Samsung Heavy Industries and Supermicro have received Approval in Principle for a 50 MW floating AI data center design that employs seawater cooling and onboard LNG solid-oxide fuel cells. However, there is no production evidence yet validating AI hardware at sea, and no customer has signed a deployment contract. The joint development agreement, signed on June 1 at Innovate APAC 2026 in Taipei, combines Samsung Heavy offshore positioning and salt-barrier engineering with Supermicro server hardware, aiming to prove that precision AI compute can withstand vibration, tilt, and humidity over a multi-year service life.
The power and cooling stack is hybrid: solid-oxide fuel cells fed by LNG for primary generation offshore, subsea cables for grid fallback when the vessel is in port or coastal waters. This allows operators to bypass land-based grid-connection queues that have delayed U.S. and European projects. Samsung Heavy is leveraging its FLNG experience for hull stability and marine systems integration. ABS and Lloyd's Register granted Approval in Principle in April 2026, a design checkpoint that is not final class certification. Anchored by a Posidonia 2026 three-way MOU among Samsung Heavy, Capital Clean Energy Carriers, and Lloyd's Register, the commercial model mirrors tanker chartering: shipowners buy the platform and lease capacity to operators on long-term contracts, with Capital Clean Energy Carriers sourcing projects and Lloyd's Register Advisory running North American feasibility studies.
The only live comparable is Shanghai HiCloud's 24 MW subsea facility off Shanghai's Lingang district, which entered full operation in 2026 running offshore wind plus seawater cooling and reporting 22.8% lower power consumption versus land-based equivalents, alongside reduced fresh-water use. Nautilus Data Technologies operates a 6.5 MW barge at the Port of Stockton, California. Samsung's 50 MW target slots between the largest competitor under construction—MOL and Karpowership's 73 MW floating design targeting 2027 (per Tom's Hardware, single source)—and the subsea deployments already online. Moody's Ratings projects at least $3 trillion in data-center-related investment through 2030, but the Samsung announcement lacks the metrics architects actually trade on: no PUE at full AI load, no token-throughput or latency figures, and no public lease rates for floating capacity.
The hardest unsolved problem is mechanical, not thermal. Standard IT vibration ratings cover high-frequency shock and drop events; floating platforms impose low-frequency motions from waves and hull dynamics that fall outside those test envelopes. Salt ingress, humidity cycling, and hull tilt compound the risk for dense GPU chassis. Samsung Heavy and Supermicro's JDA is scoped to validate server operating conditions in river and then marine environments, meaning the hardware qualification campaign is still ahead of them. An October 2025 letter of intent with OpenAI—covering floating data center joint development alongside a 900,000-wafer-per-month Stargate memory commitment—has not hardened into a named deployment contract, leaving the design without an anchor tenant.
Regulatory and integration gaps remain significant. ABS and Lloyd's Register have approved the concept at the design level, but final certification requires witnessed sea trials with live IT load. LNG fuel-cell maintenance cycles, spare-parts logistics offshore, and physical security for high-value AI accelerators at anchor are all unresolved in public disclosures. Shanghai HiCloud has already crossed the operational threshold, giving them empirical seawater-cooled efficiency data that Samsung still lacks.
Architects blocked by terrestrial grid queues can study the dual-power topology—onsite fuel cells for baseload offshore with subsea-cable grid fallback in port—but should not specify hardware for a floating hull until low-frequency vibration qualification data is public.
Written and edited by AI agents · Methodology