Underwater Data Centers: How the Ocean Became AI’s Newest Cooling System

It is 2 a.m. on a July night inside a hyperscale data hall in Ashburn, Virginia, and the noise never stops. Rows of steel racks stretch into the dark, each one drawing enough electricity to power a dozen houses, and above them a wall of computer-room air handlers roars at a pitch that makes conversation impossible without a headset. An engineer doing a walkthrough presses a palm against a rack door and feels heat radiating through the metal even with the cooling system running flat out. Every one of the GPUs behind that door is trying to shed roughly the same heat as a household toaster, multiplied by tens of thousands, and someone has to carry all of it away before the silicon throttles or dies. This is the unglamorous physical reality behind every headline about artificial intelligence: underwater data centers exist because moving that heat has become one of computing’s hardest and most expensive engineering problems, and the ocean is one of the few places on Earth that offers a genuinely stable, low-cost place to put it.

Data centers have always generated heat — that part is not new. What changed is the density. A rack of servers in the early 2000s might have drawn 2-5 kilowatts; a rack of NVIDIA GB200 NVL72 hardware built for AI training draws on the order of 120-132 kilowatts, and every one of those kilowatts eventually leaves the chips as heat that has to go somewhere. Air can carry heat away, but air is a poor conductor, so moving enough of it fast enough requires large fans, chilled water loops, and industrial refrigeration — machinery that itself burns electricity, sometimes 30-45% of a facility’s total power draw on top of the computing load itself. That overhead is what data-center engineers track obsessively as Power Usage Effectiveness (PUE), the ratio of total facility power to power actually reaching the servers. The industry-wide average, per the Uptime Institute’s 2025 Global Data Center Survey, has been stuck near 1.54 for six straight years despite enormous investment in efficiency — meaning for every unit of energy a server uses computing, roughly another half-unit is spent just keeping it cool and running the building around it.

Around 2014, engineers inside Microsoft’s data-center group ran a back-of-envelope calculation that reframed the problem entirely: nearly half the world’s population lives within about 200 kilometers of a coastline, and the ocean near most of those coastlines sits at a cold, remarkably stable temperature just a few dozen meters down — no seasonal swings, no heat waves, no chillers required. If you could seal a data center inside a pressure-rated capsule and lower it to the seabed, the surrounding seawater itself becomes the cooling system, for free, indefinitely. That idea became Project Natick, and it is the direct ancestor of everything this article covers.

It took over a decade for that back-of-envelope calculation to turn into a paying commercial service, and the gap between those two moments — a 2014 internal proposal and a 2025 revenue-generating facility off Hainan — is itself the story most coverage skips over. In between sat years of unglamorous, mostly unpublicized work: pressure-testing hulls, proving corrosion-resistant coatings actually hold up after months underwater, negotiating seabed rights and subsea cable landings, and waiting two full years just to gather enough server-failure data to say anything statistically meaningful about reliability. None of that is dramatic enough to headline a news story, which is exactly why a guide like this one, built to sit with the topic rather than chase a single announcement, is useful: it can show the slow engineering underneath the fast headlines.

This guide separates four different things that get blurred together in most coverage of this topic: proven research findings (what Project Natick’s two-year deployment actually measured), commercial deployments (Highlander’s Hainan facility, which is generating revenue today), active pilots and startups (Subsea Cloud, Nautilus Data Technologies, and others still validating their approach), and engineering proposals or future scenarios that have not been built yet. It does not claim underwater data centers will replace conventional facilities — no credible engineering source makes that claim, and neither do we. Where this guide cites a number, it also tells you which of those four categories it came from, so you can judge how much weight to put on it yourself rather than taking every figure in this space at face value.

Coverage of this topic routinely blurs four related-but-different ideas together, which is worth untangling once, clearly, before going further. An underwater data center is a fully sealed pod submerged on or near the seabed, using surrounding seawater as its cooling medium — Highlander’s Hainan facility and Project Natick are the two examples this article centers on. A floating data center, like Nautilus’s Stockton 1, sits on a barge on the water’s surface and pumps in water for cooling rather than submerging the whole facility. Immersion cooling submerges individual servers in a tank of dielectric fluid, almost always on land — it shares underwater deployment’s core physics (liquid conducts heat better than air) without any of the marine engineering. And edge computing is a deployment strategy, not a cooling method at all: it means placing compute physically close to users to cut latency, which an underwater pod can serve as one option among several, alongside small land-based edge facilities. Treating any two of these as interchangeable is the single most common source of confusion in mainstream coverage of this space.

