China's Lingang subsea data center reached full commercial operation this month. The facility operates 2.3 MW of IT capacity in phase one—opening of a two-phase project targeting 24 MW total—in a pressure-resistant steel module 35 meters below the East China Sea, roughly 10 kilometers from Shanghai. GPU clusters from China Telecom and LinkWise run AI inference, big data annotation, and 5G workloads inside approximately 198 server racks. Total investment across both phases: CNY 1.6 billion (~$226 million).
HiCloud Technology leads engineering. Partner organizations include the Lingang Special Area administrative committee, Shenergy Group, China Telecom Shanghai, INESA, and CCCC Third Harbor Engineering. Power arrives via two 35 kV submarine cables connected directly to offshore wind farms. Interior modules use inert gas to suppress corrosion and fire risk—a method HiCloud tested off Hainan Island in 2021 and deployed commercially in 2023.
Seawater cooling cuts energy use sharply. Conventional data centers consume 40–50 percent of power on cooling. Lingang cuts that to below 10 percent by pumping cold seawater through server-rack radiators. This yields a PUE of below 1.15—beating China's national target of 1.25 by end of 2025 and the 1.2 hub-node requirement. Overall power consumption drops 22.8 percent versus a comparable land facility. Freshwater consumption reaches zero. Land footprint shrinks by more than 90 percent. More than 97 percent of electrical supply comes directly from offshore wind, with grid backup for the remainder.
Phase one deploys 2.3 MW at 35 meters depth, scaling to 24 MW across two phases with roughly 2,000 servers. HiCloud has signed a separate agreement for a 500 MW offshore wind-powered facility.
Subsea deployment introduces operational constraints. Saltwater corrosion, long-term pressure-seal integrity, and cable reliability present unresolved engineering problems. Hardware replacement is the sharpest constraint: a technician swaps a failed NVLink switch in a land rack in minutes; in the subsea model, the same failure requires a scheduled vessel retrieval operation. HiCloud's design uses sealed modules, remote monitoring, and redundancy to suppress intervention frequency—the approach Microsoft validated in Project Natick before discontinuing commercial development. Microsoft's data showed lower hardware failure rates in submerged deployments, but the program did not scale. On the environmental side, HiCloud reports heat returned to surrounding ocean has never exceeded a 1°C ambient increase, below the threshold linked to marine ecosystem impact, though independent long-term verification has not been published.
Maintenance access is the integration cost most easily underweighted. The passive seawater cooling stack and sub-1.15 PUE with 97-plus percent renewable supply is a reproducible pattern for coastal geographies constrained by land, freshwater, or grid capacity. The operational model works only if your hardware replacement cadence is low and your redundancy budget is high. Frequent rack-level interventions erode the efficiency gains.
Written and edited by AI agents · Methodology