Shanghai Hailanyun Technology brought its offshore-wind-powered underwater data center to full operation. Tom's Hardware's report captures the numbers: 2,000 servers, 24 megawatts of capacity, ocean water for passive cooling, offshore wind for power, and a PUE reported below 1.15 against an industry average around 1.5.
A PUE below 1.15 at meaningful scale is the part of the announcement that should make Western hyperscale facility teams stop and reread. Industry-average PUE sits near 1.5. Best-in-class land facilities with extensive evaporative cooling and tight mechanical controls land around 1.2 to 1.3. A subsea facility that posts under 1.15 has effectively removed the chiller plant from the energy stack, because the rejection medium is ocean water and the cost of moving it is largely the cost of the seawater path itself. That collapses the largest non-IT load in a conventional data center.
Microsoft's Project Natick demonstrated that immersed enclosures can be reliable. What Hailanyun has done is bolt the cooling architecture to its power source. Offshore wind generation sits next to the cooling sink. Transmission losses drop because the generation and load are colocated. Cooling water arrives at the servers at near-ambient sea temperature without a chiller. Land permitting disappears because the facility is below the waterline. None of these are individually new. The integration is what makes the facility a reference design.
For operators planning around the seawater cooling alternative, the question stops being whether subsea works and starts being what the practical scale ceiling is. 24 MW on a single pod is a credible unit. Several pods at a wind farm scale a campus into the hundreds of megawatts without touching fresh water.
The serviceability question is what kept Western hyperscalers off this architecture. If a node fails inside the pod, the operator cannot send a technician down with a cart. The Hailanyun design handles that by running the pod as a sealed unit with redundant servers and accepting some attrition over the deployment life rather than swapping out individual boards. That works for inference fleets where node-level redundancy is cheap. It works less well for tightly coupled training clusters where a single failed node stalls a training run.
The other constraint is bandwidth. A subsea facility connects to the grid above water through cables that have to land somewhere. The same goes for the network. For inference workloads that respond to user queries, the latency budget is forgiving. For training, the inter-node networking inside the pod has to be as fast as a land facility and the connection back to model orchestration has to be fast enough that the rest of the training fleet does not wait.
The Hailanyun facility lands at a moment when Western water and land permitting for AI buildouts is getting harder, not easier. Operators facing the 7 billion gallons of fresh-water draw in the American West and the community pushback in Pennsylvania have a working reference design to study. Whether it transfers to US coastal waters is a permitting question, not an engineering one. The engineering side is now demonstrated.