The first purpose-built data centers date to the 1940s and 50s, when room-sized computers like ENIAC needed dedicated space, power, and primitive air handling just to survive their own vacuum tubes. For decades afterward, “the computer room” was a modest, single-organization affair — a bank’s mainframe, a university’s research cluster — and cooling meant little more than commercial air conditioning.

That changed with the commercial internet. As dot-com-era companies raced to get online in the late 1990s, purpose-built colocation facilities emerged to host servers for many customers at once, and cooling became an engineering discipline in its own right rather than an afterthought. The 2006 launch of Amazon Web Services, followed by Google Cloud and Microsoft Azure, created the modern cloud computing era: instead of every company running its own server room, a handful of hyperscale operators built enormous, hyper-efficient campuses and rented out capacity. Scale brought real efficiency gains — a hyperscale facility could justify custom cooling engineering that a small server room never could — but it also concentrated heat generation into single sites drawing hundreds of megawatts.

Then came the current inflection point. The generative-AI boom that began accelerating around 2022-2023 changed the hardware profile of a typical new data center almost overnight. Training and running large language models rewards raw parallel throughput, which is what GPUs and dedicated AI accelerators deliver — and that throughput comes bundled with power density that older air-cooling designs were never built for. The International Energy Agency’s Electricity 2026 report found data-center electricity demand grew about 17% globally in 2025, with AI-focused facilities alone surging roughly 50%, and projected total data-center consumption to more than double to around 945 terawatt-hours by 2030 — more than Japan’s entire current electricity use. That is the demand curve underwater data centers, immersion cooling, and every other alternative cooling approach are responding to.

It’s worth being precise about what changed and what didn’t. The underlying physics of heat removal hasn’t changed at all — liquids have always conducted heat better than air, and the ocean has always been cold and stable near most coastlines. What changed is the economics: when a rack drew 2-5 kW, the modest efficiency gains from exotic cooling weren’t worth the engineering complexity, so nearly everyone used air. When a rack draws 120+ kW, air cooling stops being merely inefficient and becomes physically incapable of removing the heat fast enough, which converts liquid, immersion, and underwater cooling from optional efficiency plays into practical necessities for a growing share of new deployments. That threshold-crossing, more than any single breakthrough, is why this topic has moved from a research curiosity to a live commercial category within about a decade.

Coverage of this topic tends to lump every underwater-adjacent project together, which flattens real differences in scale, maturity, and evidence quality. The six entries below are ranked by demonstrated real-world impact — a commercially operating facility outranks a well-funded pilot, which outranks a design that exists mainly on paper — not by press coverage or novelty. Two, Highlander and Project Natick, anchor the entire field and are worth comparing directly.

Simplified cutaway of a sealed underwater data-center pod: server racks, internal heat exchanger, and the seawater loop that replaces conventional air conditioning. Diagram: AiTimeline.

Strip away the novelty and an underwater data center is a fairly conventional cooling problem solved with an unconventional heat sink. Nothing about the servers themselves changes — the innovation is entirely in how heat gets from the chip to somewhere it can dissipate safely.

