In May 2026, a 1,950-ton steel structure was lowered into the East China Sea, about 10 kilometers off the coast of Shanghai. It contained 192 server racks, roughly 2,000 high-performance servers, and a cooling system that uses seawater as its primary heat sink, with no freshwater consumption. The facility is powered almost entirely by an offshore wind farm 500 meters away.
According to Chinese state media and industry reports, it is the world's first commercial underwater data center directly powered by offshore wind. It cost 1.6 billion yuan ($226 million) to build. The facility is the commercial version of a concept Microsoft developed for nearly a decade before abandoning it in 2024 when the economics did not work.
The Shanghai Lingang facility is not a research experiment. It began full commercial operations in May 2026 and is already serving AI training workloads, cross-border data processing, and financial computing for clients including China Telecom and Shanghai Instrumentation. It is the second-generation product from HiCloud Technology, a private Chinese engineering firm that has been refining the concept since 2021.
What Microsoft Proved — and Why It Stopped
Microsoft's Project Natick was the first serious attempt to put data centers underwater. In 2018, the company deployed a container-sized pod containing 864 servers to the seabed off Scotland's Orkney Islands, about 36 meters deep. The pod ran for two years on renewable energy from tidal and wind sources.
The results were better than expected. When Microsoft retrieved the pod in 2020, it found that the servers inside had a failure rate of just one-eighth that of a comparable land-based data center. The sealed, oxygen-free, temperature-stable environment eliminated many of the failure modes that plague conventional facilities. The Power Usage Effectiveness (PUE) — a measure of how much energy goes to computing versus cooling — dropped to 1.07, far below the industry average of around 1.5.
The technology worked. The economics did not.
Microsoft confirmed in June 2024 that it had shut down Project Natick and would not pursue underwater data centers as a commercial product. The reason, according to people familiar with the decision, was that customers showed little interest in scaling the model. They preferred expanding traditional land-based facilities, which were cheaper and faster to upgrade. The sealed pod design — which cannot be opened for maintenance or hardware upgrades — was fundamentally incompatible with the pace of AI chip development, where processors are replaced every one to two years.
Microsoft kept the project as a research platform. It did not build a commercial underwater data center.
What China Did Differently
HiCloud Technology, the company behind the Shanghai Lingang facility, took the same technical concept and rebuilt the business model around it.
The company's first-generation underwater data center, deployed off Hainan's Lingshui coast in 2022, proved the cooling concept. According to HiCloud's disclosed specifications, each pod contained 360 kilowatts of IT load and was powered from the shore. It worked, but the economics were marginal.
The second-generation Lingang facility changed two critical variables. First, it scaled each pod from 360 kilowatts to 2.3 megawatts — a 6.4-fold increase in capacity per unit. Second, and more importantly, it switched from shore power to direct offshore wind power, connecting the data center to a 200-megawatt wind farm 500 meters away via a composite fiber-optic and power cable.
The direct connection eliminates the transmission losses and grid fees that make land-based renewable-powered data centers expensive. The wind farm and the data center are co-located in the same marine area, sharing a substation and operations and maintenance crews. The power is generated, transmitted, and consumed within a single integrated system.
The result is a facility that achieves a PUE of 1.15 or lower, consumes zero freshwater, and occupies 90 percent less land than a comparable land-based facility. The green power supply rate exceeds 95 percent. Annual electricity savings, at full 24-megawatt capacity, are estimated at 61 million kilowatt-hours.
The Cost Arithmetic
The key question for any data center is not whether it can be built, but whether it can be built at a competitive cost.
HiCloud's general manager, Su Yang, told Cailianshe that during the construction phase, the per-kilowatt cost of an underwater data center can approach that of a comparable land-based facility. The capital expenditure is not dramatically higher. The savings come during operations.
The operating cost advantage is substantial. Cooling accounts for 30 to 40 percent of a conventional data center's total energy consumption. At the Lingang facility, cooling energy is compressed to below 10 percent of total consumption. The overall system uses 30 to 40 percent less energy than a comparable land-based facility. The unit computing energy cost is more than 20 percent lower than traditional cross-border data centers.
The project's payback calculation depends on the wind power price. Offshore wind in China's coastal regions is increasingly competitive with grid electricity, and the direct-connection model avoids the transmission and distribution charges that can add 30 to 40 percent to industrial power bills. Over a 15-to-20-year lifecycle, the total cost of ownership for an underwater data center is lower than its land-based equivalent.
The prefabrication model contributes to the cost advantage. Each data pod is built on land using modular construction, then lowered into the sea. The second-generation pod has a design life exceeding 25 years and can be built and tested in 90 days. The Lingang facility was officially launched in June 2025, completed in October, and began commercial operations in May 2026 following successful trials.
The Upgrade Problem — and How China Solved It
Microsoft's core objection to commercial underwater data centers was upgradeability. A sealed pod on the seabed cannot be opened. AI chips are replaced every 12 to 24 months. The sealed design meant that every hardware refresh required retrieving the pod, opening it, and redeploying it — a process that was neither cheap nor fast.
