Underwater Data Centers in 2026: How Highlander Hailanyun, Subsea Cloud, NetworkOcean, and Microsoft's Project Natick Legacy Are Cooling AI Compute With the Ocean
- Internet Pros Team
- May 28, 2026
- AI & Technology
For two decades, every solution to the AI data-center cooling problem has assumed the data center sits on land. In 2026, that assumption is finally cracking. Highlander's Hailanyun off Hainan and Shanghai, Subsea Cloud's Jules Verne pods, the NetworkOcean Bay Area pilots, and the operational lessons from Microsoft's Project Natick are converging on one provocative idea — submerge the racks. With Blackwell-class NVIDIA GPU servers crossing 130 kW per rack, freshwater drought killing cooling-tower expansion, and US grid interconnect queues stretching to 2030, sinking compute into the ocean has gone from quirky moonshot to a credible class of infrastructure backed by working megawatts.
Why the Ocean Is Suddenly Interesting to Hyperscalers
The math is harsh and getting harsher. A single NVIDIA GB200 NVL72 rack draws roughly 120-132 kW; Rubin-class racks coming in 2026-2027 are projected to push past 250 kW. Air alone cannot move that heat — even direct-to-chip liquid cooling has to dump waste heat somewhere. On land, that "somewhere" is typically evaporated water (cooling towers consume 1.8-2 million liters per megawatt per year) or chillers that recompress refrigerant at significant electrical cost. Both options have run into walls in 2026: drought-stricken regions like Arizona and Spain have begun denying permits, and Virginia's Loudoun County paused new data-center approvals while the grid catches up.
The ocean offers something land cannot: an effectively infinite, free, 4-15°C thermal reservoir that is uniformly cold below 100 meters anywhere on Earth. A sealed pressure vessel submerged in that reservoir can run server inlet temperatures within ASHRAE A2 limits using a closed-loop heat exchanger, with zero freshwater consumption and an annual PUE near 1.05 — the same neighborhood as the best Nordic free-cooled sites, but available anywhere with a coastline.
"We are not arguing the ocean is a better data center than a perfectly designed land facility in Iceland. We are arguing it is a better data center than the only land option you can actually permit, power, and water in the next 24 months — and that bar is the one that matters in 2026."
Who Is Actually Building Underwater Data Centers Today
| Project | Approach | 2026 Status |
|---|---|---|
| Highlander Hailanyun (China) | Steel pressure-vessel modules dropped 35 m down off Hainan and (later) the Lingshan Islands near Shanghai. Each module integrates with offshore wind power and is connected by subsea power and fiber cables to a coastal landing station. The flagship project advertises a PUE of 1.15 at scale. | The world's largest commercial UDC deployment by a wide margin — multiple modules in the water off Hainan, the Shanghai-adjacent expansion underway, and additional sites announced for Sanya and the Bohai Sea. |
| Subsea Cloud (USA) | Modular "Jules Verne" pods filled with a dielectric fluid in which servers are fully immersed. The pods are deployed at depths of 100-300 m, sized for hyperscaler-grade colocation, and engineered for 12-year unattended operation. | Commercial deployments in the Gulf of Mexico and the Pacific Northwest, plus partnership conversations with offshore wind operators for North Sea and Asia-Pacific sites. |
| NetworkOcean (USA) | Small-footprint, near-shore capsules targeting AI inference at the edge. The first units were piloted in San Francisco Bay, deliberately close to compute-hungry hyperscaler campuses to demonstrate sub-millisecond latency to land-based front-ends. | Y Combinator-backed startup running US Bay-Area pilots, with subsequent deployments planned at coastal data-center hubs near Los Angeles, Singapore, and the UK south coast. |
| Microsoft Project Natick (legacy) | The original 2018 Orkney Islands experiment that sank 864 servers in a nitrogen-filled steel cylinder for two years. Not a commercial product, but the 1/8th server-failure rate vs. an equivalent land control group is still the most cited datum in the field. | Officially retired as a research program; the engineering lessons live on inside Microsoft's liquid-cooling roadmap and across nearly every UDC startup founded since. |
| NTT, HydroDC, Atlantis Computing (concept & pilot) | A growing list of telcos and ocean-engineering firms exploring floating, near-shore, and bottom-mounted concepts — some paired with floating offshore wind or wave-energy generation. | Mostly pre-commercial pilots and design studies through 2026, but most carry serious engineering staff and credible offtake conversations with AI infrastructure customers. |
The Four Engineering Wins
Strip away the novelty and the case for underwater data centers rests on four hard engineering wins that are difficult or impossible to replicate on land:
Free, Year-Round Cooling
Sub-100 m ocean water sits at 4-15°C regardless of latitude or season. A closed-loop seawater heat exchanger can keep server inlet temperatures inside the safe ASHRAE envelope without chillers — driving PUE toward 1.05 and slashing the electrical bill that pays for cooling, often 30-40% of a land data center's annual opex.