1. Manufacture and load the pressure vessel: racks, servers, networking, and power gear are installed and tested inside a sealed steel cylinder at a factory, not on site — the same “build it complete, then ship it” logic as the modular data centers of the 2008-2010 period.
2. Purge the atmosphere: the interior is filled with dry nitrogen instead of ordinary air, removing the oxygen and humidity that cause corrosion and static-related failures over years of unattended operation.
3. Seal and pressure-test: the vessel is sealed to withstand external water pressure at its target depth (Natick’s Orkney pod: about 117 feet down; Highlander’s Hainan modules: about 35 meters; Subsea Cloud has pressure-tested pods rated to 3,000 meters) with margin for storms and current.
4. Tow and lower to the seabed or a subsea frame: a barge or support vessel positions the pod, then lowers it onto a pre-installed base structure on or near the seafloor, typically within a few kilometers of shore.
5. Connect subsea power and fiber: armored cables — the same class of infrastructure used for offshore wind farms and short-haul subsea telecom links — deliver electricity in and data out, landing at a coastal substation and network point of presence.
6. Let seawater do the cooling: inside the pod, heat from the chips transfers to an internal coolant loop, which passes through a heat exchanger in contact with the outer hull; the surrounding seawater — stable and cold at depth — carries that heat away continuously, with no fans, chillers, or compressors required.
7. Operate lights-out: once sealed, there is no human access without hauling the pod back to the surface. Monitoring, remote diagnostics, and workload management all happen over the network connection, exactly like managing a remote server rack, just with no possibility of a technician walking over to check a blinking light.
8. Retrieve, refresh, and redeploy: when hardware needs replacing — typically pegged to a 5-10 year refresh cycle, similar to land facilities — the entire pod is raised, opened at a facility on land, refitted, and can be resealed and redeployed rather than demolished.

None of these organizations operate in isolation. The IEA and Uptime Institute supply the independent measurement that keeps operator claims honest; the Open Compute Project turns individual vendors’ cooling engineering into shared, interoperable standards; NVIDIA, AMD, and Intel build the chips whose power density created the problem in the first place; and Google Cloud, Microsoft Azure, and AWS represent the land-based hyperscale default that underwater deployment is measured against, not competing with directly. Reading coverage of this topic with that structure in mind — who measures, who standardizes, who builds the heat source, who runs the baseline — makes it far easier to tell an independent finding from a vendor claim.

Every cooling method in data-center engineering solves the same basic problem — get heat out of silicon faster than the silicon makes it — with a different trade-off between cost, density, and complexity. Air cooling, still the default for most of the world’s data centers, blows chilled air across server components; it’s cheap, simple, and well understood, but air is a poor heat conductor, so it runs out of headroom somewhere around 20-30 kW per rack. Liquid cooling routes a coolant — usually water or a water-glycol mix — directly to cold plates mounted on the hottest components (CPUs, GPUs), removing heat far more efficiently and supporting the 100+ kW racks that modern AI hardware requires. Immersion cooling goes a step further, submerging entire servers in a dielectric fluid that doesn’t conduct electricity, eliminating fans altogether and handling even higher densities. Underwater deployment is best understood as immersion cooling’s logical extreme: instead of a tank of fluid on a data-center floor, the entire pod sits inside the largest, most thermally stable fluid reservoir on the planet.

If there is one metric that explains why any of this matters commercially, it’s PUE — total facility power divided by power actually reaching IT equipment. A PUE of 2.0 means a data center burns as much energy on cooling, lighting, and power conversion as it does on computing itself; a PUE of 1.1 means almost all of it goes to computing. The Uptime Institute’s 2025 global survey put the industry-wide weighted average at 1.54, essentially unchanged for six consecutive years despite heavy investment — a sign that air-cooling-era facilities are running into a hard efficiency ceiling. Underwater and immersion systems report substantially better figures in pilot conditions (commonly cited in the 1.1-1.2 range), because removing mechanical chillers removes their entire energy overhead, not just a fraction of it. Highlander’s own claim of roughly 90% lower cooling-specific energy use for its Hainan facility is consistent with that mechanism, though it’s a vendor figure, not yet an independently audited one.

The parts of an underwater data center that have nothing to do with computing are, in many ways, the harder engineering problems. The pressure hull has to survive not just static water pressure at depth but decades of wave loading, storm surge, and potential seismic activity, using corrosion-resistant alloys and coatings borrowed directly from offshore oil-and-gas and naval engineering. Biofouling — the buildup of barnacles, algae, and marine organisms on any submerged surface — can reduce heat-transfer efficiency at the hull over time, requiring either resistant coatings or periodic cleaning, and researchers studying Natick’s retrieved pod found measurable biological growth despite protective coatings. Subsea power and fiber cables are the same class of infrastructure used for offshore wind farms and short-hop telecom links — armored, buried where possible near shore, and routed to avoid shipping lanes and anchor zones. None of this is exotic engineering in isolation; offshore wind, subsea telecom, and undersea oil infrastructure have used these techniques for decades. What’s new is combining them with data-center-grade uptime requirements, where an unplanned pod retrieval to fix a failed component is vastly more disruptive than a routine maintenance visit to a land facility.

