AI for Energy and China’s Underwater Data Center
24 megawatts is the headline number in Shanghai’s new underwater facility, and that single figure says a lot about where AI for energy is headed. The project, built by HiCloud Technology and China Communications Construction, is not just an engineering experiment; it is a live test of how countries may support AI demand with lower cooling loads, tighter power efficiency, and more renewable supply. According to reporting translated by WIRED en Español, the site pairs offshore wind power with seawater cooling at a moment when compute growth is reshaping infrastructure decisions.
China opens the first offshore-wind underwater data center
The project sits in Shanghai’s Lin-gang Special Zone, inside the China Pilot Free Trade Zone, with modules submerged about 10 meters underwater. Its initial capacity is 24 MW, and the operating model is designed to use seawater as a natural cooling system rather than depend primarily on conventional air-conditioning.
That matters because cooling remains one of the least flexible costs in AI infrastructure. In a standard onshore facility, cooling systems can account for 40 to 50 percent of total electricity demand, as the source article notes. By contrast, the Chinese project aims to keep cooling energy below 10 percent of total usage, a sharp reduction if achieved in regular operation.
The facility is also notable for its energy source. HiCloud had already launched a commercial underwater data center in Hainan in 2023, but the Shanghai complex is the first reported to operate with offshore wind. For enterprise leaders tracking AI implementation services and infrastructure risk, that is the more important signal: compute expansion is increasingly tied to where clean power and efficient cooling can be secured together, not separately.
Why the project matters for AI infrastructure demand
The broader context is simple: AI growth is turning power planning into a board-level issue. A recent UNCTAD report referenced in the source coverage says only 32 countries host AI-specialized data centers, and about 90 percent of that infrastructure is concentrated in China and the United States. That concentration means capacity is not only a software question. It is also a siting, grid, and procurement question.
For enterprises in technology infrastructure, energy, and manufacturing, this changes how AI strategy should be evaluated. The key constraint may not be model choice. It may be access to reliable electricity, acceptable operating cost, and enough thermal efficiency to keep expansion viable.
Another practical takeaway is that infrastructure design now affects downstream AI business automation decisions. If compute-intensive workloads become more expensive in high-heat, high-grid-stress regions, companies will need to be more selective about where training, inference, analytics, and AI automation agents are deployed.
Get one practical AI-program note a week. Subscribe to the Encorp newsletter.
The numbers behind the efficiency claim
The strongest reason this story deserves attention is the density of measurable claims. According to the Chinese government’s project statement, the underwater site is designed to use more than 90 percent offshore wind power, reduce overall energy use by 30 to 40 percent, and cut refrigeration energy demand from 40 to 50 percent of total power use to under 10 percent, versus traditional onshore data centers.
A second key metric is PUE 1.15. As Google explains in its data center efficiency overview, power usage effectiveness measures total facility power divided by IT equipment power, with 1.0 representing the theoretical ideal. A target of 1.15 puts the Shanghai project in state-of-the-art territory on paper.
Three numbers stand out most:
- 24 MW initial capacity, enough to make this a serious infrastructure asset rather than a lab pilot.
- Under 10 percent of power devoted to cooling, versus the 40 to 50 percent often seen in conventional designs per the source reporting.
- More than 90 percent offshore wind power, tying compute expansion directly to renewable power procurement.
Those figures also explain why this is more than an environmental story. They point to operating economics. Lower cooling overhead improves margins and can stabilize long-term capacity planning. Better PUE makes AI data analytics and high-availability workloads easier to budget. And renewable integration may reduce exposure to fossil-fuel price swings, though it introduces its own intermittency and grid-balancing trade-offs.
There are also constraints. Underwater maintenance is more complex than servicing a standard land-based hall. Site selection is narrow. Insurance, repair logistics, and subsea component reliability will matter far more than they do in a normal enterprise build. In other words, this is an important signal, not a universal template.
What China’s policy shift says about the AI race
This project makes more sense when read alongside China’s broader energy policy. The source article notes that a new energy law took effect last year with priority given to renewables and hydrogen, while electricity-market reforms from June 2025 require solar and wind power to be traded through market mechanisms or auctions rather than older fixed tariff structures.
That policy shift matters because AI infrastructure needs long-duration planning. Power contracts, site development, and integration services all depend on confidence that generation capacity will exist where compute demand is growing. China appears to be treating this as an industrial strategy issue, not just a sustainability issue.
The comparison with the United States is not that one side is building data centers and the other is not. It is that they are emphasizing different routes to supply security. China is trying to reduce reliance on external fossil inputs while scaling renewables and nuclear options. That gives projects like the Shanghai deployment a larger strategic role: they test whether AI-era infrastructure can be both more efficient and more domestically controllable.
This is also where International Energy Agency analysis on electricity and data centers becomes useful context. AI demand is pushing utilities, operators, and large buyers toward more detailed forecasting of peak loads, resiliency needs, and transmission constraints. In practice, AI integration services now depend as much on power realism as on software architecture.
How this compares with conventional data-center design
A useful way to read this development is not as a binary choice between underwater and onshore designs, but as a benchmark against legacy assumptions. Traditional facilities compete on location, tax incentives, and fiber access. Newer AI-oriented facilities also have to compete on cooling path, water consumption, and power procurement.
| Design factor | Conventional onshore data center | Shanghai underwater model |
|---|---|---|
| Cooling method | Mechanical air conditioning | Seawater cooling |
| Cooling energy share | Often 40-50% per source article | Under 10% by design |
| PUE target | Varies widely by age and site | 1.15 target |
| Land use | Significant campus footprint | More than 90% lower by project claim |
| Power source mix | Grid-dependent, often mixed | More than 90% offshore wind power by project claim |
That comparison has direct implications for enterprise procurement. Buyers should care more about where workloads run, what the energy mix looks like, and whether vendors can report performance metrics consistently. This is especially true for firms scaling AI implementation services across multiple business units.
For operators thinking more tactically, this is where a service such as AI Smart Energy Management for Facilities fits best: the same disciplines that matter in a subsea data center, such as load prediction, anomaly detection, and continuous efficiency monitoring, also matter in ordinary enterprise estates. The difference is not the need for operational visibility. It is the level of engineering complexity.
What enterprises should watch next
Three indicators now matter more than they did a year ago: PUE targets, renewable power share, and regional cooling constraints. Those are no longer side metrics for sustainability teams. They are becoming part of mainstream AI strategy and vendor due diligence.
Enterprises should also watch whether projects like this remain isolated national showcases or start to influence commercial design standards in Asia, Europe, and North America. If more facilities pursue alternative cooling and tighter energy integration, buyers may begin to treat infrastructure efficiency as a core selection criterion rather than an ESG footnote.
The trend is clear even if this exact design remains niche. AI for energy is moving from abstract policy talk to physical infrastructure choices measured in megawatts, PUE, and procurement terms. The companies that plan around those constraints early will make better decisions about where AI workloads belong and what they should cost.
Martin Kuvandzhiev
CEO and Founder of Encorp.io with expertise in AI and business transformation