Underwater and Off-Grid: China's Bet on Subsea Data Centers
Submerged, seawater-cooled data centers sound like science fiction. China is building them at scale. Here's what's real, what's not, and whether the ocean is actually a viable answer to AI's cooling crisis.
The cooling problem in modern data centers is not subtle. An AI training cluster running at 100 kilowatts per rack generates enough heat to be genuinely dangerous if you get the thermal management wrong. The industry's conventional answer has been more chillers, more cooling towers, more water, more land. Northern Virginia is already maxing out the grid. Singapore put a moratorium on new builds for two years because of power and water concerns. The hyperscale operators are running out of easy options, and everyone knows it.
So when you hear that China has sunk data center modules into the ocean and is running them commercially, your first instinct might be to file it under "interesting but probably not serious." We had the same instinct. We've been watching this space long enough to know that "revolutionary cooling breakthrough" announcements usually age about as well as a fish left in the sun. But subsea data centers are different. They're real, they're deployed, they're generating commercial revenue, and the physics behind them are sound enough that Microsoft, the US Navy, and now multiple Chinese operators have independently concluded they're worth serious capital.
The question worth answering isn't whether this technology exists. It's whether it's ready, scalable, and genuinely better, or whether it's an elaborate solution to a problem that better liquid cooling on land could handle more cheaply.
The Problem It's Actually Solving
To understand why anyone would sink computing equipment into the ocean, you have to start with the thermal economics of modern AI workloads. A standard enterprise server from 2015 ran at maybe 5 to 10 kilowatts per rack. Today's GPU dense AI training racks are pushing 50, 80, sometimes 100 kilowatts per rack. The roadmap from NVIDIA and AMD suggests that number is heading toward 120 to 150 kilowatts per rack within the next two to three years.
Traditional air cooling hits a practical wall somewhere around 20 to 30 kilowatts per rack. Above that, you're either wasting enormous amounts of energy pushing cold air at a problem that air can't fully solve, or you're moving to some form of liquid cooling. Liquid cooling works well but adds complexity, maintenance requirements, and in the case of water cooled systems, real pressure on freshwater resources. A large hyperscale facility can consume millions of gallons of water annually for cooling, which is not an abstract concern when you're siting facilities in water stressed regions.
China's data centers collectively consumed 250 billion kWh of electricity in 2023, of which roughly 80 billion kWh went purely to cooling, and that cooling water demand is projected to reach 248.35 billion liters annually by 2026, growing at an 8.6% CAGR. ESG News The ocean solves several of these problems simultaneously. Seawater is an effectively infinite cooling resource. It's already cold, especially at depth. It doesn't need to be pumped from a municipal supply, treated, or returned to a river system. The ocean absorbs heat at enormous scale without any meaningful temperature increase at the facility level. And the ocean doesn't cost anything per liter the way freshwater infrastructure does.
China DC Electricity: Cooling vs. Total vs. Subsea Potential
Annual power consumption (billion kWh), current baseline vs. hypothetical nationwide subsea cooling scenario
Source: New Atlas (via Autonocion), 2024. Subsea savings figure is a hypothetical scenario, not a confirmed projection.
What subsea deployments add on top of passive cooling is the physical isolation of the pressure vessel itself. The sealed, nitrogen filled module design used in projects like Microsoft's Project Natick eliminates humidity, one of the primary causes of hardware failure in conventional facilities. In Natick's data, the failure rate of equipment inside the sealed subsea module was roughly one eighth the rate observed in comparable land based deployments. That's not a marginal improvement. That's a fundamentally different reliability profile.
How the Technology Actually Works
The engineering of a subsea data center is not as exotic as it sounds, once you break it down. The basic architecture is a pressure vessel, essentially a sealed steel cylinder or module, filled with dry nitrogen or another inert atmosphere, with server racks mounted inside. The module is designed to be pressure resistant at depth, watertight, and capable of being lowered to the seafloor and retrieved for maintenance.