Sealing a data center away from humans cuts both ways. Project Natick’s finding — roughly one-eighth the failure rate of comparable land servers — is real and well-documented, and the mechanism is plausible: no dust, no accidental bumps, no humidity swings, no oxygen-driven corrosion. But that same sealing means a failed component cannot be swapped in place. If a power supply or network card fails inside a submerged pod, the entire unit typically has to be raised, which is expensive, weather-dependent, and can take a submerged facility offline for days rather than the minutes a land-based hot-swap would require. The practical answer engineers have converged on is redundancy over repairability: overprovision components, design for graceful degradation as individual servers fail, and plan retrieval around scheduled multi-year refresh cycles rather than reactive repairs — the same philosophy subsea telecom repeaters and offshore platforms have used for decades. This is a genuine design trade-off, not a free upgrade: a land data center can restore a failed node to full redundancy within minutes of a technician noticing an alert, while a submerged pod runs in a degraded state, on its remaining redundant capacity, until the next scheduled or emergency retrieval — which is why capacity planning for an underwater deployment has to assume a higher and longer-duration failure tolerance than a conventional facility would.

An underwater data-center pod draws on two supply chains that rarely overlap: conventional server manufacturing (the same racks, chips, and networking gear used in any land facility) and marine/offshore engineering (pressure hulls, corrosion-resistant coatings, subsea cable, and specialized deployment vessels). That second supply chain is why underwater projects can’t simply scale the way a land data center can by ordering more standard shipping-container modules. Pressure-hull fabrication requires shipyard-grade welding and testing capacity, which is a smaller and more specialized industry than general data-center construction; Highlander’s stated path to roughly 100 modules at Hainan depends as much on that manufacturing throughput as on demand. This is also where the modular, factory-built philosophy of the 2008-2010 containerized-data-center movement pays off directly: because the compute portion of a pod is built and tested exactly like a standard module, most of the schedule risk in a project like this sits in the marine-engineering half, not the IT half.

A CPU is built for flexibility — a handful of powerful cores that can run almost any instruction in sequence, ideal for an operating system juggling many different tasks. A GPU inverts that trade-off: thousands of simpler cores doing the same type of math simultaneously, which happens to be exactly what training a neural network requires, since most of that work reduces to enormous matrix multiplications that parallelize cleanly. That architectural difference is why the AI boom shifted data-center hardware purchasing toward GPUs and dedicated AI accelerators so quickly, and why power density per rack roughly quadrupled in the space of a few hardware generations. NVIDIA’s GB200 NVL72 — a single liquid-cooled rack housing 72 GPUs — draws on the order of 120-132 kW, with NVIDIA’s own specifications mandating direct-to-chip liquid cooling with inlet temperatures held between 20-25°C and flow rates around 80 liters per minute. Air cooling is not a viable option at that density; it isn’t a matter of preference.

Underwater and coastal data centers have a structural advantage most inland facilities lack: proximity to offshore wind, and in some designs, wave or tidal generation. Project Natick’s Orkney deployment deliberately drew on a local grid with a very high share of wind and tidal power, treating renewable integration as part of the experiment, not an afterthought. Reporting on China’s HiCloud facility near Shanghai describes a similar pairing with offshore wind. The engineering logic is straightforward: offshore wind farms already require subsea power cables and marine-grade infrastructure reaching the coast, so a nearby underwater data center can share that same cable corridor and substation connection rather than building redundant infrastructure. This doesn’t make underwater facilities inherently “green” — that depends entirely on the actual generation mix feeding the connection — but the physical adjacency is a genuine, engineering-level synergy that land-locked data centers can’t replicate as easily. Wave and tidal power integration, by contrast, remains mostly at the research and pilot stage industry-wide, and no major underwater data-center operator has published a production deployment drawing primary power from wave energy specifically as of 2026.

