What Immersion Cooling Means for the Next Generation of Data Centers

Every few years, the data center industry hits a thermal wall.

The last one arrived quietly around 2015, when rack densities crept past 10 kilowatts and operators started rethinking airflow management.

The current wall is not quiet at all. A single NVIDIA H100 GPU draws around 700 watts. The newer B200 pushes past 1,000 watts.

Pack eight of those into a server chassis, stack three or four chassis in a rack, and you are looking at 40 to 56 kilowatts of heat pouring out of a single cabinet — roughly the thermal output of twenty kitchen ovens running simultaneously.

Air cooling was never designed for this.

Traditional raised-floor data centers, with their perforated tiles and hot-aisle containment, were engineered for an era when 5 to 8 kilowatts per rack was considered dense.

Even with refinements like in-row cooling units and rear-door heat exchangers, air-based systems struggle to economically manage racks above 25 to 30 kilowatts.

The physics are unforgiving: air is a poor thermal conductor, and moving enough of it to absorb heat at GPU-scale densities requires enormous fan energy, oversized ductwork, and constant mechanical attention.

The result is bloated power usage effectiveness (PUE) numbers — typically between 1.4 and 1.7 for conventional air-cooled facilities — meaning that for every watt delivered to a server, nearly half a watt is spent just keeping it cool.

Meanwhile, the electricity appetite of data centers globally has become difficult to ignore.

The International Energy Agency estimates that data centers consumed roughly 415 terawatt-hours in 2024, about 1.5 percent of global electricity demand, growing at 12 percent annually.

U.S. data centers alone now account for over 4 percent of national electricity consumption, and that figure could triple by the end of the decade.

AI workloads are the primary accelerant.

Electricity consumption from AI-optimized servers is projected to climb from 93 terawatt-hours in 2025 to 432 terawatt-hours by 2030, a nearly fivefold increase.

Against this backdrop, any technology that meaningfully reduces the energy overhead of cooling is not an incremental improvement — it is an operational necessity.

Submerging the Problem

Immersion cooling takes a conceptually simple approach to a complex thermodynamic challenge.

Instead of blowing air across hot components, you submerge them directly in a thermally conductive, electrically non-conductive fluid — a dielectric liquid that absorbs heat far more efficiently than air ever could.

The servers sit in open tanks or sealed enclosures, bathed in fluid that draws heat away from processors, memory, and power delivery systems uniformly and without the hot spots that plague air-cooled racks.

There are two primary variants.

Single-phase immersion cooling uses a fluid — often a synthetic hydrocarbon or engineered dielectric oil — that remains liquid throughout the process. Heat transfers from components into the fluid, which is then circulated to an external heat exchanger where it cools before returning to the tank.

The fluid never boils, never changes state. This simplicity is its chief advantage: tank designs are straightforward, fluids are relatively affordable at around $20 to $50 per gallon, and existing server hardware often requires only minor modifications.

Single-phase systems routinely handle rack densities up to 200 kilowatts while achieving PUE values between 1.02 and 1.03.

Two-phase immersion cooling is more thermodynamically aggressive. It uses a low-boiling-point fluorinated fluid that vaporizes on contact with hot components.

The phase change itself — liquid to gas — absorbs enormous amounts of energy, making it exceptionally efficient at removing heat from dense configurations. The vapor rises to a condenser at the top of the sealed enclosure, returns to liquid, and drips back down.

Two-phase systems can support rack densities exceeding 250 kilowatts with PUE values as low as 1.01.

The tradeoff is cost and complexity: fluorinated coolants run $200 to $500 per gallon, the enclosures must be sealed to prevent fluid loss, and maintenance requires careful handling to minimize coolant evaporation.

Both approaches share a critical advantage over air: they eliminate fans.

No fans means no fan failures, no vibration-induced wear on solder joints, and no acoustic load. It also means no need for raised floors, hot-aisle containment panels, or the elaborate ductwork that consumes valuable floor space in traditional facilities.

What the Numbers Actually Show

The efficiency gains are not theoretical.

When a facility migrates from standard air cooling with a PUE of 2.0 to immersion cooling at 1.02, the overall computing efficiency increases by over 96 percent.

Even against best-practice air cooling — facilities already optimized to a PUE of 1.12 — immersion cooling delivers a roughly 10 percent efficiency gain.

In energy terms, running identical workloads on identical hardware costs 37 percent less in electricity with immersion cooling compared to air.

Immersion cooling cuts energy use by more than 40 percent and extends hardware life by approximately 30 percent.

Those two figures interact in ways that compound over time. Lower operating temperatures reduce electromigration in chip interconnects, slow capacitor degradation, and prevent the thermal cycling — the constant expansion and contraction from temperature swings — that is one of the leading causes of solder joint failure in electronics.

Data from production deployments between 2023 and 2025 shows quarterly server failure rates of just 0.6 to 1.4 percent in immersion-cooled environments.

Some submerged facilities report failure rates one-eighth those of comparable air-cooled operations.

Hardware that lasts longer does not just save on replacement costs. It delays the capital expenditure cycle for refresh, reduces electronic waste, and keeps validated, known-good configurations in production longer.

That is a meaningful consideration for enterprises running inference workloads where consistency matters more than chasing the newest silicon.

The total cost of ownership picture is equally compelling.

An analysis of a 10-megawatt AI data center found that immersion cooling reduced TCO by approximately 39 percent over a decade — roughly $110 million in savings compared to air cooling.

