Plans to put data centers on or under the sea have been announced one after another in 2026. In May, Panthalassa announced a large funding round, and in July, Mocean Energy unveiled "Blue Core," its concept for an offshore data center. In the United States, a regulator accepted a preliminary permit application for a seabed data center in August, and in October, Ocean Power Technologies and Stratum Holdings signed a letter of intent. Challenges for offshore operation have also been pointed out, including drift, saltwater corrosion, and marine fouling. Comparing the seven projects we could identify, progress varies widely, from facilities already running as demonstrations to projects still at the concept stage. The scale announced for a project does not necessarily match its prospects of being realized.
Comparing the progress of seven offshore and seabed data center projects
For the seven projects we could identify, we organized the power source, announced scale, and stage actually reached. We distinguish between planned output or capacity and the scale of facilities already operating or being demonstrated.
| Project | Installation method / power source | Announced scale / target | Stage reached (date) | Main sources |
|---|---|---|---|---|
| Shanghai Lingang "HiCloud" | Seabed, directly connected to offshore wind power | Planned capacity 24 MW (first phase 2.3 MW), total investment 1.6 billion yuan (about $228 million) | First phase reported operating (May 2026) | Chinese media, developer |
| Panthalassa | Offshore floating platform, wave power | Aims for commercial deployment in 2027 | Plans AI inference demonstration with Ocean-3 in 2026 (announced May 4) | The company |
| Mocean Energy "Blue Core" | Offshore floating platform, wave and solar power | 500 kW to 1 MW per unit | Developing shipboard AI server racks; small-scale demonstration planned for 2027 | The company |
| Ocean Power Technologies (OPT) and Stratum Holdings | Offshore buoy, wave power | Concept to run AI inference without a power transmission cable to shore | Non-binding letter of intent (October 26) | Both companies |
| DeepGreen Western Passage | Seabed, tidal power | Initially 51 MW; after revision, 10–15 MW in the second phase | FERC accepted preliminary permit application (notice dated August 19) | Federal Register, local press |
| Mitsui O.S.K. Lines, Hitachi, Hitachi Systems | Converted used vessel moored in port | Aims to begin operation in 2027 or later | Basic agreement (March 30) | Each company |
| Yokohama Port, Osanbashi Pier | Floating unit in port, solar power and batteries | Operated on renewable energy only | Demonstration under way (began March 25, through end of fiscal 2026) | City of Yokohama and others |
Note: Based on published materials as of November 17, 2026. This is not an exhaustive survey. In addition, the Shanghai Lingang, Yokohama Port, and Mitsui O.S.K. Lines/Hitachi projects are not explicitly described as "AI data centers" in published materials.
The only project currently reported as operating is the 2.3 MW first-phase seabed data center at Lingang in Shanghai, China, while a floating data center demonstration is also under way in Yokohama port in Japan. The three wave-powered projects, by contrast, are at the stage of aiming for AI inference demonstrations at sea and future commercial deployment. Panthalassa says earlier prototypes demonstrated computing technology at sea, but reports differ in how they describe what was actually demonstrated. DeepGreen is in the middle of the permitting process, and the Mitsui O.S.K. Lines and Hitachi plan is at the basic agreement stage.
For the Lingang project, Chinese media and the developer explain that the data center is placed about 10 meters below the sea surface, between the first and second phases of an offshore wind farm, using seawater cooling and power supplied directly from the wind farm. Materials from June 2025 said a 2.3 MW demonstration facility would be built in the first phase, with expansion to 24 MW including the second phase. English-language materials from the Shanghai municipal government also say operation began in May 2026, but the 24 MW figure is the planned capacity across both phases. It is therefore necessary to distinguish between the scale confirmed as operating in published materials and the planned future capacity.
The Yokohama Port demonstration was started on March 25 by Nippon Yusen Kabushiki Kaisha, NTT Facilities, Eurus Energy Holdings, MUFG Bank, and the City of Yokohama. It is not commercial operation but an effort to verify whether a data center can run stably in an environment exposed to the salt damage and vibration unique to the sea.
Funding and output targets run ahead in three wave-powered projects
In May 2026, Panthalassa announced it had raised $140 million in a Series B round led by Peter Thiel. The company, founded in 2016, is developing offshore nodes that feed power from wave energy to onboard AI chips and send computing results to shore via low-Earth-orbit satellites. On November 6, it announced the opening of a factory in part of the former Kaiser shipyard in Vancouver, Washington, shifting from building nodes individually to assembly-line production. It is expected to employ about 50 people once staffed. Local public broadcaster OPB reported that the company has raised more than $300 million in 2026 and had been aiming to raise another $225 million in the summer.
Although fundraising and manufacturing facilities are progressing, the company's own announcement still positions the demonstration of AI inference at sea as a future plan. According to the May 4 announcement, prototypes deployed in 2021 and 2024 demonstrated technology for power generation, propulsion, autonomous navigation, and computing at sea. The AI inference demonstration itself is planned for "Ocean-3," scheduled for deployment in the North Pacific in 2026.
