Liquid cooling has, over the past two years, become the default headline in offers from data center integrators. Let's not confuse trend with reality — for the majority of EU deployments in 2026, traditional precision air cooling (CRAC + cold aisle containment) is still both technologically and economically the correct choice. Where this breaks is defined by one parameter: rack density in kW. This article splits the decision around it.
Where the break-even lies — the one number that decides
Rack density is the reason anyone considers liquid cooling at all:
- **< 10 kW/rack** — air cooling, trivially. Standard precision AC (Vertiv Liebert PCW, Stulz CyberAir, Schneider InRow) with hot/cold aisle containment. Fully sufficient, no reason to consider liquids.
- **10–25 kW/rack** — air cooling **still possible**, but requires more aggressive cold aisle containment, higher flows, supplementary in-row cooling (rear-door heat exchangers). Cost +20–40% over standard configuration.
- **25–40 kW/rack** — air cooling **reaches the limits of physics**. Requires rear-door heat exchangers (CoolIT, Motivair) or overhead in-row units. Costs per kg start exceeding the liquid cooling alternative.
- **40–80 kW/rack** — direct-to-chip liquid (DLC) is **practically mandatory**. Air cooling here means 18–25 °C inlet air, which is no longer ASHRAE A1 compliant, and the thermal gradient across chip + heatsink + heatpipe + air is unattainable.
- **80–130 kW/rack** — single-phase **immersion or DLC + immersion hybrid**. This is the territory of fully populated NVIDIA HGX H200/B200 racks.
- **130+ kW/rack** — **two-phase immersion** (3M Novec, Asperitas, Submer SmartPodX). The only long-term sustainable solution.
If your densest rack is < 25 kW (typical enterprise IT, most virtualization workloads, databases, web servers), liquid cooling will cost you 2–3× more CAPEX without operational benefit. The investment **never returns**.
Three real use cases where liquid cooling wins
1. AI/ML training clusters with NVIDIA H100/H200/B200
An 8× H100 SXM5 rack draws 12–14 kW at full FP16 training. A fully populated SuperPOD rack with 8× HGX H200 (no free U) draws 35–42 kW. B200 and B300 in 2026 push this further — estimates of 50–65 kW per node. With more than 4 such racks in a room, **air cooling physically isn't enough**, because CRAC inlet air needs a temperature below 15 °C, which is beyond ASHRAE A1 and requires extreme chiller setups.
2. HPC simulation clusters with 64+ core CPUs + accelerators
Threadripper PRO 7995WX (350 W), Xeon Platinum 8592+ (350 W), Intel Granite Rapids HEDT (up to 400 W) at full load = 12–18 kW just in CPUs. Add 4× H100 PCIe = 12 kW in accelerators. Per-rack 35 kW. For this we would deploy **DLC** (Direct Liquid Cooling) with coldplates on CPU + coldplates on GPU.
3. Edge data centers with extreme density on a small footprint
Telco edge, retail edge, manufacturing edge. The client needs 100 kW IT load in a 40 ft container. Air cooling requires 35–40% of footprint for cooling infrastructure. Liquid cooling 10–15%. The container fits, or doesn't — that's what decides.
Three use cases where the client wants liquid cooling and SHOULDN'T have it
1. Regular enterprise data center
50–200 racks, average density 5–8 kW, peak 12 kW. Traditional CRAC + cold aisle containment + hot aisle containment + variable speed fans is 95% of liquid cooling performance for 40% of the cost. **PUE 1.3–1.4 in EU climate without free cooling, 1.15–1.25 with free cooling.** Better than most liquid cooling installations that "advertise" with PUE 1.1 but actually (after accounting for pumps, secondary CDU loop, evaporator) land at 1.2–1.25.
2. Hybrid DC with 5 AI racks and 50 regular server racks
The client hears "AI cluster, you need liquid". Reality: build a separate sub-room for the AI cluster with liquid cooling, leave the main hall on air cooling. Hybrid architecture. Mixed-mode where one CRAC piping system feeds 5 liquid racks and 20 air racks is an operations nightmare (different setpoints, different maintenance scheduling, suboptimal PUE).
3. Client with uncertain roadmap
"Maybe in 2 years we'll add AI servers". Build the proper air-cooled infrastructure now with **a reservation in the floor for liquid piping** (reserve 800–1,200 mm of straight path from chiller plant to AI hall). Install liquid cooling later, when you have specific GPUs purchased. Premature investment in liquid cooling for uncertain workload = dead capital assets for 18–30 months.