Heat from the servers transfers through heat exchangers to the outside of the vessel. Seawater flowing past the hull absorbs that heat passively in some designs, or through dedicated heat exchange loops in others. The key innovation is the thermal pathway: instead of spending energy on chillers and cooling towers, you're leveraging the thermal mass and temperature of the surrounding ocean. The deeper you go, the colder the water, but you also need more robust pressure engineering. The two deployments with published depths have struck that balance very differently. Microsoft's Natick Northern Isles lay at 117 feet, about 36 metres, on the seafloor off Orkney. HiCloud's filing for the Lingang project gives a mean water depth of about six metres, with the pod standing on piles and most of its structure above the waterline; Chinese coverage of the launch generally says about ten metres.
Power delivery is the other core engineering challenge. You need to run high voltage cables from shore or from an offshore platform to the module. This is well understood technology: offshore wind and oil and gas have been running subsea power cables for decades. The data connections run alongside the power cables, using fiber optic lines that are again standard offshore infrastructure. The novel part isn't the cable technology. It's the integration of data center operations at the other end of those cables.
Maintenance is where the honest engineering conversation gets complicated. Land based data centers have technicians walking the floor continuously. Subsea modules don't. The design philosophy has to shift from reactive repair to planned replacement. In Natick, Microsoft designed the modules to run for five years without human intervention, then be retrieved and either serviced or decommissioned on a replacement cycle. This works if the hardware failure rates are low enough, which the sealed nitrogen environment helps with, but it changes the operational model fundamentally. You're not running a data center the way you'd run a building. You're running it more like a satellite: plan for the full mission duration, accept that you can't fix everything in flight, and design accordingly.
China's Deployments: What's Actually in the Water
This is where China's approach diverges from the exploratory research phase that Microsoft represented with Natick. China hasn't just tested a module. It's commercializing the concept at scale, and doing so with the urgency that comes from a government that has decided this is a strategic infrastructure priority.
The Shanghai Lingang project gives the clearest window into what China's subsea buildout actually looks like in hardware terms. HiCloud's marine-use assessment filing to the Lingang Special Area authority describes a single vertical data pod — 14 metres in diameter, 17 metres tall, 402 tonnes of steel and equipment — holding 192 IT racks across four decks at an installed capacity of 2.3MW. It stands on four 2.0-metre steel piles in about six metres of water, 13 kilometres off the Nanhuizui shoreline, with 24.2 metres of structure above mean sea level. HiCloud — Lingang marine-use assessment filing The 24MW in wide circulation is the total planned across two phases rather than what is built: the Ministry of Transport's launch notice of 14 February 2026 gives 24MW in two phases, with phase one at 2.3MW and total project investment of ¥1.6 billion. PRC Ministry of Transport The efficiency numbers are the part worth pausing on: Lingang saves an estimated 61 million kWh per year compared to an equivalent land-based facility, cuts electricity consumption by 22.8%, reduces land use by 90%, consumes zero liters of freshwater, and is designed for a PUE below 1.15 — a design target in the filing, not a measured figure. ESG News Ninety-five percent of its power comes from green sources. Total build cost comes in at no more than 50% of what a comparable land-based data center would require. CGTN These are not incremental improvements. They're a different cost and resource structure entirely.
What makes Hainan the right location is a combination of factors that are specific to China's geography and grid situation. The South China Sea off Hainan offers relatively shallow coastal waters, warm surface temperatures (which means you need to go to modest depth to access meaningfully cooler water), and proximity to Hainan's growing technology sector and the free trade zone that the central government has been developing there. The island also has renewable energy ambitions: Hainan has targeted 50 percent renewable power by 2025, which pairs well with the energy efficiency argument for subsea cooling.
The project's commercial rationale Data Center Knowledge is partly about cooling efficiency and partly about regulatory arbitrage. China's inland data center capacity is increasingly constrained by power quotas imposed by provincial governments trying to manage grid load. Coastal and offshore deployments exist in a different regulatory space, and early movers are taking advantage of that ambiguity.
Beyond Highlander, the Chinese government has backed subsea data center research through the State Oceanic Administration and through funding directed at universities and state owned enterprises studying offshore compute infrastructure. State level endorsement Global Times has accelerated the pace of development considerably. When the Chinese government identifies a strategic technology, the timeline from research to deployment compresses dramatically compared to purely commercial development cycles in the West.
There's also a military and dual use dimension worth acknowledging directly. Offshore computing infrastructure has obvious applications beyond commercial cloud services. Subsea nodes can support maritime surveillance, autonomous undersea systems, and communications infrastructure that doesn't depend on land based networks. This isn't a reason to dismiss the commercial deployments, but it's context for understanding why the Chinese state is particularly interested in this technology relative to purely market driven actors.