Those savings come from multiple vectors: lower electricity bills, reduced mechanical infrastructure, smaller physical footprints (immersion-cooled facilities can be up to 85 percent more compact), and fewer maintenance interventions.

Upfront capital costs for immersion are comparable to air cooling at equivalent densities, and as rack density increases, immersion pulls ahead decisively.

At 40 kilowatts per rack, capital costs drop to around $6 per watt, well below the $7 per watt typical of air-cooled configurations.

The Water Question

Energy consumption gets most of the attention, but water may be the more consequential resource issue for the industry.

A large data center can consume 5 million gallons of water per day for cooling — drawn from municipal supplies, surface water, or underground aquifers.

U.S. data centers collectively used 66 billion liters of water in 2023, more than triple the amount from a decade earlier.

More than 160 new AI data centers have been built across the U.S. in the past three years, many in regions already facing water stress.

Immersion cooling largely sidesteps this problem. Closed-loop dielectric fluid systems do not consume water in the cooling process.

Some implementations use water in secondary heat rejection — transferring heat from the dielectric fluid to a building-level cooling loop — but the volumes are a fraction of what evaporative or chilled-water systems require.

Studies have found that immersion cooling can reduce water consumption by up to 91 percent compared to conventional air cooling.

For operators building in water-constrained regions like the American Southwest, or for municipalities pushing back against the water demands of new facilities, this is a material advantage in both permitting and public perception.

The sustainability case extends beyond resource consumption.

Denser facilities mean fewer facilities overall to deliver the same compute capacity, which translates to less land use, less construction material, and shorter supply chains for physical infrastructure.

GPU-dense racks cooled by immersion can deliver in a single building what might otherwise require multiple air-cooled halls — a meaningful consideration as available power and real estate near population centers become scarce commodities.

Where the Industry Stands Today

Immersion cooling is no longer a research curiosity.

Companies like GRC, Submer, and Asperitas have deployed production-scale systems across enterprise, colocation, and hyperscale environments.

Microsoft demonstrated two-phase immersion as early as 2021 and reported cooling energy reductions of up to 95 percent.

Alibaba, Google, and Meta have all tested immersion cooling for AI and machine learning clusters.

In 2025, GRC launched an automated coolant lifecycle management platform with real-time fluid purity monitoring and predictive maintenance, a sign that the technology has matured past the proof-of-concept stage and into the operational tooling phase.

The market trajectory reflects this maturation. Single-phase direct-to-chip cooling is becoming standard for AI workloads now, while two-phase immersion is expected to reach mainstream adoption by 2027.

The convergence of several pressures — GPU power densities that air cannot handle, electricity costs that operators cannot absorb, water constraints that regulators will not ignore, and sustainability commitments that enterprises must honor — is making immersion cooling less of a technology choice and more of an infrastructure inevitability.

That said, adoption is not without friction.

Maintenance workflows change: technicians accustomed to sliding a server from a rack and swapping a drive must now work with fluid-submerged hardware. Staff training is required.

Some OEM warranties may not cover immersion-modified equipment, though this is changing as more vendors certify their hardware for submersion.

And the dielectric fluids themselves, particularly the fluorinated compounds used in two-phase systems, face increasing regulatory scrutiny in Europe and parts of the U.S. due to concerns about per- and polyfluoroalkyl substances.

The industry is responding — mineral oil and synthetic hydrocarbon alternatives for single-phase systems avoid these concerns entirely — but operators evaluating two-phase deployments need to factor in evolving chemical regulations.

Building for What Comes Next

The data center of 2030 will look fundamentally different from the one built in 2020.

Rack densities will routinely exceed 50 kilowatts; some AI training clusters are already pushing past 120 kilowatts per rack.

At those levels, air cooling is not just inefficient — it is physically inadequate.

The question is not whether liquid cooling will become the norm, but which form it will take and how quickly operators will make the transition.

Immersion cooling represents the most complete answer to the thermal challenge because it addresses the entire system simultaneously — processors, memory, power delivery, and storage all benefit from uniform fluid contact, rather than requiring separate liquid loops for each component.

It also decouples facility design from climate: an immersion-cooled data center in Dallas performs identically to one in Minnesota, because outdoor air temperature is irrelevant to a sealed fluid loop.

That geographic flexibility matters for operators choosing sites based on power availability and network connectivity rather than ambient temperature.

For mid-market companies building private AI infrastructure — organizations that need GPU-dense compute without hyperscaler budgets — the economics of immersion cooling are particularly attractive.

A 2.5-megawatt facility using immersion cooling can deliver the compute density of a much larger air-cooled operation, with lower ongoing energy costs and a smaller physical footprint.

Advanced immersion cooling lowers power use and operating costs enough to enable competitive pricing and better unit economics.

That matters enormously when the goal is making private AI infrastructure accessible to organizations that have traditionally been priced out of the conversation.

The thermal wall is real, and it is getting taller with every new generation of accelerator hardware.

Immersion cooling does not just solve the immediate heat problem. It reshapes the cost structure, the environmental footprint, and the design constraints of the facilities that will power the next decade of AI.

The operators who recognize this early will build infrastructure that remains viable long after the current generation of GPUs has been replaced by whatever comes next.

The ones who do not will find themselves retrofitting — expensively, disruptively, and on someone else’s timeline.