HiCloud's design addresses this in two ways.
First, the pods are modular and replaceable. The Lingang facility uses multiple independent data pods, each with its own power and cooling connections. When a pod's hardware becomes obsolete, it can be disconnected, retrieved, and replaced with a new pod. The marine infrastructure — the foundation, the cable, the wind connection — remains in place. The upgrade cost is the cost of building a new pod, not rebuilding the entire facility.
Second, the pods are designed for retrieval. Unlike Microsoft's fixed seabed installation, HiCloud's pods are deployed on a retrievable platform. The company has not disclosed the exact retrieval mechanism, but the design allows for periodic hardware refresh without decommissioning the entire site.
This does not fully solve the upgrade problem. A land-based data center can swap a GPU in hours. An underwater data center must retrieve, open, and redeploy a pod — a process measured in weeks. But it makes the model viable for AI workloads that are stable enough to justify a multi-year deployment, even if it is not suitable for the fastest-moving frontier training clusters.
The Scale Ambition
HiCloud is not stopping at 24 megawatts. The company has plans for a 500-megawatt underwater data center cluster — more than twenty times the Lingang facility's capacity. That scale would make it one of the largest single data center projects in the world, comparable to the largest hyperscale campuses under development in the United States.
The expansion plan builds on the Lingang facility's integrated model. Each new pod added to the cluster shares the same wind connection, the same marine space, and the same operations infrastructure. The marginal cost of adding capacity is the cost of the pod itself, not the cost of new transmission lines or new land acquisition.
The market context supports the ambition. China's AI computing demand is growing faster than its supply. The country's intelligent computing capacity reached 2,185 EFLOPS by June 2026, up 177 percent year over year. The coastal provinces — Shanghai, Guangdong, Zhejiang, Jiangsu — host the densest concentration of AI companies and the highest demand for low-latency compute. They also have the least available land and the tightest energy constraints.
The underwater data center solves both problems at once. It puts the compute where the users are without consuming land. It uses the ocean as a free cooling source. And it connects directly to offshore wind, avoiding the grid congestion that delays land-based projects.
What the Project Proves
The Shanghai Lingang facility does not prove that underwater data centers are the future of AI infrastructure. The model has real limitations: hardware upgrades are slow, maintenance is expensive, and the addressable market is limited to coastal regions with access to offshore wind.
What it does prove is that the concept Microsoft abandoned is commercially viable under the right conditions.
Microsoft's Natick project was technically successful but commercially stranded. It had no customer demand because it had no cost advantage. The facility used tidal and wind power from the Orkney grid, but it did not have a direct connection to a dedicated offshore wind farm. It was a research pod, not a production facility.
HiCloud's Lingang facility has a direct connection to a 200-megawatt wind farm. It shares the marine area, the substation, and the operations crew with the wind operator. The power is generated and consumed within a single integrated system. That integration is what makes the economics work.
The company that figured this out is not an American hyperscaler. It is a private Chinese engineering firm that started with a 360-kilowatt pilot off Hainan four years ago and iterated its way to a 24-megawatt commercial facility. The technical challenges proved manageable. The business model was the harder problem.
Microsoft demonstrated that the concept could work technically. The Shanghai facility demonstrates that it can work commercially.
Sources: CNR (April 3, 2026); China Daily (February 10, 2026); China Telecom (April 16, 2026); China News Service (February 10, 2026); Shanghai Securities News (February 10, 2026); C114 (June 9, 2026); The Paper (April 2, 2026); Baidu Baike — Northern Isles Data Center; 36Kr (May 20, 2026); Data Center Dynamics (May 20, 2026); Cailianshe (July 17, 2026); Anadolu Agency (June 12, 2026); WIRED (June 10, 2026).
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.
Company specifications come from public disclosures; independent verification is limited.
Cost and payback figures are point-in-time estimates; actual results vary.
Comparative claims rely on public reporting, not operational data.
The project's multi-year performance has not been independently audited.
Capacity and PUE figures may not be directly comparable across firms.
AI infrastructure is changing quickly; claims can become outdated soon.
Sources
- CNR (April 3, 2026)
- China Daily (February 10, 2026)
- China Telecom (April 16, 2026)
- China News Service (February 10, 2026)
- Shanghai Securities News (February 10, 2026)
- C114 (June 9, 2026)
- The Paper (April 2, 2026)
- Baidu Baike — Northern Isles Data Center
- 36Kr (May 20, 2026)
- Data Center Dynamics (May 20, 2026)
- Cailianshe (July 17, 2026)
- Anadolu Agency (June 12, 2026)
- WIRED (June 10, 2026).
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.; Company specifications come from public disclosures; independent verification is limited.; Cost and payback figures are point-in-time estimates; actual results vary.; Comparative claims rely on public reporting, not operational data.; The project's multi-year performance has not been independently audited.; Capacity and PUE figures may not be directly comparable across firms.; AI infrastructure is changing quickly; claims can become outdated soon.