Panthalassa

Microsoft already proved the hard part: a sealed server cylinder dropped on the seafloor off Orkney ran for two years with a fraction of the failure rate of servers on land, cooled for free by cold seawater. But it still drew its electricity through a cable to shore. Panthalassa's pitch is the part Microsoft left out. Its roughly 85-meter steel node — about the height of Big Ben, mostly underwater — uses the rise and fall of waves to push water through an internal turbine, so the same structure makes the power, cools the chips, and beams the results back up by satellite. Cut the cable and you cut the grid interconnect, the land permit, and the freshwater all at once — the four things currently bottlenecking onshore AI buildout.

A kilowatt in the wave is not a kilowatt at the GPU — one analyst sees a deployed node losing half its useful output within a year, fouled and drifting, before anyone services it.

The catch is that wave energy is a startup graveyard. Devices have been bankrupting their backers since the 1990s, for a brutal reason: the ocean rarely breaks a machine outright, it just grinds its efficiency down — fouled grilles, cooling derate, drifting moorings — until the economics stop penciling. One analyst projects a deployed node sliding to roughly half its useful output within a year without major servicing. A node with even a trickle of residual drift wanders kilometers a day, so a 'stationary' data center quietly migrates.

Which is why the real bet here isn't that the engineering works on the first try. It's that pairing the generator with a high-value AI compute load — inference, not training — finally makes wave power worth enough to survive the failures and iterate. Thiel's $140M is buying iterations that earlier wave startups never could afford. Nothing is in the water yet that you could rent: a prototype is slated for the northern Pacific later in 2026, with commercial units promised in 2027.