TL;DR
- Immersion cooling submerges entire servers in a non-conductive dielectric fluid that absorbs heat directly from every component, including DIMMs, NICs, and VRMs.
- Two flavours: single-phase (fluid stays liquid, pumped through a heat exchanger) and two-phase (fluid boils at chip temperature, vapour condenses on a coil and rains back).
- Achievable rack densities are 100-250+ kW with very low PUE (often 1.03-1.10) because there are no server fans and no CRAH units in the cooled space.
- Trade-offs are real: custom tank infrastructure, dielectric-fluid logistics, service complexity, and — for some two-phase fluids — environmental and regulatory pressure around PFAS.
Overview#
Immersion cooling is the most thermally aggressive method commercially deployed in data centres. Where direct-to-chip liquid cooling places a cold plate on the hot dies and leaves the rest of the server to chassis air, immersion drops the whole motherboard into a tank of dielectric fluid that contacts every surface.
It is not new — Cray supercomputers used immersion in the 1980s, and crypto-mining drove a wave of single-phase deployments in the mid-2010s. What has changed is that hyperscaler interest (Meta, Microsoft) and the thermal demands of GB200-class systems have re-legitimised immersion as a mainstream option, while regulatory scrutiny on PFAS-based two-phase fluids has begun to reshape vendor choices.
Single-Phase vs Two-Phase#
The two technologies behave very differently from a facilities, fluids, and operational perspective.
| Property | Single-phase | Two-phase |
|---|---|---|
| Fluid state | Stays liquid throughout | Boils at chip; condenses on coil |
| Typical fluid | Synthetic hydrocarbon or PAO | Engineered fluorocarbon (some PFAS-class) |
| Tank type | Open-top or sealed, pumped | Sealed, gravity-driven |
| Heat rejection mechanism | Pumped through plate heat exchanger | Vapour condenses on chilled coil |
| Rack density ceiling | ~100-150 kW | ~200-250+ kW |
| Fluid cost | Low to moderate | High |
| Regulatory exposure (2026) | Low | Elevated — PFAS restrictions tightening in EU/UK |
| Servicing | Drip, drain, service | Tank seal must be broken under controlled conditions |
Several historically dominant two-phase fluids are PFAS-class chemicals. The EU's universal PFAS restriction proposal (under REACH) and similar UK consultations may materially constrain availability through the late 2020s. Confirm fluid sourcing and end-of-life disposal plans before committing to a two-phase deployment.
Typical Operating Specifications#
| Parameter | Single-phase | Two-phase |
|---|---|---|
| Fluid temperature (working) | 30-50 °C | Boil point typically 49-60 °C |
| Heat capture | ~95-100 % | ~100 % |
| Fan power | Zero (fans removed) | Zero (fans removed) |
| PUE achievable | 1.03-1.10 | 1.02-1.05 |
| Pump power | Modest — moves fluid through HX | Minimal — passive condensation |
| Tank footprint | Larger than rack — horizontal | Sealed vertical or horizontal tank |
Trade-offs#
- Density: immersion exceeds DLC at the top end — useful for GB200 NVL72 and beyond, marginal below 80 kW per rack.
- PUE: immersion routinely hits the lowest published PUE figures in the industry because fans and CRAH units are eliminated.
- Service model: pulling a server means lifting it from a fluid tank, draining it, and accepting some fluid loss. Hot-swap density and parts replacement are slower than air or DLC.
- Compatibility: not every component is rated for immersion. Spinning disks, mechanical fans, and certain optics fail or behave erratically. Server vendors increasingly offer 'immersion-ready' SKUs that bypass these issues.
- Fluid management: leak detection, viscosity tracking, particulate filtering, and biocide control all become operational disciplines.
- Regulatory exposure: as noted, two-phase PFAS fluids face restrictions; single-phase synthetic hydrocarbons do not, but they are flammable above their flash point and need fire-suppression engineering accordingly.
When to Use Immersion#
- You are deploying GB200 NVL72 or denser, and DLC alone cannot meet the thermal budget within the rack envelope.
- Site water is constrained — immersion's near-100 % heat capture means dry coolers can reject the load without evaporative cooling.
- You need PUE below 1.10 in a hot climate where free cooling on air is impractical.
- Acoustic constraints matter (immersion has no server fans, so the data hall is near-silent).
- Edge sites where containerised, sealed immersion tanks can be shipped pre-fluid-filled and dropped onto a pad.
Operational Pitfalls#
- Component qualification: confirm every NIC, NVMe drive, and optical transceiver is immersion-rated. Optics fluid wicking is a known failure mode.
- Fluid contamination: ingress of water, particulates, or solvents changes dielectric properties. Closed tanks and disciplined service procedures are essential.
- Service ergonomics: lifting a fluid-soaked server requires hoists and drip trays. Plan service bays and aisle widths accordingly.
- Fire codes: single-phase synthetic hydrocarbons have a flash point typically above 175 °C, but local fire codes may still require special suppression. Two-phase fluids do not burn but displace oxygen if released.
- End-of-life: spent fluid is a regulated waste stream. Build the disposal contract before the first fluid arrives on site.
References
- Open Compute Project — Immersion Cooling Sub-project · OCP
- ASHRAE TC 9.9 — Liquid Cooling Guidelines · ASHRAE
- ECHA PFAS Restriction Proposal (REACH Annex XV) · European Chemicals Agency