Immersion Cooling for AI Workloads:
Single-Phase vs Two-Phase
and the Infrastructure Implications
AI GPU racks at 50–100 kW cannot be cooled by air. Immersion cooling removes heat directly from the chip — but single-phase and two-phase systems have radically different facility infrastructure requirements.
The AI infrastructure buildout is breaking the assumptions that data centre MEP engineering has been built on for 30 years. GPU clusters for large language model training regularly exceed 50 kW per rack. The most dense AI accelerator configurations reach 100–120 kW per rack. Air cooling — even in the most optimised configurations, with best-in-class containment, high-efficiency CRAH units, and precision underfloor delivery — cannot remove heat at this density. The physics simply do not permit it.
Immersion cooling removes heat directly from electronic components by submerging them in a thermally conductive, electrically non-conductive dielectric fluid. Heat transfers from the component directly to the fluid — which is then cooled externally. There are no fans, no airflow paths, no hot aisle containment strategies. The thermal resistance between chip and coolant is orders of magnitude lower than between chip and air. This is why immersion cooling is the only technically viable path for the highest-density AI workloads — and why every hyperscaler building AI infrastructure at scale is evaluating or deploying it.
Why Air Cooling Fails Above 20–25 kW/Rack
The fundamental constraint is the heat capacity of air. Air has a specific heat capacity of approximately 1.006 kJ/kg·K and a density of 1.2 kg/m³ at standard conditions. Moving enough air through a rack to remove 50 kW of heat while maintaining an acceptable temperature rise requires airflow volumes that create unmanageable velocity, noise, and pressure drop conditions in the server room.
The airflow arithmetic: To remove 50 kW with a 15°C supply/return air temperature differential requires approximately 2,800 m³/hr of airflow through a single rack. A standard rack is 600mm wide × 1000mm deep. The resulting air velocity through the rack is over 12 m/s — generating noise exceeding NC 65 and back-pressure that defeats server fan designs. This is why ‘high-density’ air-cooled data centre designs typically cap at 20–25 kW/rack, not 50–100 kW.
Single-Phase vs Two-Phase Immersion: The Technology Choice
Two distinct immersion cooling technologies are in commercial deployment. They share the same fundamental principle — submerging electronics in dielectric fluid — but differ in how the fluid absorbs and rejects heat.
| Parameter | Single-Phase Immersion | Two-Phase Immersion |
|---|---|---|
| Cooling mechanism | Sensible heat (fluid temperature rises) | Latent heat (fluid boils and condenses) |
| Typical fluid | Mineral oil, synthetic ester, engineered hydrocarbon | Fluorocarbon or fluoroketone (3M Novec / Vertiv) |
| Fluid GWP | Low (mineral oil) to moderate | High (fluorocarbons) to low (fluoroketones, GWP≈1) |
| Operating temperature | Fluid at 20–40°C | Fluid boils at 49–60°C |
| Heat rejection temperature | 40–50°C (warm water cooling possible) | 55–65°C (high-grade heat recovery feasible) |
| Pump requirement | Warm fluid circulated by pump | Natural circulation by vapour/condensate cycle |
| Capital cost | Lower — simpler tanks and CDU | Higher — sealed vapour containment; condensers |
| IT equipment compatibility | Wide — most servers with minor modification | Narrower — sealed enclosures required for some designs |
| Fluid consumption / loss | Low — open tanks with management | Minimal — sealed system with recovery |
| Deployments at scale | Many — hyperscaler pilots and production | Growing — newer technology; fewer large deployments |
Single-Phase Immersion: Infrastructure Implications
In single-phase immersion cooling, servers are submerged horizontally in open tanks filled with dielectric fluid. The fluid circulates by pump through an external cooling distribution unit (CDU) where it transfers heat to a facility chilled water or warm water loop. The fluid returns cooled to the tank.
The Tank and CDU
Each tank holds 10–20 servers (depending on tank size and server orientation). The CDU couples the warm dielectric fluid loop to the facility cooling water loop via a plate heat exchanger. No refrigeration is required — the facility water loop can be at 40–45°C, enabling cooling tower free cooling in most climates year-round.
Water-Side Economiser Integration
Because single-phase immersion operates at higher fluid temperatures (outlet at 40–50°C), the facility cooling water can be supplied by a cooling tower without mechanical refrigeration in ambient temperatures up to approximately 30–35°C wet-bulb. Delhi, Bengaluru, and Pune achieve significant free cooling hours annually — dramatically reducing PUE.
Server Preparation
Servers must have spinning hard drives replaced with SSDs (mineral oil damages bearing seals), fan connectors capped or fans removed (no airflow required), and any conformal coating or moisture barriers verified. Most major server manufacturers now offer immersion-ready configurations.