The Energy and Water Advantages: How Real Are They
The Power Usage Effectiveness (PUE) figures claimed for subsea deployments are genuinely compelling, if you believe them. Conventional land based data centers average a PUE somewhere between 1.3 and 1.6, meaning they use 30 to 60 percent more total power than their IT equipment alone consumes, the overhead going primarily to cooling. Best in class hyperscale facilities in cold climates achieve PUEs around 1.1 to 1.15. HiCloud's filing gives the Lingang pod a design target of PUE below 1.15 — 系统 PUE 小于 1.15 — and the Ministry of Transport's launch notice describes it the same way, as a design figure. No audited measurement of this facility's PUE has been published, and the operator co-authored the Chinese industry standard the project is assessed against. Some subsea designs claim approaches toward 1.07; the same filing cites 1.076 as a general figure for underwater data centres rather than a measurement of this pod. The passive cooling load on the power bill really is much smaller, which is what makes the target plausible — but a target is what it is.
The water consumption argument is even cleaner. A land based facility cooling with evaporative towers consumes between 1.5 and 3 liters of freshwater per kilowatt hour of IT load. The Lingang subsea facility uses zero liters of freshwater ESG News, the ocean is the cooling medium, and it stays in the ocean. China's data center sector is on track to consume 248.35 billion liters of freshwater annually for cooling by 2026, growing at an 8.6% CAGR Mordor Intelligence, which puts the zero-freshwater claim in sharp relief. For regions under water stress, or for operators trying to meet sustainability commitments, this is not a minor benefit. It's potentially the decisive factor.
The energy argument is a bit more nuanced. You do save energy on cooling. But you spend energy on the subsea cable infrastructure, on the pressurization and monitoring systems, and on the logistics of deployment and retrieval. The net energy math still favors subsea in most analyses, but the margin narrows when you account for the full infrastructure cost. The honest number at Lingang is a 22.8% reduction in electricity consumption compared to an equivalent land-based facility ESG News, meaningful, but not transformative on its own.
Subsea vs. Land DC: Key Efficiency Metrics at Shanghai Lingang
Percentage improvement of subsea deployment vs. equivalent land-based facility
Source: ESG News; New Atlas (via Autonocion), 2024. Figures reflect Shanghai Lingang demo-phase deployment.
The Engineering Risks Nobody Talks About in the Press Releases
We've seen enough "revolutionary" data center announcements to know that the limitations section is where the real story lives. Subsea data centers have genuine ones.
The corrosion environment is brutal. Seawater is one of the most hostile environments for metal and electronics that exists. The pressure vessel protects the equipment inside, but the exterior of the vessel, the cable connections, the heat exchange surfaces, and all the penetrations through the hull are fighting a continuous battle against salt, biological fouling, and electrochemical corrosion. This is solvable with the right materials and coatings, offshore energy has been doing it for decades, but it adds cost and requires rigorous inspection cycles that aren't free.
The maintenance model is the deeper challenge. Land based facilities can respond to a hardware failure within hours. A subsea module requires a vessel, a dive team or ROV, and a recovery operation that might take days and cost hundreds of thousands of dollars to execute. The sealed module philosophy works only if the failure rate inside is low enough that planned five year retrieval cycles are sufficient. Microsoft's Natick data is encouraging on this point, but Natick ran for two years in Scottish waters with relatively modest compute density. Running AI training workloads at 80 kilowatts per rack for five years continuously is a different stress test.
Connectivity redundancy is another honest concern. Land based facilities can have multiple fiber paths, multiple grid connections, and multiple physical access routes. A subsea module on a single cable run has limited redundancy. A ship dragging an anchor across the cable (which happens more often than you'd hope to fiber infrastructure in busy shipping lanes) can take out the connection entirely. The South China Sea has some of the world's busiest maritime traffic. Routing cables around shipping lanes adds distance and cost, and doesn't eliminate the risk.
Finally, the decommissioning question has no good answer yet. What do you do with a subsea module at end of life? Retrieving it is expensive. Leaving it on the seafloor is environmentally and legally problematic. The industry hasn't developed a standard answer, and regulators in most jurisdictions haven't finished thinking through the rules. This is a real long term liability that operators are carrying, even if it's not showing up in the current economics.
