AI infrastructure · Energy and delivery · Spotlight

Data Centers Are Looking Offshore. The Investment Case Is Still Unproven

Published November 15, 2026 · Sources reviewed November 15, 2026

Wave-powered nodes, subsea pods, converted ships and moored barges are different bets. Each must prove it can deliver reliable, connected computing at a cost customers will pay.

The latest case for putting data centers at sea starts with a familiar terrestrial problem: large campuses need land, power, water, permits and local consent. On November 6, Panthalassa opened a factory in Vancouver, Washington, to build the wave-energy nodes it hopes to deploy offshore. That factory is tangible. The offshore computing business it is meant to support still has to prove itself. [S1], [S3]

“Offshore data center” is becoming a broad label for systems with very different designs. Some try to generate power and run compute on the same platform. Others put compute pods underwater and pair them with a separate energy system. Still others propose converting ships or mooring modular data halls on barges. Each relocates some terrestrial constraints; none makes power reliability, connectivity, maintenance, permitting or customer economics disappear.

Four designs, four different milestones

Panthalassa’s Ocean-3 is the most integrated version: an autonomous platform that uses wave power for onboard AI inference and is designed to send results by low-Earth-orbit satellite. Panthalassa announced a $140 million Series B in May and said it planned a northern Pacific pilot in 2026, with commercial deployments targeted for 2027. PNNL’s technical summary still describes Ocean-3 as under development. As of this review, we found no public confirmation that an Ocean-3 compute node had been deployed. The earlier Ocean-2 trial lasted three weeks and tested power-converter systems; it did not demonstrate Ocean-3’s computing performance. [S1], [S2]

The Vancouver factory changes a different part of the evidence. OPB reported that the site marks Panthalassa’s first effort to move from custom fabrication toward an assembly line. The company said the factory could employ about 50 people when fully staffed. That is evidence of an industrial step and an intended manufacturing path, not a production run, delivery rate, or cost per finished node. [S3], [S14]

Mocean Energy’s Blue Core combines wave power, offshore solar, storage and AI server racks in a separate design. Its current company page lists onboard compute as in development and a small-scale demonstration target for 2027. That is a company milestone to watch, not an operating computer at sea. Similar language about low-cost power and seawater cooling appears in developer materials; FJ has not found independent operating data establishing those economics. [S4]

DeepGreen Western Passage takes a site-specific subsea route. Its Maine proposal describes 34 compute pods and up to 16 tidal turbine-generators rated at 300 kilowatts apiece. Their combined 4.8-megawatt nameplate rating is a calculated equipment total; it is neither continuous output nor IT capacity. FERC accepted the preliminary-permit application for filing. That moves the application into review; it does not itself grant the requested study permit. Even an issued preliminary permit would not authorize construction. Eastport’s opposition and the project’s open legal and corporate questions are tracked in FJ’s project profile. [S5], [S16], [S6], [S15]

A ship or barge can make a different argument: reuse maritime assets and move modular data halls closer to coastal markets. Mitsui O.S.K. Lines, Hitachi and Hitachi Systems signed an MOU to study converting a used ship, including demand verification and feasibility work; their release says operations could begin in 2027 or later. It does not announce a selected vessel, a customer contract or a construction commitment. Seatrium’s separate 30-MW barge concept received Bureau Veritas Approval in Principle in July. The classification review covered basic design elements; it is an early engineering milestone, not a construction approval or operating project. [S7], [S8]

The distinction matters for investors: an Ocean-3 node, a tidal-powered subsea site, a converted vessel and a moored barge are not interchangeable units in a market forecast. They differ in power source, connection to shore, maintenance model, use of seawater, licensing path and likely workload. Announced megawatts cannot be compared until those boundaries are clear.

The sea offers energy. It does not promise firm power.

Wave energy is an available resource; a data center needs electricity at the moment its servers need it. A wave converter’s nameplate, peak output and annual yield are different measures. A compute platform must also account for storage, power conversion, redundancy and the possibility that generation, servers or communications are unavailable at the same time. These are engineering requirements inferred from the designs, not results published for Ocean-3 or Blue Core.

