Floating and subsea data centers: what is actually operating in 2026

The distinction between using water and consuming it is the real water-use argument for marine siting, and it is genuinely strong. But the constraint that kills projects is power. Autonomous floating facilities need enormous onboard generation, fuel storage and redundancy, which is precisely what ended the Nautilus Maine project.

Elliott Bay’s demand data is the best public sizing available: baseline inquiries cluster at 30 MW, most fall in the 50 to 80 MW range, and AI-focused inquiries push 100 to 300 MW — well beyond anything built anywhere.

A United States floating data center without passengers, tethered to shore, is not a Coast Guard-inspected vessel. Most owners instead pursue voluntary class certification through ABS or DNV, and both societies are actively developing rule sets specific to floating data centers. The requirements resemble offshore infrastructure: periodic inspection, maintenance planning and lifecycle management.

On the Jones Act, this is analysis rather than settled law. No data-center-specific Customs and Border Protection ruling or litigation exists. The analytically correct framing is that 46 U.S.C. section 55102 restricts transportation of merchandise between United States points to vessels that are US-built, US-owned and US-crewed, so a permanently moored, shore-tethered facility that transports no merchandise sits largely outside it, while repositioning a barge between US ports, or supplying it from US points, creates coastwise exposure. The practical bite is that a US-built hull costs a multiple of a foreign one. Take maritime counsel before relying on any definitive claim, including this one.

Why the geography is lopsided: marine permitting and environmental scrutiny, particularly of thermal discharge into sensitive ecosystems, are materially more stringent in Western jurisdictions than in China. That, more than technology, explains why the only operating capacity is Chinese.

Aikido Technologies has the most interesting design in the field: a floating offshore wind platform with integrated compute, sized at a 15 to 18 MW turbine per unit supporting roughly 10 to 12 MW of AI-grade compute in three data halls of 3 to 4 MW each, with battery storage and seawater cooling. A 100 kW proof of concept is planned in Norway later in 2026, with United Kingdom commercial deployment targeted for 2028.

Yokohama Port in Japan began a demonstration on 25 March 2026 at Osanbashi Pier, running through the end of Japan’s 2026 fiscal year — a containerized data center on a floating mini-platform with solar and battery at 100% renewable, with Yokohama City, NYK Line, NTT Facilities, Eurus Energy Holdings and MUFG Bank. It is explicitly testing salt-damage resistance, vibration-related operational stability and renewable energy management, the three unresolved engineering questions.

Mitsui O.S.K. Lines, Hitachi and Hitachi Systems signed a memorandum of understanding in March 2026 to develop floating data centers converted from second-hand vessels, a Karpowership-style model, with operations possible from 2027.

Microsoft’s Project Natick was discontinued in 2024, leaving one finding that still matters: submerged servers failed at one-eighth the rate of the land-based control group.

Starcloud raised a $170 million Series A at a $1.1 billion valuation in March 2026, led by Benchmark and EQT Ventures, and launched Starcloud-1 in November 2025 with an Nvidia H100, claimed as the first terrestrial-grade GPU operated and trained in orbit. Starcloud-3 is planned as a 200 kW, three-tonne spacecraft for Starship, and the company has filed a Federal Communications Commission application for an 88,000-satellite constellation, targeting roughly five cents per kilowatt-hour once Starship flies regularly at $500 per kilogram.

Google’s Project Suncatcher, announced 4 November 2025, will launch two Trillium-generation TPU satellites with Planet Labs by early 2027, toward a theoretical 81-satellite cluster in a one-kilometre radius at 650 kilometres altitude, economic only at roughly $200 per kilogram launch cost by 2035. Google names its own obstacles: thermal management, on-orbit reliability and high-bandwidth-memory radiation sensitivity. Lonestar Data Holdings flew an 8-terabyte solid-state drive and a Microchip PolarFire FPGA to the lunar surface on Intuitive Machines’ Athena mission, targeting off-world disaster recovery.

The honest assessment: total operating space-based compute is a handful of GPUs. Against 7,481 MW under construction in United States colocation alone, space is a technology-watch item, not a capacity source this decade, and Starcloud’s economics depend on a Starship launch price that does not yet exist.

Two reasons, both concrete. First, the Keppel ratio — 25 MW on 9,870 square metres of land plus 7,580 square metres of sea space — is a real number for port authorities and waterfront owners evaluating whether a berth, a slip or a filled parcel has a compute use. Second, the pattern across Shanghai, Singapore, Stockton and Yokohama is that marine compute needs landside real estate anyway: substation, switchgear, fiber landing, maintenance and staging. The waterfront parcel behind the mooring is the transaction.