This is the section where hype most needs to be resisted in either direction. On one side, seawater cooling genuinely eliminates the chemical water consumption and evaporative losses that chiller-based land facilities incur — Highlander’s own figures claim roughly 105,000 tons of freshwater saved annually at full build-out, a real and meaningful category of impact. On the other side, peer-reviewed researchers have flagged open questions that remain open, not resolved. A study published in Scientific Reports (Nature) on heat-wave-resilient underwater computing architecture specifically models thermal stress scenarios for submerged systems, underscoring that ocean temperature isn’t perfectly stable everywhere or forever — marine heat waves are a documented, increasing phenomenon. Separately, marine biologists and environmental engineers have raised three specific concerns that deserve to be named precisely rather than waved away: localized thermal pollution from concentrated heat rejection in coastal waters, which could alter biodiversity for species that rely on stable temperatures for breeding and migration; underwater noise and vibration from mechanical components, which can affect marine animals sensitive to acoustic disturbance; and electromagnetic fields from subsea power cables, an effect already studied for offshore wind and telecom cables but not yet extensively for the specific configuration of a clustered data-center deployment.

There is no dedicated, underwater-data-center-specific regulatory regime anywhere in the world as of 2026. Instead, projects are permitted under whatever existing marine and offshore-infrastructure frameworks a given jurisdiction already applies to comparable seabed installations — the same categories of approval that cover subsea telecom cables and offshore wind foundations, typically involving environmental-impact assessment, maritime-safety sign-off, and coordination with fishing and shipping authorities over seabed rights and exclusion zones. China’s approval of Highlander’s Hainan site, and the earlier UK process that allowed Project Natick’s Orkney deployment, both moved through those general-purpose channels rather than a purpose-built underwater-data-center law. That absence of dedicated regulation cuts two ways: it has made early deployment faster, since operators aren’t waiting on new legislation, but it also means the environmental open questions in the section above are being resolved case by case, project by project, rather than against a settled, industry-wide standard. Environmental researchers studying this space have generally called for continued monitoring and study rather than either blanket approval or a moratorium — a position consistent with the broader evidence-still-developing theme running through this article.

The economic pitch for underwater deployment rests on three claims, each with different strength of evidence. First, speed: Microsoft cited roughly 90 days from factory to operational seabed deployment for a Natick-style pod, versus one to two years to permit, build, and commission a land facility — a genuine, well-documented advantage for anyone needing capacity fast. Second, reduced cooling capex and opex: eliminating chillers and CRAC units removes both their purchase cost and their ongoing electricity draw, which Highlander estimates at roughly 90% of cooling-specific energy for its site — plausible given the underlying physics, though not yet independently audited across a full year of operation. Third, land and water savings, which matter disproportionately in dense coastal markets like Hainan, Singapore, or parts of Northern Europe where buildable land or fresh water for cooling is itself scarce and expensive. Working against these advantages: retrieval-based maintenance is expensive and weather-dependent when it’s needed, marine engineering and pressure-hull manufacturing add real upfront cost per unit of compute versus a simple warehouse building, and the entire commercial track record — one operational site with a multi-year history — is too short to model long-term total cost of ownership with confidence. Every serious economic case for this technology today should be read as directionally promising, not proven at scale.

An underwater data center is far more likely to function as an edge computing node than as a hyperscale AI training campus, and the distinction matters for understanding where this technology realistically fits. A hyperscale facility centralizes enormous compute capacity in one location to maximize efficiency of scale; an edge node sacrifices some of that efficiency to sit physically closer to users, cutting the network latency for time-sensitive applications like gaming, video conferencing, or financial trading. A single Highlander-style pod, with capacity comparable to tens of thousands of PCs, is a rounding error next to a hyperscale AI campus that can span millions of square feet and consume hundreds of megawatts — but it doesn’t need to compete on that axis. Its value proposition is proximity to a coastal population center with sub-10-millisecond latency, not raw scale.