Zero Freshwater Consumption
Land data centers can drink 1.5-2 million liters of potable water per megawatt per year through evaporative cooling. As regulators in drought-stressed regions cap or revoke water permits, a sealed UDC that uses zero potable water becomes the only deployable option in many counties.
Higher Hardware Reliability
Project Natick's sealed, nitrogen-filled enclosure produced an 8x lower server-failure rate than its dry-land control. Without oxygen, humidity, dust, or human hands, the dominant failure modes of a data center largely vanish. Twelve-year unattended service intervals are no longer fanciful.
Speed and Permittability
A prefabricated UDC pod can be welded on shore, towed to site, and energized in under 12 weeks. Compare that to 3-5 years for a permitted, transmission-connected hyperscale land campus. In a market where AI capacity is measured in megawatts-per-quarter, the deployment-time delta dwarfs the engineering risk.
The Honest Problems Nobody Has Fully Solved
Marine impact. Discharging warm water near coral, seagrass, or shellfish habitat is a legitimate concern. Hailanyun has published thermal-plume monitoring data showing localized rises of 1-2°C within meters of the discharge — small but non-zero. Permitting under the US EPA NPDES regime, the EU's Marine Strategy Framework Directive, and China's coastal-zone management rules is still being written for this use case.
Biofouling and corrosion. The ocean is a relentless industrial environment. Anti-fouling coatings, sacrificial anodes, copper-nickel hulls, and ROV-based scrubbing have all been borrowed from offshore oil and submarine cable engineering, but operating-cost data over a true 10-year cycle is still thin.
Service and repair. A failed rack at 35 m depth is not a forklift away. The economics only work if hardware reliability is high enough that you accept losing a small percentage of servers between scheduled retrievals, much like a hyperscaler accepts dead nodes in a hyperscale fleet. Predictive maintenance and digital twins of the pod become non-negotiable.
Jurisdiction. Whose laws apply 12 nautical miles offshore? Which country's data-residency rules govern a pod in international waters connected by fiber to two continents? These are unsettled questions that contract lawyers, regulators, and the EU AI Act enforcement teams will spend years untangling.
What Underwater Data Centers Mean for the AI Infrastructure Stack
- Cooling is now an architectural decision, not a facilities one. The choice between air, immersion, direct-to-chip, and seawater closed-loop is increasingly made at the silicon roadmap level — Blackwell, Rubin, MI400, and Trainium 3 all assume liquid by default.
- Coastal real estate becomes strategic. Expect competition for shoreline plots near offshore wind farms, subsea cable landing stations, and metropolitan AI-inference markets — the new digital ports of the AI economy.
- The water-energy nexus is now an AI story. The same data center that previously fought a town over freshwater rights can sidestep the fight by going subsea. Regulators and ESG teams should expect to be asked the question.
- Edge AI inference may live offshore. NetworkOcean's near-shore model points to a future where latency-sensitive inference is hosted in capsules a few miles from the coast rather than in inland mega-campuses.
- The deployment cycle compresses. Twelve-week pods reshape the cap-ex calendar. AI customers will increasingly demand "modular megawatts" available in quarters, not years — and underwater is one of the few credible answers.
The Bottom Line
Underwater data centers are not going to replace the hyperscale land campus. The grid-connected, terrestrial, multi-hundred-megawatt site will remain the workhorse of the AI economy for the foreseeable future. But the moment that workhorse cannot be built fast enough, powered cleanly enough, or cooled cheaply enough to meet the demand curve of frontier AI, an alternative becomes inevitable. In 2026 that alternative has names, working units, and revenue.
For business and infrastructure leaders the message is straightforward: when you evaluate cloud regions, AI inference partners, or your own colocation strategy, "underwater" is no longer a tech-demo footnote — it is a deployable class of capacity with real PUE numbers, real reliability data, and a permitting story increasingly preferable to the one onshore. The companies that learn to procure, operate, and govern subsea compute over the next 24 months will have access to megawatts that their competitors will spend years trying to permit. The AI race is, quietly, becoming a maritime one.