DOE’s support for open-water wave testing shows why prototype headlines are an incomplete benchmark. The agency describes demonstrations lasting six months to two years as a way to reduce risk and learn about installation, operations and maintenance. Its marine-energy overview lists salt, sediment, severe wave forces, costly offshore servicing, permitting and limited test access among the sector’s challenges. Those issues affect equipment survival and service costs before a developer can establish cost per useful compute hour. [S9], [S10]

Cooling is a real potential advantage, but “the ocean is cold” is not an operating plan. Heat still has to pass through a designed system into seawater. The system must manage fouling, corrosion, heat-exchanger performance, environmental monitoring and repairs. DOE’s discussion of marine-energy testing emphasizes measuring environmental interactions, noise and wildlife response; any project-specific impacts require site and design evidence, not a generic claim that offshore means lower impact. [S9], [S10]

Connectivity may define which computing fits

Panthalassa says its Ocean-3 design will run inference onboard and send results by satellite. That architecture could fit workloads where a relatively small result can be transmitted after substantial processing at sea. It is an analytical inference, not a disclosed customer use case. Workloads that need frequent exchange of large datasets, very low latency or tight integration with existing cloud regions may not fit as readily. The relevant test is not whether a satellite link exists, but whether a customer can complete a valuable workload at its required speed, reliability and cost.

A converted vessel or a jetty-moored barge might preserve more conventional connectivity, but the public announcements do not establish a full power-and-network design, port agreement, interconnection arrangement or customer. The closer an offshore design sits to shore infrastructure, the more it may rely on the very grid, port, cable and local permissions that a deep-ocean pitch is meant to avoid. The more self-contained it becomes, the more it must prove onboard generation, storage, network performance and autonomous operations.

Technical success has not guaranteed a commercial business

Microsoft’s Project Natick is a useful precedent, with a limit. The company built and retrieved a subsea research data center to examine reliability and operations below sea level. In 2024, Microsoft told Data Center Dynamics that Natick was no longer an active effort and that it was not building subsea data centers, while retaining lessons for other research. That history does not settle the case for newer wave-powered or ship-based systems. It does show that making an underwater data center work as an experiment and choosing it as a commercial product are separate decisions. [S11], [S12]

The same separation applies to capital. Panthalassa has disclosed a $140 million financing round, and NAVER announced an investment in the company. Those disclosures establish investor participation, not a compute purchase, supplier award or project revenue for either investor. Public-market investors need the size and terms of the exposure, plus evidence of materiality, before translating a venture-financing headline into earnings. [S1], [S13]

Investor lens: price delivered compute, not the concept

The core comparison is an all-in cost for reliable, customer-usable computing. It would need to include node or vessel capital, power-generation equipment, storage and conversion, server replacement, network service, installation and recovery vessels, marine insurance, port or mooring fees, environmental monitoring and downtime. No reviewed disclosure provides that full cost for the projects above. Company descriptions of “cheap” or “free” energy are targets or promotional claims until operating data and accounting boundaries support them. [S1], [S4], [S7], [S8]

  1. Measured operation: months of power, compute and connectivity data across real sea conditions, with capacity basis and downtime disclosed.
  2. Customer fit: an identified workload and commercial agreement that explain latency, data-transfer volumes, service levels and who pays.
  3. Repeatable service: inspection, repair, recovery, server replacement and insurance costs demonstrated in the target operating environment.
  4. Permitted deployment: a site-specific path through federal, state, local, maritime and environmental reviews. A filing, memorandum or design review is not an operating authorization.
  5. Manufacturing and finance: finished-unit costs, output rates, funding for deployment and a route to cash generation, separated from company-wide capital raises.

Offshore computing may become useful for selected workloads, especially where its power or cooling system can be located with the compute and the customer can accept the network model. The next investable proof point is narrower than the vision: a deployed system that repeatedly delivers paid computation, with a disclosed uptime and cost that survive servicing, storms and replacement cycles. Until then, the ocean is a setting for competing infrastructure experiments—not a shortcut around the economics of building a data center.

Sources and claim boundaries · 16 documents

Company schedules and product descriptions are attributed as announced. Government notices establish filing and process, not engineering performance. FJ’s analysis distinguishes design milestones from deployment, and capacity from power output or IT load.

  1. [S1] Panthalassa / PR Newswire · 2026-05-04

    Panthalassa announced a $140 million Series B, a 2026 northern Pacific Ocean-3 pilot plan and 2027 commercial-deployment target. These are company statements and forward targets; not evidence of completed deployment, customer contract, output or commercial service.

    Source role: Company financing and development announcement

  2. [S2] Pacific Northwest National Laboratory / U.S. Department of Energy · Publication date not stated

    PNNL says Ocean-2 underwent a three-week trial in February 2024 and that Ocean-3 is under development as an autonomous floating platform designed to host data centers. Ocean-2 testing is not Ocean-3 compute performance.

    Source role: Government technical summary

  3. [S3] OPB / Courtney Sherwood · 2026-11-07

    Reports Panthalassa unveiled its Vancouver, Washington factory on November 6, 2026, as the company’s first effort to move from custom work toward assembly-line manufacturing; roughly 50 staff are a future full-staffing figure, not a verified current count or production rate.