Is This Actually Ready for Production AI Workloads?
Here's our honest read after tracking this space carefully.
The technology works. The physics are sound, the basic engineering is proven (if not yet at scale), and the reliability data from Natick and the early Chinese deployments is genuinely encouraging. This is not vaporware. It's not a science project. It's a real operational approach that is generating commercial revenue today.
But "works" and "ready for your production AI training cluster" are not the same sentence yet.
For inference workloads, where latency matters, coastal subsea nodes actually have an interesting advantage: you can put compute closer to population centers that sit near coastlines without needing urban land. Southeast Asia, coastal China, the US East Coast and West Coast all have dense population within reach of reasonably shallow coastal waters. A low power inference node in a subsea module a few kilometers offshore could serve urban latency requirements with better cooling economics than an equivalent land based facility in an expensive urban market.
For training workloads, which is where the serious heat density challenge lives, subsea deployments aren't the primary answer yet. The combination of high rack density, continuous high power operation, limited maintenance access, and the need for rapid hardware upgrades as GPU generations evolve every 18 to 24 months doesn't fit the subsea operational model cleanly. You don't want to be retrieving a module to swap H100s for B200s on the ocean floor on a rushed timeline.
The more likely near term role of subsea infrastructure in the AI buildout is as a complementary capacity layer rather than a replacement for land based hyperscale. China's approach, treating subsea as an overflow and edge capacity solution for coastal markets while keeping core training infrastructure onshore, is probably the right framing for the next five years.
What China Gets Right About This Bet
The strategic logic of China's subsea investment is coherent even if the technology isn't fully mature. China is building AI infrastructure under real constraints: power quotas, water stress in many inland provinces, land costs in tier one cities, and political pressure to keep AI compute within national jurisdiction. Subsea deployments address several of these simultaneously.
More importantly, China is playing the long game. The global underwater data center market sits at $3.2 billion in 2025 and is projected to reach $14.8 billion by 2034 at an 18.6% CAGR Dataintelo, and the country that develops the operational expertise, the regulatory frameworks, the cable installation capability, and the module manufacturing supply chain will be positioned to capture a disproportionate share of that growth. Even if today's deployments are more demonstration than production scale, the institutional knowledge being built is genuinely valuable.
The broader engineering research base IEEE Spectrum that Microsoft's Natick project generated has been largely published and is freely available. China's developers have studied it carefully, and the Hainan deployment shows clear design influences from Natick's learnings. This is the advantage of being a fast follower with state backing: you inherit the proof of concept research and move directly to commercialization.
The West's response to this has been mostly silence. Microsoft shelved Natick after completing the research phase. There are startups in the US and Europe working on subsea concepts, but none with China's combination of state support, coastal geography, and manufacturing scale. If subsea data centers become a meaningful part of the AI infrastructure stack over the next decade, China will have a substantial head start.
The Verdict: Science Project or Genuine Infrastructure?
We don't usually deal in absolutes, but this one's worth stating clearly. Subsea data centers are not a science project. They're not the cooling breakthrough that replaces everything that came before. They are a genuine, deployable, commercially viable approach to a specific subset of the cooling problem, particularly well suited to coastal markets, water stressed regions, and edge compute applications.
The honest limitations, maintenance complexity, rack density ceilings, connectivity redundancy, and decommissioning uncertainty, are real engineering challenges that will take years of operational experience to fully work through. Anyone telling you these are solved is selling you something.
But the direction of travel is clear. Heat density is going up. Freshwater constraints are getting worse: China's cooling water demand alone is compounding at 8.6% annually toward a quarter-trillion liters per year. Land costs in desirable markets are rising. The ocean isn't going anywhere. China has made a strategic bet that the combination of these pressures will eventually make subsea the obvious answer for coastal markets, and they're building the capability to be ready when that tipping point arrives.
Whether they're right about the timing is the open question. Whether the physics and engineering eventually support large scale subsea AI infrastructure, we're confident they do. The math works. The technology exists. The only question is how long it takes for the operational model to mature enough to handle the full brutality of modern AI workloads running continuously at densities that would have seemed absurd five years ago.
That's not a science project. That's an infrastructure bet with a credible thesis. We'd rather be tracking it honestly than dismissing it, or overselling it.