Float the Data Center, Swap the Reactor Later

The bottleneck for new AI compute is no longer chips. It is the four-to-seven-year wait for grid interconnection, a gap measured in years while new accelerators ship in months. A Y Combinator-backed startup, Atomarine, is pitching a structural answer: shipyard-built platforms that separate hull, power vessel, and cooling loop into swappable modules.

The company pegs a first barge cluster at 75 MW each, scalable to 450 MW, with company-stated roughly 1.1 PUE seawater cooling (PUE being the industry shorthand for how much energy a data center spends on itself rather than on compute) and a phased power architecture: first power vessels run on natural gas, then a compact marine reactor takes the same berth. The 1.5 GW-per-year cadence, 20–40-year life, and 4–5× faster deployment claims are Atomarine's modeling, not deployed numbers, a distinction HN commenters in thread 49108039 pressed, especially on maritime-nuclear regulation and cross-waters sovereignty.

Atomarine pegs demand as a 35–49 GW U.S. firm-power shortfall by 2030. The adjacent Seatrium 30 MW offshore data center pilot (a separate entity) is parallel, not corroborating. No reactor vendor, shipbuilder, utility, or customer is on record.