Immersion cooling in a data center means submerging servers directly into a thermally conductive, electrically non-conductive dielectric fluid instead of relying on air or water to carry heat away. The fluid contacts the hardware directly, absorbs heat from processors and other components at the point of generation, and is then cooled and recirculated. The result is a fundamentally more efficient heat transfer path than any air-based system can provide.
How Immersion Cooling Works and Why It Is Different from Air Cooling
Conventional air cooling methods in data centers rely on fans to move cooled air across hot components, then exhaust the heated air out of the rack or room. The thermal conductivity of air is low, roughly 0.026 W/m·K, which forces data center operators to maintain large temperature differentials, use substantial fan power, and keep server density within limits that air can manage. At 30 kW per rack and above, those limits become genuinely binding.
Dielectric fluids have thermal conductivity values orders of magnitude higher than air, and heat transfer scales with fluid mass flow rate rather than gas velocity. Every component in a submerged server transfers heat simultaneously, including memory, voltage regulators, and storage. Fans are removed entirely, eliminating a failure point and recovering their power for compute.
The heated fluid exits the tank through a heat exchanger connected to a facility water loop or dry cooler. Because tank temperatures can run as high as 45-50°C without harming hardware, the facility side can often use adiabatic or dry cooling instead of mechanical chillers, reducing both capital cost and energy draw.
Single-Phase vs Two-Phase Immersion Cooling
Single-phase immersion cooling keeps the dielectric fluid in liquid form throughout the entire cycle. The fluid enters the tank as a liquid, absorbs heat from the submerged hardware, exits still as a liquid (just warmer), and passes through a heat exchanger before returning to the tank. The process is continuous and mechanically simple. Common single-phase fluids include mineral oils, synthetic hydrocarbon-based fluids, and engineered fluids such as those sold commercially by companies including Submer and GRC (Green Revolution Cooling).
Two-phase immersion cooling operates differently. The fluid is chosen specifically because it boils at a low temperature (typically 49-60°C) when in contact with hot components. As it boils, it absorbs latent heat of vaporization from the hardware surface, which is a very efficient heat transfer mechanism. The vapor rises to the top of the tank, condenses on a coil or chilled lid, and falls back into the liquid bath as a recovered fluid. Two-phase systems can handle extremely high heat flux densities because boiling is an inherently efficient way to move large amounts of heat rapidly.
The fluorinated fluids historically used in two-phase systems, including the 3M Novec family, face serious regulatory pressure. Their classification as PFAS (per- and polyfluoroalkyl substances) has triggered restriction proposals under the EU’s REACH framework, pushing operators toward lower-GWP hydrofluoroether alternatives and non-fluorinated fluid chemistries. The regulatory picture is still developing as of mid-2026. Anyone planning a two-phase deployment should treat fluid selection as a long-term strategic decision, not purely a thermal performance one.
| Factor | Single-Phase Immersion | Two-Phase Immersion |
|---|---|---|
| Heat transfer mechanism | Sensible heat (no phase change) | Latent heat of vaporization |
| Fluid type | Mineral oil, synthetic hydrocarbon, engineered fluid | Low-boiling-point fluorocarbon or hydrofluoroether |
| Thermal efficiency | High; supports 100+ kW/rack | Very high; handles extreme heat flux (200+ kW/rack in some deployments) |
| System complexity | Lower; simpler pump and heat exchanger circuit | Higher; vapor management, condenser coils, sealed tank |
| Fluid cost and availability | Generally lower; more sourcing options | Higher; specialized fluids, regulatory risk on fluorinated variants |
| Typical use case | AI training clusters, HPC, dense compute at scale | Highest-density workloads, specialized compute, legacy HPC installations |
Immersion Cooling vs Direct-to-Chip Liquid Cooling
Direct-to-chip liquid cooling (cold plate cooling) routes liquid through a metal cold plate on the processor die and carries heat to a facility water loop. Memory, storage, and power supplies still cool via air. Most deployments today are hybrid: cold plates on CPUs and GPUs, residual air for everything else.
The case for direct-to-chip is compatibility. Servers designed for direct-to-chip cooling are standard form-factor machines, available from major OEMs, with existing warranty and support structures intact. You do not need to modify hardware to fit a tank, you do not need to source and manage large volumes of dielectric fluid, and your maintenance procedures remain close to what your operations team already knows. NVIDIA’s current generation of H100 and H200 SXM modules ships with liquid cooling support, and major server vendors including Dell, HPE, and Supermicro offer cold-plate-cooled AI server configurations.
Immersion wins on whole-system thermal coverage. Because every component is submerged, there are no thermally uncooled parts of the server. In practice, this means immersion systems can push rack densities and sustained utilization levels that direct-to-chip systems cannot reach because the residual air path becomes a bottleneck at high loads. If you are running AI training clusters at very high utilization, 24 hours per day with GPU thermal design powers exceeding 700W per card, immersion provides a more complete thermal solution. The tradeoff is the upfront investment in tank infrastructure, hardware compatibility verification, and a maintenance workflow that is simply different from working with standard rack servers.
