The proof of concept off Shanghai proves the engineering works. Whether it tells us anything about AI infrastructure, and whether anyone has measured what it does to the ocean, are different questions.
Last week, China switched on what state media and the operator are calling the world's first commercial underwater data center. The facility, built roughly six miles off the coast of Shanghai's Lin-gang Special Area by HiCloud Technology and China Communications Construction, sits about 30 feet below the surface of the East China Sea, draws power from a nearby offshore wind farm, and uses passive submersion in ocean water for cooling. The numbers being floated by Gizmodo, which first reported the commissioning, are striking: a roughly $226 million price tag and, per a Chinese government release published last October, power consumption roughly one-fifth that of a comparable land-based facility.
The engineering is real. The framing is not.
The facility's maximum capacity is 24 megawatts. For context, that is about what a mid-sized suburban office park draws at peak. The gigawatt AI campuses being built by Microsoft, OpenAI, and a handful of Chinese hyperscalers run 40 to 80 times larger. Anyone citing this as a blueprint for the next generation of AI infrastructure is comparing a proof-of-concept to the actual problem.
The efficiency claim, the one that gets the headlines, deserves its own scrutiny. The "one-fifth" power figure comes from Chinese government and operator sources, and the methodology behind it is not laid out in detail in the English-language announcement. The mechanism itself is plausible: cold deep water is a far more effective heat sink than air-cooled servers, and pairing the load directly with offshore wind eliminates transmission losses and lets the operator claim a lower-carbon footprint per compute hour. None of that is in dispute. What is in dispute is whether the savings generalize to a facility large enough to matter, in a marine environment capable of absorbing the heat, and at a cost that survives a real procurement.
That last point is where the science gets uncomfortable. Peer-reviewed work published in Nature Scientific Reports warned that submerged data centers risk creating localized thermal plumes, deoxygenated water, and conditions favorable to harmful algal blooms. A similar proposal off San Francisco Bay, the NetworkOcean project, ran into organized opposition from marine scientists and regulators over exactly these concerns. The Chinese project, according to the available reporting, has not yet published thermal-effluent monitoring data or third-party ecosystem assessments. "It works" is not the same as "it works without ecological cost."
This matters because the macro picture is not subtle. The United Nations University Institute for Water, Environment and Health projected in its 2026 report on the environmental cost of AI that data-center water demand could reach 9.3 trillion liters per year by 2030, roughly the current total water demand of sub-Saharan Africa. Even if China's underwater experiment is a clean win on every dimension, the most effective decarbonization lever is still the one the same body of research flags: not building data centers at the scale the AI buildout currently projects. Twenty-four megawatts of submerged compute, working perfectly, is a rounding error against that trajectory.
There is a real story here, and it is not "China flips the switch first." It is an unverified power-savings claim, an unmodeled marine-thermal trade-off, and a proof-of-concept three orders of magnitude too small to test the thesis anyone actually cares about. The facility is on. The hard part is still ahead.