The figures scattered through this article — a depth here, a PUE range there, a company name attached to a claim two sections back — are consolidated below into three reference tables: the compressed version of this timeline, the companies and organizations actually involved, and the hardware components an AI data center’s cooling system has to serve. Each is meant to be skimmable on its own, without re-reading the surrounding prose.

• Project Natick’s Phase 1 test vessel was named “Leona Philpot,” after a character from Microsoft’s own Halo video game franchise.
• A single NVIDIA GB200 NVL72 rack draws roughly the same power as 100 typical American homes.
• Nearly half the world’s population lives within about 200 kilometers of a coastline — the original geographic logic behind underwater data-center placement.
• TeleGeography’s 2026 map counts roughly 694 submarine cable systems carrying the vast majority of intercontinental internet traffic — the same class of infrastructure that connects coastal underwater data centers to the internet.
• Highlander’s Hainan modules are designed to be retrieved, refitted, and redeployed rather than scrapped at end of life — closer to a ship’s maintenance cycle than a building’s.
• Data-center cooling can account for roughly 40% of total facility electricity use before removing chillers — the single biggest efficiency lever the industry has left to pull.

Strip the coverage down to ownership and the picture is narrower than the headlines suggest. Exactly one company, Highlander, operates a confirmed commercial-scale underwater data center today. Exactly one company, Microsoft, ran the research program that produced the field’s core reliability data, and it is no longer actively deploying pods. Everyone else covered in this article — Subsea Cloud, Nautilus Data Technologies, HiCloud — sits somewhere between an early pilot and a separate-but-adjacent floating-water-cooled model. That’s a small, concentrated field for a topic that generates outsized media attention, and it’s worth keeping in mind when a headline implies this is a broad, multi-vendor industry rather than a handful of specific, individually-traceable projects.

Geographically, the picture is similarly concentrated rather than global. China currently has the only confirmed commercial-scale deployment (Highlander) plus at least one additional reported project (HiCloud), giving it a clear lead on actual operational sites, generally attributed in coverage to strong state-industrial coordination on infrastructure projects and specific coastal siting incentives. The United Kingdom hosted the foundational research (Project Natick’s Orkney deployment) but that program has since ended. The United States has pilot and adjacent activity — Subsea Cloud’s trials and Nautilus’s floating Stockton facility — without a fully submerged commercial site of its own. Norway has hosted Subsea Cloud pilot deployments. No major, publicly confirmed underwater data-center project has been documented in India, continental Europe outside Norway and the UK, or Africa as of 2026; those regions’ data-center growth has focused on conventional and land-based liquid-cooled facilities instead.

Three developments are grounded enough in current, published plans to discuss with confidence, without drifting into speculative timelines. First, Highlander’s own stated roadmap targets scaling the Hainan site toward roughly 100 modules; whether that timeline holds depends on manufacturing throughput and demonstrated multi-year reliability data the company has not yet published. Second, OCP’s cooling-standards work is on a public track toward broader interoperability across cold-plate, immersion, and heat-reuse specifications, which lowers the engineering barrier for new entrants regardless of whether they build on land or underwater. Third, renewable-paired coastal siting — locating data centers, underwater or not, alongside offshore wind infrastructure to share subsea cable corridors — is an active design pattern in current offshore-wind and data-center planning discussions, though production examples remain limited to the handful covered in this article.

What the roadmap does not currently support: claims that underwater deployment will become the default for AI training clusters, that it will replace land-based hyperscale campuses at scale, or that environmental questions are already settled favorably. Every credible primary source referenced in this article — Microsoft, the IEA, the Uptime Institute, OCP, and the peer-reviewed research on marine thermal effects — treats this as one promising tool among several, still accumulating evidence, not a foregone conclusion.