However, Latitude Media reported that the early prototypes "Ocean-1" and "Ocean-2" were intended mainly to verify the power generation mechanism and carried no data center components. Because Panthalassa's announcement describes technology demonstrations across multiple prototypes together, it is impossible to tell from the two accounts alone which unit performed how much computing. The company says each unit can generate up to 1 MW and estimates manufacturing costs at $1 million to $1.5 million, but this figure does not include transportation or maintenance costs.
At Mocean Energy of Edinburgh, UK, there is a wide gap between the power output reported from its test unit and its future target. According to reports, in sea trials that ended in spring 2024, the test unit's output was about 10 kW. By contrast, the company's target for an offshore data center combining wave and solar power is about 500 kW per node, a gap of roughly 50 times on paper.
The test unit reportedly generated power even in waves 40 feet (about 12 m) high, while output fell in small waves. However, it is unclear whether the test unit and the commercial Blue Core node share the same design, and the reports do not make clear whether the roughly 10 kW is a rated output or an actual measurement. The roughly 50-fold gap is therefore a comparison of figures given under different conditions and does not directly represent an improvement factor in generation performance.
Mocean also has a concept of grouping multiple nodes into clusters of up to 100 MW. By simple arithmetic, that would take 200 nodes at 500 kW, or 100 nodes assuming the maximum 1 MW per unit stated in the July announcement. However, the AI server racks for shipboard use are listed as "in development" on the company's website, and a small-scale demonstration is planned for 2027. The July announcement also set out a plan to begin the first offshore demonstrator and initial commercial deployment from 2030.
Several figures have also been given for long-term deployment targets. In its July 2026 announcement, Mocean set a target of deploying 6 GW over 10 years, which it said corresponds to about 2% of projected AI data center demand. Meanwhile, there are reports that co-founder Cameron McNatt is aiming for 10 GW, and a concept of deploying 1 GW over five years has also been reported. It is not clear from published materials whether these refer to the same scope or period.
The plan by Ocean Power Technologies (OPT) and Stratum Holdings is at an even earlier stage. The letter of intent the two companies signed on October 26 is not legally binding. Under the concept, OPT would provide a limited, non-exclusive license to its patents for the "PowerBuoy" wave energy buoy along with an existing prototype buoy, while Stratum would provide GPUs, including NVIDIA B300, and system integration technology. Stratum is expected to pay OPT a monthly fee, with the aim of verifying whether AI inference can be run at sea without a power transmission cable to shore.
Why send computing results to shore instead of power
The three projects combining offshore wave power with data centers share a common aim: avoiding the cost of laying subsea power cables. Bryson Robertson of the University of Victoria in Canada estimates that subsea cables carrying power from the open ocean to shore can cost hundreds of millions of dollars even over a distance of just a few miles. Panthalassa uses the power it generates directly in onboard AI chips and sends only the computing results to shore by satellite. By transmitting data instead of large amounts of power, the approach aims to eliminate the need for large subsea transmission facilities.
However, without a connection to the power grid, fluctuations in wave power output must be absorbed on board alone. A land-based data center can make up shortfalls from the grid, but for a standalone offshore node, batteries and similar equipment are the main means of adjustment. Climate technology consultant Michael Barnard estimates that running 1 MW of net IT equipment would require average generation of 1.3 to 2 MW, including cooling, power conversion, and propulsion to hold position. Moreover, if the generating equipment's capacity factor is 25–35%, rated output on the order of several megawatts would be needed. In other words, the maximum 1 MW generating capacity per unit cannot simply be treated as capacity able to support 1 MW of IT load.
Reducing the power needed for cooling is also cited as an advantage of offshore data centers. Mocean expects cooling to account for about 5% of total energy consumption at sea, versus 15–35% at land-based data centers. OPT and Stratum also believe cold seawater can cut cooling power consumption, but both state in their announcement that this effect has not yet been demonstrated. The developer of the Lingang project claims it cut power consumption by 22.8% compared with conventional facilities and held PUE to about 1.15. PUE is a metric obtained by dividing a data center's total power consumption by that of its IT equipment, but we could not find independent verification of these figures.
On equipment reliability, drift, saltwater corrosion, and marine fouling are cited as common challenges. For seabed systems, there is a precedent in Microsoft's "Project Natick." In the second phase of that project, a sealed data center module placed on the North Sea floor had a server failure rate one-eighth that of a land-based control group. However, this was the result of a sealed facility fixed to the seabed and does not indicate the reliability of floating data centers rocked by waves. Microsoft acknowledged in 2024 that Project Natick had already ended.
Wave power itself also faces commercialization difficulties. In 2025, Finland's AW-Energy went bankrupt and the US company AquaHarmonics ceased operations. Robertson points out that wave power has not attracted as much investment as other renewable energy sources, and that there have been cases where the failure of just one or two prototypes made it hard for a company to survive. Realizing offshore AI data centers requires not only computing equipment but also maturity in the wave power technology that supplies its electricity.
