發生了什麼事
Fortune, republishing reporting by Nir Kshetri and The Conversation, describes data-center projects in China, Japan, Singapore, South Korea, Portugal and the United States that use underwater or floating infrastructure to support AI computing. The approaches rely partly on seawater cooling and, in some cases, offshore renewable power. The report says China’s Hainan facility entered full commercial operations in May 2026, while other projects remain tests, proposals or facilities under construction.
Fortune published the report on August 25, 2026, under the byline of Nir Kshetri and The Conversation. It frames the development as a response to the land, energy, water and carbon pressures associated with the AI boom. The report describes several forms of ocean-based infrastructure rather than one standardized technology: sealed modules placed on the seafloor, data-center containers mounted on floating platforms, and land-based facilities that draw seawater for cooling. These projects are presented as attempts to address constraints around AI infrastructure, not as evidence that a mature offshore data-center industry already exists.
The report traces the idea to Microsoft’s underwater data-center research program, launched in 2015. In 2018, Microsoft placed a sealed data center containing 864 servers on the seafloor near Scotland’s Orkney Islands. After two years, Microsoft reported that servers in the underwater unit failed at about one-eighth the rate of comparable land-based servers. The company hypothesized that the sealed environment reduced exposure to oxygen, humidity and changes and limited physical disturbance from maintenance work. Fortune says Microsoft ended the project in 2024 and did not disclose why; outside analyses cited in the report suggest regulatory requirements and the difficulty of upgrading or replacing equipment may have contributed.
Fortune reports that China’s Hainan underwater facility, described as possibly the first wind-powered underwater data center, launched in June 2025 and began full commercial operations in May 2026. The $226 million project reportedly uses seawater cooling and consumes at least 30% less electricity than traditional data centers. The report also describes a floating-platform test near Yokohama, Japan, using shipping containers, solar panels and batteries; a four-story floating facility that Keppel began building in Singapore and expects to open in 2028; and planning in Ulsan, South Korea, for an underwater center that could house more than 100,000 servers. Other examples include a proposed tidal-powered facility in Maine and Portugal’s SIN01 facility in Sines, which uses Atlantic seawater for cooling.
The source does not independently verify the projects’ performance figures, commercial durability or environmental effects. The reported savings are largely claims from companies, contractors or project descriptions cited by Fortune. Some initiatives remain proposals or tests, and the article does not provide comparable measurements across facilities, details of the computing workloads being run, or a full accounting of construction, cabling, renewable-power and costs.
為什麼這很重要
AI data centers require substantial electricity, cooling and physical space. Moving some infrastructure offshore could reduce pressure on land and treated freshwater and may avoid local opposition to large onshore facilities. But ocean deployment does not eliminate energy use or carbon emissions, and heated discharge, difficult repairs, environmental permitting and unclear long-term maintenance costs could create new public risks.
AI computing is increasing demand for large data centers, whose impacts extend beyond the servers themselves. They need electricity for computation, additional energy for cooling, water or other cooling media, and large parcels of land connected to power and communications networks. Fortune reports that ocean-based designs could reduce demand for land and treated freshwater, while offshore wind, solar or tidal generation could reduce reliance on grid electricity or fossil fuels in specific projects. Those potential benefits are relevant to communities facing competition over land, water and power supplies.
The siting question also has a public-acceptance dimension. Fortune cites a March 2026 Gallup poll in which 70% of Americans opposed building AI data centers in their communities. Offshore facilities could reduce some direct neighborhood impacts, such as land conversion, noise and visible industrial development. At the same time, the report notes that more than half of the world’s population lives within 120 miles of a coast, suggesting that offshore placement would not necessarily put computing far from users. That proximity could help network performance, although the source provides no measured or service-quality comparison.
The environmental tradeoff is not simply land versus ocean. Seawater cooling can reduce freshwater demand and the electricity required for refrigeration, but returning warmer water to the sea may affect local oxygen levels, pH and marine organisms. Fortune reports that the engineering contractor for China’s Hainan facility measured a seawater increase of less than 1 degree Celsius near the site, while Portugal’s SIN01 returns water about 1 degree Celsius warmer. Those observations describe localized measurements, not a definitive assessment of ecosystem impact. The report warns that cumulative heat from multiple facilities could produce thermal pollution in places where marine species depend on stable temperatures.
The source also places the issue in a broader climate context: oceans are already warming, and UNESCO estimates that about 60% of marine ecosystems are degraded or used unsustainably. Ocean infrastructure could therefore shift some burdens from populated land areas into ecosystems already under stress. The public-interest significance depends on whether operators can demonstrate that reduced freshwater use or lower operating electricity outweighs impacts from construction, cables, anchors, vessels, equipment and heated discharge. Fortune does not provide a complete lifecycle comparison with land-based AI data centers, so the net environmental result remains unknown.
互動機制：它實際上是如何運作的
以互動方式探索這項發展背後的基礎技術。
What should a useful AI forecast state?
接下來看什麼
The key questions are whether ocean-based facilities can operate reliably at commercial scale, whether their environmental claims hold up under independent measurement, and how regulators assign responsibility for marine impacts. Watch for evidence on lifecycle emissions, seawater changes, effects on marine ecosystems, repair logistics, equipment-replacement cycles, insurance and permitting, as well as whether projects progress beyond pilots and proposals.
Maintenance is the clearest operational test. Fortune reports that a failed computer in an underwater facility cannot be repaired or replaced on site; the entire sealed module may need to be brought to the surface. That could make routine upgrades slower and more expensive than in a land-based center, where technicians can reach individual servers. Watch for published data on failure rates after the initial Microsoft experiment, frequency, replacement procedures, downtime, corrosion, cable damage and the ability to install newer hardware as AI systems change rapidly.
Regulation and accountability will determine whether proposed projects become durable infrastructure. Offshore facilities may require environmental permits and oversight involving coastal authorities, maritime regulators, energy agencies and fisheries interests. The source says regulatory concerns may have been among the reasons Microsoft did not expand its project, but Microsoft did not state its reasons publicly. Future reporting should identify which agencies approve each facility, what marine monitoring is required, who pays for remediation after a failure, and how operators disclose , emissions and ecological data.
The commercial claims need independent testing. The reported 30% electricity reductions for China’s facility and the planned Ulsan project are meaningful only if they use comparable workloads, weather conditions, power sources and accounting boundaries. A fair comparison should include the energy and emissions associated with manufacturing sealed structures, building offshore renewable equipment, laying cables, transporting modules, retrieving failed hardware and decommissioning the facility. It should also distinguish reduced cooling electricity from total data-center energy consumption.
Finally, watch whether projects advance from demonstration to repeatable deployment. Yokohama’s test is scheduled to continue through March 2027, and Keppel’s Singapore facility is scheduled to open in 2028, while the Maine and Ulsan projects described by Fortune are proposals or planning efforts. Useful evidence would include sustained operations, customer workloads, transparent environmental measurements and financial results. Until that evidence is available, ocean-based AI data centers are best understood as a set of active experiments and infrastructure bets, not a proven replacement for land-based facilities.