Islands and left it for two years. The result was startling: of 855 servers underwater, only six failed, a 0.7 percent failure rate against 5.9 percent for the identical servers on land. Eight times more reliable. It used zero fresh water and ran at a power usage effectiveness of 1.07, far better than the industry's 1.67 average.
It worked. And then, in June 2024, Microsoft shelved it. Two reasons. The pods were sealed, so you could not get inside to service the hardware or swap in new AI chips. And they ran off the shore's grid, so they never escaped the very constraint that matters most.
Two gaps. That is all that stood between a proven idea and a deployable one.
Closing the two gaps
That is the work behind Thermosea. Submerged halls you can actually enter, service, and upgrade, cooled directly by the sea, and powered by their own geothermal plant instead of a shore grid that has nothing to spare. The geothermal choice is deliberate. It is the kind designed to avoid the induced earthquakes that have shadowed some enhanced-geothermal drilling, so the power source does not trade the grid fight for a new one with the people living nearby. Solve serviceability, solve power, the two things that made Microsoft walk away, and everything Natick proved still holds.
I want to be precise about where this stands, because the honest version is the stronger one. The sea is already humming with servers. That part is real and happening now. But most of what is going into the water repeats the exact mistakes that shelved Natick: sealed pods nobody can service, still tethered to the shore grid they were built to escape. People are deploying the dead end and calling it the future. Our servers are not in the sea yet. They will be soon enough, and when they go in, they go in without the two flaws that sank the last attempt. We are at feasibility, and by my numbers it gets cheaper than land-based compute over time, because the two most expensive and most contested inputs, power and cooling, stop being line items you fight for and become conditions you sit inside.
Why this matters most for Africa
Here is why this is not a side note for the continent. It is the center of it.
The places that most need sovereign compute, that cannot keep shipping their citizens' data offshore, are often the places the grid locks out first. Kenya did not lack vision. It lacked 1,000 spare megawatts, and there is no version of the next ten years where a fast-growing African grid casually frees up a third of its capacity for one building.
So the continent has two roads. It can wait decades for grids to catch up, and keep renting cloud capacity from someone else's country in the meantime. Or it can leapfrog, the way mobile money leapfrogged bank branches, by adopting the model that does not depend on the infrastructure it does not have. A data center that brings its own power and its own cooling and needs almost no land is not a luxury for a coastal African nation. It is the most logical thing in the world.
There is a sovereignty layer to this too, and it is the part African governments should be thinking about hardest. When your data lives in someone else's country because that is the only place with the power to host it, you do not fully control it. You rent access to your own information, under another nation's laws, subject to another nation's outages and politics. A coastal country that can sink its own off-grid compute into the water next to its own cities does not have that problem. It owns the building, the power, and the data, all at once. For a continent with thousands of kilometers of coastline and a hard ceiling on grid power, that is not a small advantage. It is the difference between hosting the AI era and renting a seat at it.
The sea does not care how strong your grid is. That is the entire point.
The honest question
So I will end with the question I keep asking, and I mean it as a real one.
The most sophisticated companies on earth are staring at space and at nuclear reactors to solve the AI data center crisis. Both are expensive, both are slow, and only one of them scales past a handful of rich firms. Meanwhile the sea sits there, cheaper, colder, already off the grid, with a Microsoft experiment on record proving it works.
Why is almost nobody looking down?
Tell me what I am missing. Or tell me I am right. Either way, I think we are trying to cool the future in the wrong place.
Frequently asked questions
Why are AI data centers running out of power? AI compute grew faster than grids could expand. New grid connections in major hubs now wait four to seven years, and the US PJM queue alone holds over 2,600 gigawatts of pending requests. Power, not land or capital, is now the main limit on data center growth.
What happened to the Microsoft data center in Kenya? The one-billion-dollar Microsoft and G42 project in Kenya stalled because the national grid, around 3,000 megawatts total, could not supply the gigawatt the facility ultimately needed. Kenya's president said powering it would mean "switching off half the country." Talks also broke down over capacity-payment guarantees.
Do underwater data centers actually work? Yes, by the available evidence. Microsoft's Project Natick ran 855 servers on the seabed for two years with one-eighth the failure rate of land servers, zero fresh water use, and a power efficiency well above the industry average. Microsoft shelved it because the sealed pods could not be serviced or upgraded, not because the concept failed.
Why did Microsoft cancel Project Natick? Two reasons: the pods were sealed shut, making it impossible to repair hardware or install newer AI chips, and they drew power from the shore grid. Those are engineering and siting gaps, not failures of the underwater idea itself.
Are nuclear small modular reactors the answer for AI power? For a few large companies in permissive countries, possibly. But reactors require secured fuel, buried waste, and national-security-grade licensing, and most populations resist hosting them. That makes SMRs hard to replicate worldwide, which is why they are not a general solution for most of the planet.