Efficiency: PUE, Heat Reuse, and the Fan Elimination Effect
A data center’s Power Usage Effectiveness (PUE) is the ratio of total facility power to IT power. A PUE of 1.0 would mean every watt entering the building powers compute; nothing is lost to cooling overhead. Air-cooled facilities typically run PUEs between 1.3 and 1.6 at scale. Well-optimized direct-to-chip liquid-cooled facilities can achieve PUEs in the 1.1-1.2 range. Immersion-cooled facilities, particularly those with warm-fluid operation and no mechanical chilling requirement, can operate at PUEs below 1.1, with some published figures approaching 1.03-1.05 under favorable ambient conditions.
Two further efficiency gains compound the PUE story. Fan elimination: a standard 1U server’s fans draw 50-150W under load; immersion servers run fanless and that power goes to compute. Heat reuse: systems engineered for warm-fluid operation produce 40-70°C output usable for district heating, greenhouse heating, or process heat. Several European facilities have made heat reuse the central economic justification for their immersion build-out, not PUE alone.
Where Immersion Cooling Is Being Deployed
AI training infrastructure is the most active segment. As operators push into NVIDIA H100 and H200 GPU clusters and look toward Blackwell-architecture systems where per-card TDP keeps climbing, the thermal case for immersion has sharpened. Broader hyperscaler AI infrastructure investment signals increasing willingness to commit to non-air thermal architectures as rack densities exceed 100 kW.
Cryptocurrency mining was an early commercial adopter. High utilization, commodity hardware, and dense deployment made the economics compelling before most enterprise data centers were willing to move. That adoption helped mature the vendor ecosystem and brought tank and fluid costs down. HPC in national laboratory settings has also adopted immersion, particularly for physics simulations and molecular dynamics workloads running at sustained high loads.
Edge deployments are a smaller but growing niche. Sealed immersion tanks need minimal external airflow and generate far less acoustic noise, which makes them viable in non-traditional environments: manufacturing floors, substations, and remote sites where a conventional rack requiring precision air conditioning would not be practical.
The Challenges You Need to Plan For
Serviceability is the first friction point most operators encounter. Pulling a server from an immersion tank means handling a fluid-coated machine, which requires protective equipment, a drainage step, and cleaning before standard bench work can begin. Single-phase mineral-oil systems are viscous at room temperature, making the workflow noticeably more involved than pulling a standard rack server. Two-phase fluorocarbon fluids are less viscous but bring their own handling and containment requirements.
Hardware warranty coverage is inconsistent. Most server vendors have not issued blanket approvals for immersion deployment; some have certified specific platforms with specific fluids, others simply void coverage if hardware is submerged. The Open Compute Project has published immersion cooling specifications that give vendors a standard to align to, but the position varies enough that you should verify the warranty stance for every platform before committing at scale.
The capital expenditure profile is front-loaded. Tanks, fluid volumes, CDUs (coolant distribution units), and facility-side heat exchanger loops all carry costs that do not apply to air-cooled rack installations. The total cost of ownership argument for immersion often becomes favorable over a 5-7 year horizon, but the initial build-out cost is higher, and that matters when capital budgets run on 12-18 month cycles.
The vendor ecosystem is also still maturing. The number of qualified suppliers for tanks, fluids, and immersion-compatible server configurations has grown over the past four years, but it remains less commoditized than the air-cooling supply chain. Lead times and the depth of operational knowledge among data center staff are real planning factors, not theoretical ones.
Frequently Asked Questions
What is immersion cooling in a data center?
Immersion cooling means submerging servers in a bath of dielectric fluid, a thermally conductive, electrically non-conductive liquid that directly absorbs heat from processors, memory, and other components. The heated fluid is pumped through a heat exchanger, cooled, and returned to the tank. No air cooling fans are required, and the thermal transfer efficiency is dramatically higher than air.
What is the difference between single-phase and two-phase immersion cooling?
In single-phase systems, the fluid stays liquid throughout, absorbing heat through temperature rise only. In two-phase systems, the fluid boils at a low temperature when it contacts hot components, absorbing heat through the phase change itself, which is a more efficient process. Two-phase handles higher heat densities but uses more expensive, specialized fluids that face increasing regulatory scrutiny over PFAS content.
Is immersion cooling better than direct-to-chip liquid cooling?
Neither is universally better; the choice depends on your density requirements and risk tolerance. Direct-to-chip cooling is easier to integrate with standard server hardware and vendor warranties. Immersion covers every component thermally and handles higher rack densities, but requires purpose-built tanks, fluid management, and modified hardware. For most first deployments, direct-to-chip is the lower-friction starting point.
What fluid is used in immersion cooling?
Single-phase systems typically use mineral oil, synthetic hydrocarbon blends, or engineered dielectric fluids from vendors such as Submer or GRC. Two-phase systems have historically used fluorinated fluids in the 3M Novec family, though regulatory pressure on PFAS compounds is driving adoption of alternative low-boiling-point fluids with better environmental profiles. Fluid selection is now as much a regulatory and supply chain question as a thermal performance one.
What are the downsides of immersion cooling?
The main challenges are serviceability complexity (hardware must be drained and cleaned before hands-on work), inconsistent hardware warranty coverage from OEMs, high upfront capital cost for tanks and fluid volume, and an immature supply chain compared to air cooling. Fluid cost and regulatory uncertainty around fluorinated two-phase fluids add further planning complexity. Total cost of ownership typically favors immersion over a 5-7 year horizon, but the break-even requires high utilization.
