Follow the compute, and follow the money. On land, the data center buildout is a $400 billion capex machine in 2025, up 75% in 2026, running toward as much as $6.7 trillion by 2030. But watch where the new money points: offshore. Google's Project Suncatcher is putting AI chips in orbit, Musk wants a space-based AI data center, Anthropic held orbital-compute talks with SpaceX, and Starcloud raised $170 million to become Y Combinator's fastest unicorn ever. On the water: Peter Thiel put $140 million into Panthalassa, China switched on an undersea data center in May 2026, and Core Power builds a 300-megawatt floating nuclear plant for the Pentagon by 2028.
The sizes: land takes the trillions, up to $6.7 trillion by 2030. Orbit grows 67% a year toward $39 billion by 2035. And the ocean, where no analyst dares to publish a number yet, I size myself at $20 to 40 billion by 2035, because water beats vacuum: free cooling, no launch costs, and a technician can actually reach it. The growth rates offshore are the prediction, not today's size.
Why are they fleeing? On land they face the backlash: protests, permits, and voters blocking data centers, my full call is card 03. On land they face the buckling power grid, card 30. And on land they face the law: taxes, regulators, courts. Offshore, none of that reaches them. I predict that by 2030 there are 2 AIs: a public AI on land, regulated and contested, and a sovereign AI at sea and in orbit that no law reaches and no switch turns off. Once that split locks in, AI as a public good is dead.
One critical note on the space dream: physics does not care about pitch decks. Cooling, radiation, repairs, and launch costs make orbit the hardest place to run computers. Low orbit is real and coming. The ocean arrives first.
Early call