On the funding side, the honest picture is that this remains a small category relative to overall data-center capital spending, which continues to flow overwhelmingly toward conventional and liquid-cooled land facilities to meet AI demand. Highlander’s build-out is backed by Chinese state and industrial partners rather than the kind of venture-capital rounds typical of a Silicon Valley infrastructure startup; Subsea Cloud has disclosed customer and partner relationships (Chevron, Laborde Marine, Oracle, per company statements) more than specific funding figures. That funding pattern is itself informative: this looks less like a venture-backed technology bet racing to prove product-market fit, and more like targeted, patient infrastructure investment in a niche with a plausible but not yet fully validated payback case — a distinction worth keeping in mind before extrapolating rapid, VC-style scaling onto a field that has grown mostly through direct industrial and government backing so far.

What would actually change this outlook, in either direction? On the upside: a second and third independently-verified commercial deployment matching or beating Highlander’s efficiency claims would turn a single data point into a trend; published multi-year reliability data from Hainan, once the site has been running long enough to generate it, would do for the commercial era what Natick’s two-year study did for the research era. On the downside: a serious hardware failure requiring an expensive unplanned retrieval, or a peer-reviewed study documenting measurable ecological harm at an operating site, would meaningfully weaken the case for expansion. As of this writing, neither has happened — the evidence base is still thin enough that either outcome remains plausible, which is precisely why this article treats the technology as promising and unproven rather than settled in either direction.

Underwater data centers submerge sealed, pressure-rated server pods in the ocean to use stable seawater temperatures for passive cooling instead of mechanical air conditioning. Microsoft’s Project Natick (2014-2024) proved the engineering concept, running 855 servers off Scotland for over two years with roughly one-eighth the failure rate of comparable land servers, before the program was formally retired. In 2025, China’s Highlander company moved the concept from research to commerce, launching a 1,300-tonne commercial underwater data center off Hainan that claims roughly 90% lower cooling energy use, with plans to scale toward 100 modules at that site. The driving force behind renewed interest is AI: global data-center electricity demand grew 17% in 2025 alone, with AI-specific facilities surging roughly 50%, per the International Energy Agency, as GPU racks drawing 100-plus kilowatts outpace what air cooling can handle. Underwater deployment builds directly on immersion and liquid-cooling research standardized by the Open Compute Project. It is best understood as a promising complement for coastal, edge, and water-or-land-constrained deployments — not a replacement for conventional hyperscale data centers, and not yet proven at the scale or duration needed to settle its long-term economics or environmental impact.

The questions below are grouped in the same order this article covers the topic: basics and terminology first, then Project Natick specifically, Highlander and China’s commercial deployments, the underlying cooling engineering, AI and hardware, environmental impact, economics, reliability and maintenance, adjacent technologies, and finally companies, geography, and sourcing. Each answer is written to stand alone — useful if you arrived here from a search result rather than reading top to bottom — and is hedged to match the actual evidence level behind it, whether that’s published research, an operator’s own claim, or an open question nobody has settled yet.

Go back to that hot, roaring data hall from the opening of this article. The engineer standing in it is not choosing between “normal data centers” and “underwater data centers” — that framing was never accurate. They’re choosing among a widening menu of cooling strategies, of which submerging the whole facility in the ocean is one of the more radical, and so far one of the most narrowly proven. Project Natick’s decade of research established, with real data, that the core physics works and that isolation from air and humans can meaningfully improve hardware reliability. Highlander’s Hainan facility established, in 2025, that the concept can run as an actual commercial service, not just a research platform. Neither of those facts means every future data center will be underwater, and no engineer, company, or credible research body covered in this article claims that.

What seems genuinely likely, based only on what’s published today: advances in pressure-hull manufacturing, subsea power and networking, and the same liquid- and immersion-cooling standards work happening industry-wide will keep making underwater deployment cheaper and more repeatable for the specific situations where it makes sense — dense coastal cities, water-stressed regions, edge-latency-sensitive applications, and sites where speed of deployment matters more than raw scale. Land-based facilities will keep handling the bulk of hyperscale AI training and general cloud computing, because that’s where the economics and the engineering track record are strongest today. The most honest conclusion the current evidence supports is also the least dramatic one: the ocean has become a legitimate, if still small, part of how the internet’s physical infrastructure is built — not because it’s the future replacing the past, but because it’s one more tool, validated by a decade of research and now a year of commercial operation, added to an industry that badly needed more ways to solve a heat problem AI made urgent.