Microsoft's Project Natick put a data center on the ocean floor in 2018, and the idea has refused to die since. Seabase's State of Subsea AI Infrastructure 2026 brief argues that the category has shifted from curiosity to a serious option for coastal AI workloads, though the gap between research pods and commercial fleets remains wide.
The brief separates what's actually been demonstrated from what's only been pitched. That distinction matters because headlines keep collapsing sealed research vessels, floating barges, nearshore submerged modules, and marine-cooled hybrids into one story. They are not one architecture, and they are not at the same stage of maturity.
The stakes are timing. AI inference demand is moving compute toward coastal metros and cable landings, because multi-step agent pipelines compound network delay in ways that make geography matter more than it did for training. Seabase's brief puts it plainly: training can concentrate where power is cheap, but inference cares where users are.
Rack power density is pushing everyone toward liquid cooling
Dense AI racks have moved from classic enterprise densities into tens of kilowatts, with high-density designs pushing toward and beyond roughly 80 to 100 kW in some planning envelopes. Those numbers come with caveats about generation and cooling method, and the brief treats them as living ranges rather than fixed facts.
Each accelerator generation raises rack power, thermal density, and networking requirements. That pressure is one reason liquid cooling and unusual siting ideas keep returning. The practical question for ocean systems is not whether GPUs get hotter. It is whether the facility model can service, power, and network them near demand.
On land, direct liquid cooling adoption is climbing. An S&P Global survey of enterprise data center decision-makers found 21% plan to shift to liquid cooling over the next year, up from 13% in 2024. Ocean concepts use seawater as the stable heat sink for those liquid loops, leaning on the resulting temperature differential and heat-transfer efficiency.
Coastal geography is the real differentiator
Inference, agents, and interactive applications increasingly care about geography. Multi-step pipelines compound network delay. That is why coastal metros, cable landings, and regional capacity keep appearing in serious plans, according to the brief.
Subsea fiber remains the backbone of international traffic. Landing density and route diversity shape which coastal regions can host low-latency, high-bandwidth AI infrastructure. Proximity to a cable landing station helps, but terrestrial routing and peering still decide outcomes.
China coastal and Hainan-linked projects show industrial and state interest beyond Western research demos, though independent verification of scale and customer use still varies by source. Subsea Cloud has been associated with submerged data-center concepts and public advocacy for underwater siting. Nautilus Data Technologies focuses on water-based cooling and marine or barge-adjacent facility concepts, a reminder that water-cooled and underwater are often blended in headlines.
The serviceability problem nobody has solved
The brief's open questions are the honest part. Can modules be manufactured and redeployed like a product line? What maintenance tempo is economically sustainable? How do sites secure firm power without isolating themselves from fiber? It's not clear these will hold up under commercial load, and the brief doesn't pretend otherwise.
Higher rack power increases the importance of heat-exchanger design, materials, biofouling management, and shore power quality. Accelerator refresh cycles are the harder problem: marine designs need to handle new GPUs without rare vessel campaigns.
Regulatory scrutiny covers thermal discharge, acoustics, materials, seabed disturbance, cable routes, and emergency recovery. Continuous monitoring is becoming part of the credibility test, not an optional add-on. Environmental telemetry shared with regulators or research partners is on Seabase's watchlist for the next quarter.
What to watch next
Seabase's quarterly watchlist prioritizes announcements that change architecture, regulation, or demonstrated operations: new underwater or floating installations with retrieval evidence, partnerships between compute operators and marine contractors, power or small modular reactor announcements tied to coastal AI campuses, cable system and landing-station expansions near demand centers, and national or municipal AI infrastructure strategies that mention marine siting.
Repeatable module retrieval and redeployment campaigns, published interconnect and fiber paths for nearshore sites, and customer-facing capacity claims with measurable SLAs are the signals that would separate real progress from concept marketing. Seabase itself is one participant in this category, emphasizing modular subsea infrastructure near coastal demand, power-source flexibility, and terrestrial fiber.
Why coastal metros keep winning the inference race
Coastal metros keep appearing in serious plans because latency compounds. A multi-step agent pipeline that calls a model, then a tool, then a model again turns small delays into visible lag. That is why regional capacity near cable landings keeps showing up in infrastructure roadmaps, and why marine siting near those same landings gets attention it would not have earned for training workloads.
Seabase's brief frames this as a geography problem rather than a technology one. The hardware matters less than where it sits relative to users and fiber. For operators evaluating ocean compute, the question is whether the marine advantage in cooling and land scarcity outweighs the cost of servicing equipment at sea.
Related coverage
Related coverage
For more on data center infrastructure trends, see S&P Global's 2026 data center services report and the Seabase industry brief.
Regulators will likely ask for continuous environmental telemetry before permitting multi-year operations, and operators who cannot provide it will struggle to move beyond pilot projects. That telemetry requirement is one more reason the category's progress will be measured in quarters, not headlines.
The open question is which workloads actually benefit from marine siting versus land edge sites. Seabase's brief lists it among the unresolved questions, and the answer will likely depend on latency sensitivity, power availability, and maintenance budgets rather than on cooling alone. Operators weighing these tradeoffs should expect to spend more on retrieval campaigns and environmental monitoring than on the servers themselves in the first year.
The next twelve months should clarify whether subsea compute earns a permanent place in the AI infrastructure stack or remains a niche option for coastal operators with unusual constraints.
Internal links: cloud and edge computing coverage and archive.