183 towns said no, so AI compute is heading offshore, underground, and into orbit

On July 18, 142 protests against data center development were held across 42 states, and 183 U.S. towns now have moratoriums or bans on the facilities, according to the U.S. Data Center Moratorium Tracker ^{[1] [2]}. That is 84 percent of states with at least one protest on a single day ^[1], and the response from developers has been displacement rather than delay: into the open ocean, into disused mines, and into orbit ^[3].

The sentiment behind the moratoriums is not marginal. A Gallup poll in March 2026 found 70 percent of Americans surveyed oppose construction of AI data centers in their local areas ^[4]. The objections are the ones that surface in permit hearings: electricity demand, water demand, noise, and job creation promises that went unfulfilled ^[5].

So the siting maps are being redrawn toward places with no electorate. Portland-based Panthalassa is building an orb-shaped pilot, Ocean-3, to float hundreds of miles into the Pacific, converting waves to electricity with pressurized seawater driving an internal turbine and using the surrounding ocean for cooling ^{[6] [7]}. It has raised $225 million at a valuation near $2 billion before launching ^[8]. China's Shanghai Lingang undersea facility sits six miles off the coast and is wind-powered; the Chinese government says it runs on 95 percent green electricity, cuts power consumption 22.8 percent, and eliminates water use entirely ^{[9] [10]}. It had drawn $226 million in investment as of October 2025 ^[11].

Orbit is the more crowded pitch. Starcloud launched Starcloud-1 in November 2025 and says it was the first company to train an LLM in space, using solar power and the vacuum for heat rejection ^{[12] [13]}. The company claims ten times lower energy costs even after launch expenses, and plans a second satellite in October 2026 with 100 times the power generation of the first ^{[14] [15]}. Axiom Space put low-Earth-orbit data centers up in January 2026 on Kepler satellites ^[16]. SpaceX has filed with the FCC for a constellation of up to 1 million satellites forming a solar-powered data center, Blue Origin filed in March for more than 50,000 data centers in space, and Google is studying the idea as Project Suncatcher ^{[17] [18] [19]}.

For anyone sizing compute in the next two years, the numbers matter more than the physics. The only site here with a stated capacity figure is Norway's Lefdal Mine Datacenter, at 80 megawatts, according to Data Center Dynamics ^[20]. That is a converted mine 197 feet below sea level and more than 2,000 feet inside a mountain, running since 2017, roughly nine years ^{[21] [2]}. Everything else is a pilot, a prototype, or an FCC filing. Starcloud's own gap between first and second launch is 11 months ^[3]. Substituting orbital or mid-Pacific capacity into a 2027 plan is not supported by any of this.

The two constraints that do not negotiate are distance and hands. A node hundreds of miles offshore or in low-Earth orbit has a latency floor set by geography, and it cannot be serviced by driving to it. Lefdal is the instructive case precisely because it is boring: hydro and wind power, fjord-water cooling, carbon neutral, heated process water earmarked for a local salmon hatchery, and, as the operator told DCD in December 2025, not a single complaint from locals ^{[22] [23] [24]}.

Watch whether the moratorium count keeps climbing past 183, whether Starcloud's October 2026 satellite ships on time and holds its 100x claim, how the FCC handles a filing for up to 1 million satellites, and whether anyone copies Lefdal's actual trick of being invisible and locally useful on land.