DataCentersX > Stack > Cooling and Thermal Management > Immersion Cooling

DC Immersion Cooling

Immersion cooling submerges server hardware in a dielectric fluid that both electrically insulates the electronics and carries away heat. It is the only mainstream modality in which there is no intermediate transport fluid between the heat-dissipating components and the coolant: every chip, memory module, VRM, and NIC is in direct contact with the working fluid. That architectural simplicity eliminates the coverage gap that constrains direct-to-chip and pushes the density envelope to levels where rack and even row become the wrong organizing units.

Immersion is mature in cryptocurrency mining, where the economics justify the capital expense and the hardware fleet is homogeneous, purpose-built, and warranty-compatible with fluid exposure. In AI training and HPC, immersion remains at pilot and select deployment scale. The open engineering question is whether immersion eventually displaces direct-to-chip at some density threshold, or whether progressively more aggressive DTC designs (cold plates on every significant dissipator) continue to absorb the envelope first.

Single-phase versus two-phase

Immersion cooling splits into two architecturally distinct approaches based on whether the working fluid changes state during operation. The distinction drives fluid selection, tank design, pumping requirements, and regulatory exposure.

Property	Single-Phase Immersion	Two-Phase Immersion
Working fluid state	Liquid throughout; no boiling at chip surface	Boils at die surface; vapor rises and condenses at cold coil
Typical fluids	Synthetic hydrocarbons, mineral oils, silicone-based fluids	Engineered fluorocarbons with boiling points 50 to 60 degrees C
Heat transport mechanism	Pump-driven forced convection	Natural circulation driven by vapor rise and condensate return
Tank pressure	Atmospheric, open or loosely sealed	Near-atmospheric but requires vapor containment
Thermal capacity	High; well-suited to 50 to 150 kW per tank	Very high; phase-change absorbs large heat loads at constant temperature
Regulatory status	Conventional industrial fluid handling	Fluorinated fluids under active PFAS regulation

Single-phase immersion

A single-phase system consists of a sealed or open tank holding servers vertically in a bath of dielectric fluid. Fluid circulates through the tank, across the hardware, out to an external heat exchanger (where it transfers heat to a facility water loop), and back into the tank. Pumps provide the motive force; the fluid never boils and never changes state.

The dominant fluid chemistries are synthetic hydrocarbons (engineered for low viscosity, high flash point, and material compatibility), mineral oils (lower cost, higher viscosity, adequate for many workloads), and silicone-based fluids (excellent thermal stability, higher cost). Fluid selection trades off thermal performance, material compatibility with board components, environmental and handling properties, and cost per liter at the scale of a multi-megawatt deployment.

Single-phase dominates current AI and HPC immersion pilots because the fluid chemistry is mature, the regulatory status is clear, and the engineering is closer to conventional heat-exchanger design than two-phase. Submer, LiquidStack (single-phase product line), and GRC are the primary vendors building to this architecture.

Two-phase immersion

A two-phase system exploits the latent heat of vaporization. The working fluid boils at the chip surface (typically in the 50 to 60 degrees C range at atmospheric pressure), vapor rises through the tank, condenses on a water-cooled coil at the top of the enclosure, and falls back as liquid. No pumps are required in the immersion bath, and the system self-regulates: hotter chips boil more fluid, more vapor rises, more heat crosses to the condenser.

The thermal performance advantage is substantial. Phase change absorbs heat at nearly constant temperature, which holds silicon junction temperatures in a narrow band regardless of load variation. Two-phase systems can handle the highest power densities of any mainstream cooling approach and scale to rack-equivalent loads of 250 kilowatts and beyond without density-limiting.

The regulatory constraint is real and active. The engineered fluids with boiling points in the useful 50 to 60 degree C range have historically been perfluorinated compounds, a class under increasing restriction in the EU (REACH), the US (EPA), and elsewhere as PFAS regulations tighten. 3M announced in 2022 that it would exit the fluorochemical manufacturing business, which included the Novec fluids widely used in two-phase immersion. Replacement chemistries are under active development but the supply landscape remains less settled than single-phase.

LiquidStack is the best-known two-phase vendor. Several research and HPC deployments run two-phase systems, but the regulatory uncertainty has slowed enterprise and hyperscale adoption relative to what the thermal performance would otherwise justify.

Material compatibility

The engineering discipline that most distinguishes immersion from other modalities is material compatibility. Every component on the board is in continuous contact with the working fluid, which means every material in every component has to tolerate long-term immersion without degradation, dissolution, or chemical interaction.

The materials under scrutiny include printed circuit board laminates and solder masks, component encapsulants on capacitors and ICs, labels and adhesives on components and cables, rubber gaskets and cable jackets, and thermal interface materials on any heat sinks that remain in the system. Some of these (PCB laminate, solder) are generally compatible with dielectric fluids. Others (adhesive labels, certain elastomers, some encapsulants) can swell, soften, or leach into the fluid over time, degrading both the component and the fluid chemistry.

Hardware vendor warranties are the practical expression of this constraint. Standard servers are not warranted for immersion operation; deployment in dielectric fluid voids manufacturer warranty in most cases. A small and growing segment of OEM product lines ships immersion-certified, with verified material compatibility and contractual warranty coverage. This segment is the structural gate on broader enterprise adoption: until mainstream AI reference designs ship with immersion warranty coverage, immersion remains a pilot-scale modality.

Tank and facility architecture

An immersion deployment replaces the rack with a tank. Tanks are typically open-top (for single-phase) or sealed (for two-phase), holding servers oriented vertically with backplanes and connectors at the top for serviceability. Tank capacity in current deployments spans roughly 50 to 250 kilowatts per tank, with larger purpose-built systems in the megawatt range for the highest-density AI pilots.

Facility integration requires a secondary water loop for heat rejection, analogous to the facility water loop on a direct-to-chip deployment. The heat exchanger between the dielectric fluid and the facility water is the analog of a CDU; its sizing and redundancy dominate tank-level design. Leak management is different from DTC: dielectric fluid is typically non-conductive and not immediately hazardous to electronics, but spills are expensive (fluid cost) and disruptive (floor cleanup, tank refill).

Serviceability requires new operational procedures. A server pulled from a tank drips fluid, which has to be captured and returned. Components are wet during handling. Cable routing, board-level diagnostics, and component replacement all happen with fluid-coated hardware. Operations staff need immersion-specific training and the hall layout has to accommodate fluid handling equipment and wet-service areas.

Where immersion fits

Immersion cooling sits at the high-density end of the modality spectrum, with effectively unbounded per-tank density limited more by power delivery and networking than by thermal capacity. It is the correct engineering choice today for specific applications: cryptocurrency mining at scale, select HPC deployments where fluid-compatible hardware is custom-specified, research and pilot programs exploring post-DTC density, and a small number of AI deployments where the operator has the organizational capacity to manage the operational model.

Immersion is not yet the correct choice for mainstream AI training, where direct-to-chip reference designs continue to absorb the density envelope through extended cold-plate coverage. Whether that changes depends on three factors: how far DTC can be pushed before material and serviceability costs exceed immersion's, how quickly replacement two-phase chemistries reach commercial maturity, and how AI accelerator TDPs evolve over the next two generations. The competitive boundary between DTC and immersion at 500 kilowatts per rack and beyond is one of the open engineering questions in data center thermal management.

Related coverage