From 51 MW to a maximum of 15 MW: DeepGreen's seabed data center plan shrinks
One of the seabed data center plans advancing in the United States is "DeepGreen Western Passage," off Eastport, Maine. The developer, DeepGreen Western Passage SPV, applied to the US Federal Energy Regulatory Commission (FERC) on February 11, 2026, for a preliminary permit for a plan combining tidal power generation with a seabed data center. The concept was initially on the scale of 51 MW, but later amendments reduced the size of the facilities. According to the local paper The Quoddy Tides, an amendment on June 8 changed the first phase from 15 MW to 1–5 MW and the second phase from 51 MW to 10–15 MW.
In the permit review, additional explanation was requested about the scale and footprint of the facilities. According to The Maine Monitor, FERC pointed out four deficiencies in the application in a letter dated July 15 and demanded an explanation of why a 400-acre "operating area" was being sought when the actual equipment would fit within 27 acres. It also asked for a specific number of tidal turbines, which the original application had put at about 170.
After an amendment on July 30, the acceptance notice published by FERC dated August 19 listed the number of tidal turbines at a maximum of 16. The notice also served as a call for comments, with a submission deadline of October 19.
However, reports differ on when the output reduction was decided. The Federal Register notice records amendments on June 8 and July 30, consistent with The Quoddy Tides' report that the scale was reduced on June 8. The Maine Monitor, on the other hand, describes the same reduction as a July amendment. The Bangor Daily News also reported that the plan's scale and output were reduced following FERC's requests for additional information, and that a phased development approach was adopted on the advice of a staff member. It cannot be determined from published materials whether the reduction had already been completed by June 8 or was further changed on July 30.
According to the August 19 notice, the plan includes 34 seabed computing pods, each 12 m long and 4.2 m in diameter, and up to 16 tidal turbine generators rated at 300 kW each. It also calls for laying a 1.5-mile, 13.8 kV cable that bundles power and optical fiber. The plan is to operate the data center solely on self-generated tidal power, without connecting to the onshore power grid.
What matters here is that FERC's acceptance of a preliminary permit application does not amount to construction approval. Even if a preliminary permit is granted, what the developer obtains is a priority position to file a formal license application while conducting studies, and work in the waters is not authorized. The Maine Monitor reports the preliminary permit term as 48 months. For DeepGreen to actually install equipment, it must obtain separate required licenses.
Moves toward regulation are also advancing at the local level. According to The Maine Monitor, the Eastport City Council unanimously passed a 180-day moratorium on the data center plan in August. The Quoddy Tides further reports that at its regular meeting on September 9, an ordinance banning large data centers in the city was unanimously adopted. DeepGreen is reportedly at the stage of selecting equipment supply partners and has not yet placed any equipment in the water.
Environmental impacts are also a concern. The Center for Biological Diversity points out that increased vessel traffic related to the project could raise the risk of collisions with whales, and that noise and thermal discharge could affect the marine environment.
Japan opts for port-moored systems, with demonstrations clustered around 2027
Japan's two projects are being pursued in a direction that limits the challenges of offshore installation. On March 30, Mitsui O.S.K. Lines, Hitachi, and Hitachi Systems signed a basic agreement toward developing and commercializing a floating data center built from a converted used vessel. They will study demand and equipment specifications mainly in Japan, Malaysia, and the United States, aiming to begin operation in 2027 or later. Conversion work on the used vessel is expected to take about one year, and they anticipate shortening development time by up to three years compared with a land-based data center. The plan is to use seawater or river water for cooling.
For Mitsui O.S.K. Lines, a concept has also been reported in which power would be supplied to the moored vessel from shore or from a separate power-generating vessel. A port-moored approach avoids the propulsion needed for position keeping and the large-scale offshore maintenance required for floating platforms at sea. And if onshore power or a power vessel can be used, there is no need to absorb output fluctuations, as with wave power, on board alone. Measures against salt damage and vibration are still required, however.
The Yokohama Port demonstration is investigating exactly this kind of operating performance in a marine environment. A small floating unit (mini-float) installed at Osanbashi Pier carries a containerized data center, solar power equipment, and batteries. The plan is to operate the data center on renewable energy alone and, by around the end of fiscal 2026, verify the effects of salt damage and vibration on operation and the stability of power management combining generation and storage equipment.
Looking across the companies' plans, important demonstrations and commercialization schedules are concentrated from the end of fiscal 2026 through 2027. Panthalassa announced on May 4 that it will deploy Ocean-3 in the North Pacific in 2026 to demonstrate AI inference and aims for commercial deployment in 2027. Mocean plans a small-scale demonstration in 2027, and the Mitsui O.S.K. Lines and Hitachi floating data center also looks toward operation in 2027 or later. In Yokohama Port, demonstration results are due to be compiled by the end of fiscal 2026, and DeepGreen is waiting for FERC's decision.
What matters in evaluating these plans is actual operating data rather than announced output targets or investment amounts. Concrete benchmarks will be how stably Panthalassa's Ocean-3 can run AI inference at sea, how far the Yokohama Port floating data center can limit the effects of salt damage and vibration, and whether DeepGreen obtains a preliminary permit from FERC. Once such results become clear, it will be possible to compare the technical feasibility and applicable scale of seabed, port-moored, and offshore wave-powered approaches based on real operating conditions.
