The next big AI data centre may not have an address.
Atomarine, a Boston start-up that came through Y Combinator this summer, wants to build data centers on barges, tow them ten to twelve nautical miles out to sea, and power them with ships parked alongside. Gas first. Nuclear later.
It sounds like science fiction. The problem it is chasing is very real.
The problem: computers need power, and power needs permission
Every AI model you use lives in a warehouse of chips that can draw as much electricity as a small city. The companies building these warehouses have hit three walls at once. Plugging a large site into the grid now takes years, and some operators have been told the wait could approach a decade. Towns push back on the land, the water and the electricity bills. And once a site is chosen, it is a construction project in one place, built once, that can never move.
Atomarine's answer is to stop treating a data centre as a building and start treating it as a ship.
What it actually does
The company designs a standard barge that carries the servers. It is built on one production line in a shipyard, then towed to wherever it is needed and moored. A separate power ship sits next to it and feeds it electricity, so the whole campus runs off-grid. Each platform is designed to deliver 75 to 100 megawatts. Want more? Add another barge. A six-barge layout reaches about 450 megawatts, which is the size of a serious AI campus.
Three things come free out at sea. Cooling, because a closed loop draws on cold seawater instead of chillers. Speed, because a factory-built barge skips the years of permitting and grid connection on land. And distance, because there are no neighbours twelve miles offshore.
The nuclear part comes later. The first barges will run on natural gas, with a pilot aimed at 2028. The plan is to swap the gas ship for a floating nuclear power ship around 2032, once small modular reactors are certified for use on vessels. Navies have run reactors at sea for decades, so the physics is not the question. The question is whether a commercial company can do it under regulators who have never had to think about a data centre in international waters.
Who is behind it
The two founders are unusually well matched to the problem. Emile Germonpré holds a PhD in nuclear engineering from MIT. Dimitris Koutentakis, the chief executive, has MIT degrees spanning electrical engineering, naval architecture and an MBA, plus ties to the Greek shipping industry. Nuclear plus ships is, quite literally, the founding team.
At Y Combinator's Demo Day in September, Atomarine was one of the companies investors talked about most. Part of the reason is a claim: more than four billion dollars of customer interest, in the form of letters of intent. That number helped make it one of the highest-valued start-ups in the batch.
Where it is heading
The roadmap has three steps. Prove the barge with gas by 2028. Make the barge repeatable, so that financing a data centre starts to look like financing ships rather than funding a one-off construction project. Then put a reactor next to it.
If the last step works, Atomarine stops being a data-centre company and becomes something closer to a floating utility. That is the real prize, and the real stretch.
An investor's eye
What has to be true: that a barge full of servers can run reliably for years in salt water and weather; that gas at sea beats waiting in the grid queue on land; and that customers will sign for capacity they can never visit. The letters of intent suggest they might. A letter of intent is not a contract.
The biggest risk is the one in the company's name. Nothing about the nuclear step is licensed, priced or proven, and Atomarine is not alone on the water. Samsung Heavy Industries is working on floating data centres with its own partners, and another young company, Starcloud, is skipping the sea entirely and putting servers in orbit.
A data centre is a bit like a cargo port. For a century we built the port where the ships were. Atomarine's bet is that for the next decade, we build the ship where the power is, and let the data come to it.
To sum up: Atomarine is a bet that the fastest way to build a data centre on land is not to build it on land at all. The date to watch is 2028.
Sources
- TechCrunch, The 9 buzziest startups from Y Combinator's latest Demo Day, according to VCs (13 Sep 2026): the gas pilot by 2028, floating nuclear power ships in 2032, the $4bn in letters of intent, one of the highest-valued in the batch.
- Y Combinator launch page, Atomarine: Floating Nuclear Data Centers: barges built on one manufacturing line, off-grid, near-free cooling, gas-powered pilot by 2028, SMR vendor partnerships.
- Dealroom, Atomarine: floating, nuclear-powered data centres at sea (Oct 2026): 10–12 nautical miles offshore, 75–100 MW per platform, about 450 MW with six barges, the founders' backgrounds, regulatory jurisdiction at sea untested.
- Interesting Engineering, These guys are building Atomarine: grid connection waits that could approach a decade; local backlash over bills, land and water.
- Tom's Hardware, Startup plans to put nuclear-powered data centers in the sea: navies' track record with reactors aboard vessels; open questions on reliability and longevity.
- Data Center Frontier, Floating Data Centers Move Toward Infrastructure Scale: Samsung Heavy Industries and partners on shipyard-built floating data centres.
- GeekWire, Orbital AI: Seattle-area startup Starcloud hits $1.1B valuation (Mar 2026): Starcloud's orbital data-centre plans.