Types of liquid cooling — what to choose
Rear-door heat exchanger (RDHx)
- **What:** passive or active cooler built into the rack door that cools hot air after it leaves the rack
- **Capacity:** 30–60 kW per rack
- **Plus:** retrofit into existing air-cooled DC without changing servers, minimal CDU plumbing
- **Minus:** suboptimal PUE (1.3+), only reduces load on CRAC, doesn't eliminate it
- **When:** intermediate step toward higher density without changing servers
Direct Liquid Cooling (DLC) coldplate
- **What:** liquid flows through coldplates directly on CPU and GPU. Secondary loop (liquid ↔ air). Server interfaces with a rack-level CDU.
- **Capacity:** 60–150 kW per rack
- **Plus:** excellent PUE (1.1–1.15), compact, good serviceability (quick-disconnect couplings)
- **Minus:** requires servers with coldplate variant (Supermicro SYS-821GE, HPE Cray EX, Dell PowerEdge XE9680L), 2× CAPEX server vs. air variant, CDU per rack/per row
- **When:** AI training clusters, HPC
Single-phase immersion
- **What:** servers are submerged in dielectric fluid (ShellLubri, Submer SmartCool, Asperitas Adaptive)
- **Capacity:** 80–200 kW per tank
- **Plus:** best PUE (1.03–1.08), zero airflow, drastic noise reduction, high density
- **Minus:** complete change of server form factor (no fans, modified cabling), service means lifting the server out of the tank + draining fluid, compatible only with specific servers
- **When:** greenfield AI cluster, edge deployment with extreme density
Two-phase immersion
- **What:** the fluid **evaporates** on contact with the hot chip (Novec 7100, 7500), vapors condense on a cooling coil above the tank and drip back. Passive, gravity-driven.
- **Capacity:** 130–300 kW per tank
- **Plus:** most efficient, no pumps in the primary loop
- **Minus:** 3M Novec fluid costs €80–150 per liter, **3M announced discontinuation of Novec PFAS chemicals by 2025** — alternative fluids not yet mature. Plus: regulatory risk in the EU around PFAS chemicals.
- **When:** still niche, waiting for non-PFAS alternatives
Hidden costs that don't show up in the ROI calculator
1. Server vendor lock-in
Liquid cooling means servers must be compatible. Supermicro SYS-821GE-TNHR at $180k vs. air variant SYS-821GE-TN at $145k. 25% CAPEX premium that doesn't show up in the cooling ROI calculator.
2. Maintenance and operations training
Air cooling maintenance is trivial. Liquid cooling maintenance means: leak detection, monitoring fluid pH/conductivity, quick-disconnect couplings, specialized service personnel. **Training costs for the team €8,000–15,000 per engineer**, 6–12 months of hands-on practice.
3. Insurance + property risk
Liquid in IT space is a risk that insurers price in. Allianz, Munich Re require additional leak detection, automatic shut-off valves, drip trays — all CAPEX. Insurance premium may be 8–15% higher for the first 5 years.
4. Planning for fluid disposal
After 5–7 years the fluid degrades and needs replacement. 1,000 liters of dielectric at €80k + disposal €10k. Plan it into the TCO calculation.
5-minute decision tree
1. **Max rack density < 15 kW?** → Air cooling. No question. 2. **Max density 15–25 kW?** → Air cooling with rear-door heat exchanger as a migration step. 3. **Max density 25–40 kW?** → DLC or rear-door heat exchanger depending on which servers will be purchased. 4. **Max density 40–80 kW?** → DLC. No exception. 5. **Max density 80+ kW?** → Immersion (single-phase). Two-phase only if you have a 5+ year project timeline. 6. **Edge deployment with space constraints?** → Immersion or DLC depending on server compatibility. 7. **Vague AI roadmap?** → Air cooling now, plan a reserve for liquid retrofit.
Practical advice for tendering
In the tendering process always ask for **two proposals**: one with air cooling up to the maximum realistic density for your IT, one with liquid cooling. Compare: - CAPEX including servers (not just cooling infrastructure!) - 7-year TCO including maintenance, energy (PUE × IT load × energy price), insurance - Operational risk profile (number of people needed for 24/7 operation, single points of failure)
Vendors themselves will emphasize liquid cooling ROI over a 7-year horizon — and they're usually right at 30+ kW density. At < 25 kW density, air cooling wins consistently, because the vendor also factored in "opportunistic efficiency" from rising energy prices.
---
*We do data center design + build and air-to-liquid retrofit migration. If you're considering AI infrastructure investment above €500k, the first project assessment (4-hour workshop) walks through the two-stage decision matrix against your specific workloads.*