Building inference data centers on the high seas

A node is a steel structure ten to thirty meters across at the top, running seventy to a hundred meters down, which Kann likens to the height of Big Ben. It is generator, vehicle and data center at once. The hull shape forces water into a pressurized reservoir as the structure rises and falls, and that water falls back through a turbine on a mostly closed seawater circuit, which a colleague named ocean hydro: the aim is hydro’s cost structure without depending on a finite supply of rivers. A second feature of the same hull pushes water backwards, so the node always moves forward and only needs steering, like a Roomba you cannot stop. And because the design rests on cutting the cable to shore, the energy has to be consumed on board, which is why each node carries a computing cluster or an electrolyzer instead of exporting electricity. Wave energy stayed coastal historically because it assumed a cable, and usually a seafloor connection to push against. Removing both is what lets these go thousands of miles out.

Going out is the point, because that is where the resource is. Waves are wind energy accumulated over long fetches with very little loss, so they keep running after the wind stops, which he calls the world’s biggest solar battery. Coastal water gets a small, degraded fraction of it while competing with fishing and seafloor regulation. In the target regions he says a fifteen meter object intercepts about 2.5 megawatts of wave flux on average, with wave heights averaging four to four and a half meters across the year and rarely below three. Simulating against ten or eleven years of historical weather data, he gets capacity factors over 90% and, with two to four hours of battery, payload availability between 99 and 99.8%. The contrast he draws with solar is about shape rather than average: output has drawdowns lasting half a day, to perhaps 50% of the payload, and never reaches zero.

The most interesting move is economic. Panthalassa started out chasing a one-cent levelized cost of electricity and abandoned the goal, because with compute as the payload the node is only a fifth to a tenth of the total cost structure once you count replacing servers several times over its life. Making the node cheaper barely moves the answer, so the company now deliberately optimizes toward a more expensive node in exchange for higher uptime, more battery and shapes that can be stamped out of steel quickly. The power stays cheap anyway, at an optimum around three and a half to four cents per kilowatt-hour, but the claim that carries the argument is the other one: seawater provides free convective cooling, so the node replaces the power plant and the data center together, and that combined cost structure is what he says the comparison should be against.

Which leaves the objections, and Kann presses the right ones. Maintenance in open ocean is answered by designing for none, a solid steel hull with coatings and one water turbine on bearings as the only moving part. Kann points out that generators and power electronics fail on land regardless of the sea, and the response is a design philosophy rather than a result: an avionics-derived approach from people out of Raytheon and Collins Aerospace, analog logic with no firmware and no liquid capacitors. Sheldon-Coulson is explicit that maintenance-free operation is still a design goal and it remains to be seen whether they achieve it. On servers, the modeled answer is roughly one percent lower availability than land systems at the right deploy-and-recover cadence, and he argues failure rates may actually be lower at sea because the payload is sealed with nitrogen replacing oxygen, colder, and free of vibration and dust. The second answer is to change the workload, since much valuable work is CPU-heavy reinforcement learning and tool use, and newer accelerators use less of the most failure-prone memory. That also settles the market question: added satellite latency is about a hundred milliseconds, which vanishes into the time a human or an agent already waits, so the targets are long-running inference and reinforcement learning rather than tightly interconnected training clusters or anything genuinely latency-critical. The oddity Kann flags is the density inversion: a node’s economic optimum is around 400 kilowatts, roughly one rack in a hyperscale data center, inside a structure tens of meters wide and seventy deep. It works only because ocean area is effectively unconstrained.