    Source role: Independent reporting

  4. [S4] Mocean Energy · Current company page reviewed 2026-11-15

    Mocean describes Blue Core as a proposed wave-, solar- and storage-powered offshore compute unit. Its page lists onboard compute in development and a small-scale demonstration target for 2027. This is not a completed compute demonstration or independent performance evidence.

    Source role: Company technical and schedule description

  5. [S5] Federal Energy Regulatory Commission / Federal Register · 2026-08-25

    FERC accepted DeepGreen Western Passage’s preliminary-permit application for filing. The application describes 34 compute pods and up to 16 300-kW turbine-generators (4.8 MW combined nameplate by arithmetic). Filing acceptance is not an issued study permit; even an issued preliminary permit does not authorize construction. Nameplate generation is not IT load.

    Source role: Project-specific government notice

  6. [S6] Federal Energy Regulatory Commission · Page last updated 2025-02-03

    FERC says eligible hydrokinetic pilot licenses are short-term and include environmental monitoring and safeguards. This is general process guidance; the route applicable to any specific offshore concept depends on its design and site.

    Source role: Government licensing guidance

  7. [S7] Mitsui O.S.K. Lines / Hitachi / Hitachi Systems · 2026-03-30

    The three companies announced an MOU to study conversion of a second-hand ship into a floating data center, including demand verification and feasibility work, with operations targeted for 2027 or later. It is a study plan, not a vessel order, customer commitment or service start.

    Source role: Company memorandum-of-understanding announcement

  8. [S8] Bureau Veritas Marine & Offshore · 2026-07-27

    Bureau Veritas granted Approval in Principle to Seatrium’s 30-MW barge concept, a design review covering fundamental arrangement, cooling process, stability and safety considerations. AiP is early technical review, not construction approval, customer demand, financing or operating evidence.

    Source role: Independent marine classification society announcement

  9. [S9] U.S. Department of Energy · Page last updated 2026-07-20

    DOE describes longer-duration wave-energy demonstrations expected to last six months to two years to reduce deployment risk and improve evidence on installation, operation and maintenance. This describes the testing rationale, not a requirement for every project.

    Source role: Government technology-testing and funding guidance

  10. [S10] U.S. Department of Energy · 2022-07-18

    DOE identifies salt, sediment, wave forces, offshore deployment and servicing costs, permitting and test-site access as marine-energy challenges, and describes monitoring of noise, wildlife interaction and environmental conditions.

    Source role: Government sector and environmental context

  11. [S11] Data Center Dynamics / Sebastian Moss · 2024-06-17

    Microsoft told DCD Project Natick was no longer active and said the company was not building subsea data centers, while retaining research lessons. The article also reports the two-year experiment and its server-failure comparison; this article relies on the commercial-status statement, not a direct comparison with present projects.

    Source role: Independent reporting with direct Microsoft comment

  12. [S12] Microsoft Research · Project page; publication date not stated

    Microsoft describes Project Natick as a research project to determine the feasibility of subsea data centers powered by offshore renewable energy. A completed research trial is not evidence that the company adopted it as a commercial product.

    Source role: Company research record

  13. [S13] NAVER Corporation · 2026-08-13

    NAVER confirmed an investment in Panthalassa and repeated the Ocean-3 2026 demonstration and 2027 commercialization targets. The release does not disclose the investment amount or establish a customer contract or project revenue for NAVER.

    Source role: Public-company investment announcement

  14. [S14] Fjord Journal · 2026-11-15

    FJ’s project profile records the factory milestone and Ocean-3 target separately; it labels NOAA station coordinates as a regional reference only and keeps corporate verification held.

    Source role: FJ project profile

  15. [S15] Fjord Journal · 2026-10-28 review

    FJ’s profile treats Western Passage as a proposed project; the permit application and municipal opposition are not construction authorization, and its permit outcome remains unverified.

    Source role: FJ project profile based on government records and reporting

  16. [S16] Federal Energy Regulatory Commission · Page reviewed 2026-11-15

    FERC explains that a preliminary permit allows the holder to study a proposed hydropower project and preserve priority to apply for a license; it does not authorize construction. An application accepted for filing is not itself an issued preliminary permit.

    Source role: Government explanation of preliminary-permit authority

Public-source analysis; no interviews, site visits, private contract review or independent cost audit. Targets are not treated as achieved milestones. This is infrastructure research, not a securities recommendation.

Following AI infrastructure? Bookmark Fjord Journal for sourced reporting on projects, power and capital.