The Sea Eats Space: Why floating ocean data centers beat orbital compute in 2026 is a 24-page engineering working paper that compares two competing architectures for next-generation AI data center capacity: anchored floating ocean platforms (so-called Sea Stations) cooled by seawater and powered by offshore renewables, and low-Earth-orbit satellite data centers of the kind being demonstrated today by Starcloud and Axiom Space. Across seven engineering dimensions — capital cost, operating cost (PUE), internet speed (RTT), reliability, density, schedule, and risk — the paper finds that floating ocean data centers win 6 of 7 dimensions on a 36-month decision horizon, with the seventh (throughput to the user) tied or won by LEO on aggregate mesh. The analysis is grounded in 213 verified sources including the Voyager 10-K filing, Google's Project Suncatcher preprint, and the June 2024 cancelation of Microsoft's Project Natick, the single most important Tier-1 caveat in the entire document. Written in handbook style for engineers, executives, and policy people rather than academics, the paper defines every technical term (from GPU to UNCLOS to API RP 2A-WSD) before using it, tags every quantitative claim with confidence tiers (T1 = real and running; T4 = aspirational), and walks the reader through the Sea Station architecture, the seven engineering matchups, a top-10 risk register for both sea and space, six honest counter-arguments, and a 24-month action plan for operators and investors. The central honesty clause is that this window is structural, not permanent: if SpaceX achieves its $200/kg launch-cost target, orbital economics close the gap, and the paper's "sea wins" verdict inverts into "depends on the workload" in the late 2030s. For 2026, with Microsoft Natick's commercial cancelation as the cautionary precedent and Tier-1 cooling PUE of 1.05 demonstrated at small scale, the answer is clear: build the floating thing.