Monitoring and Fluid Management
Fluid cleanliness (particle count, water content, dielectric strength) must be monitored. Fluid top-up for evaporation losses. Oil analysis every 6–12 months. These are manageable maintenance activities compared to the distributed fan replacement programme of equivalent air-cooled infrastructure.
Two-Phase Immersion: Heat Recovery and Efficiency
In two-phase immersion, servers are submerged in a fluid with a low boiling point (49–60°C). The fluid absorbs heat from components, boils, rises as vapour, and condenses on a cooled condenser at the top of the sealed tank. The condensed liquid drips back onto the servers. This natural thermosiphon cycle requires no fluid pump.
Waste heat recovery advantage: Two-phase systems reject heat at 55–65°C — high enough for direct use in building heating systems, district heating networks, or absorption chillers. A 10 MW AI data centre operating at 55°C rejection temperature can export 8–9 MW of usable heat — offsetting gas consumption in adjacent buildings or processes. This transforms the data centre from a pure energy consumer to a net heat supplier.
The environmental profile of two-phase fluid is critical. Early deployments used fluorocarbon fluids (3M FC-72) with GWP of 8,000–9,000 — catastrophically high if any vapour loss occurred. The industry has shifted toward fluoroketone fluids (3M Novec 649, Solvay Galden) with GWP below 1. Fluid selection must consider both performance and long-term environmental compliance.
Facility Infrastructure Changes for Immersion Cooling
Deploying immersion cooling is not simply replacing server racks with tanks. It requires rethinking every element of the data hall infrastructure — structural loading, electrical distribution, cooling water design, containment, and fire suppression.
- 01
Structural Loading
Immersion tanks with fluid weigh 800–1,200 kg/m² — 4–6× the floor loading of conventional raised-floor data centres (typically 200–300 kg/m²). New construction can design for this; retrofit of existing facilities requires structural assessment and potentially additional column support.
- 02
Electrical Distribution
No in-rack cooling, no hot aisle containment, no CRAH units. The data hall electrical distribution simplifies significantly. Power is delivered directly to tank-mounted bus bars. Floor loading from cooling equipment is eliminated but the high-density racks draw 50–100 kW each — power distribution capacity per rack position increases substantially.
- 03
Cooling Water Distribution
High-temperature cooling water (40–45°C supply) replaces chilled water (6–7°C). Insulation requirements are different. Pipe materials must be compatible with warm-water service. The chiller plant may be eliminated entirely for single-phase deployments with adequate free cooling hours.
- 04
Fire Suppression
Mineral oil and synthetic ester are combustible — fire risk classification changes compared to a conventional data centre. Clean agent fire suppression per NFPA 2001 is typically retained, but the fire risk assessment must account for the tank construction and fluid fire point. Tank lids and containment berms are required.
- 05
Fluid Containment
Secondary containment for tank overflow and fluid spills — similar to transformer oil containment requirements. Floor containment bunding, drain sumps, and fluid recovery systems must be designed.
The KVRM Approach to Immersion Cooling Projects
- 01
Workload and Density Assessment
We begin with the IT team’s rack density projections — current, 3-year, and 5-year. Where peak rack densities exceed 20 kW, immersion is evaluated alongside hybrid (air + rear-door/direct liquid) as a solution pathway.
- 02
Fluid and Technology Selection
Single-phase vs two-phase evaluated against cooling water temperature availability, heat recovery opportunity, fluid environmental profile, and IT equipment compatibility. Life cycle cost comparison developed.
- 03
Structural and Civil Assessment
Floor loading, ceiling height (tank craneage), and drainage requirements assessed. For retrofit projects, structural survey commissioned before tank layout is finalised.
- 04
Cooling Water System Design
CDU sizing, warm water distribution design, free cooling hours calculation for the specific site location, and integration with existing chiller plant (if partial immersion deployment).
- 05
Regulatory and Fire Safety
Fire risk assessment for the new fluid type. Clean agent suppression system design updated. Insurance and PESO/CCOE regulatory requirements confirmed.
Conclusion: Immersion Cooling Is Not the Future — It Is the Present
For AI workloads above 30 kW/rack, immersion cooling is not a technology choice between options of equal merit. It is the only engineering solution that works at the densities the hardware requires. The question for data centre operators is not whether to adopt immersion cooling for high-density AI infrastructure, but how quickly they can develop the engineering capability to deploy it.
The facilities that will serve the next generation of AI infrastructure are being designed today. The engineers specifying those facilities need to understand immersion cooling not as a future consideration but as a current requirement — one that changes every element of the MEP design from structural loading to fire suppression, from cooling water temperature to electrical distribution architecture.
Designing AI Infrastructure with High-Density Cooling?
KVRM provides immersion cooling MEP design — technology selection, tank layout, warm-water cooling systems, structural assessment, and fire suppression design — for AI data centres and GPU cluster facilities.
Request a Free Consultation →