Technical Manual · Cooling Comparison
Air vs Liquid Cooling — Comparison Methodology
The editorial basis, cited numeric ranges, and physics behind the air vs liquid cooling comparison page. This is an editorial reference — every number is a cited range from published sources (ASHRAE, Uptime Institute, IEA, NVIDIA), not a dynamic solver. The companion page does not accept user inputs; it presents industry-sourced ranges with an honest uncertainty envelope.
▶ Open the comparison page01 Purpose & editorial basis
The Air vs Liquid Cooling comparison page is an editorial reference, not a dynamic calculator. It does not accept user inputs and does not produce bespoke outputs. Its purpose is to present industry-sourced numeric ranges that describe the performance envelope of each cooling architecture across seven dimensions: rack density, PUE, CAPEX, OPEX, retrofit complexity, acoustic noise, and AI/HPC readiness.
All cited figures are ranges drawn from published sources — ASHRAE liquid-cooling guidelines, Uptime Institute annual PUE surveys, NVIDIA DGX product specifications, and IEA water-stress research. No single facility is guaranteed to fall at any specific point within a range; site climate, design choices, utilisation, and workload mix shift the actual outcome. This manual documents which source each range originates from and what physics underlies the gap between the two architectures.
Editorial scope: this page documents the comparison basis. It is not a cooling-design tool, not a financial model, and not a prescriptive recommendation. Use it to orient a technology decision, then engage a qualified mechanical engineer for facility-specific design.
02 Inputs
This page is a static editorial reference — there are no user-facing inputs. The comparison page presents fixed cited ranges, not a parameterised model. The table below documents what each comparison row measures and where its numeric range originates, so readers can evaluate the basis independently.
| Row | What is measured | Air value (cited range) | Liquid value (cited range) |
|---|---|---|---|
| Rack density | Peak sustained IT power per rack (kW) | 15–20 kW/rack (with containment) | 40–120+ kW/rack (DLC); 100–200+ kW (immersion) |
| PUE | Total facility power ÷ IT power | 1.30–1.60 typical; 1.20 hyperscaler floor | 1.02–1.15 (DLC with dry cooler or free cooling) |
| CAPEX | Cooling infrastructure cost per MW IT load | $2–4 M/MW | $3–6 M/MW (including CDUs and piping) |
| OPEX | Relative cooling energy consumption | Higher — fan power + overcooling losses | 30–50% energy reduction vs air baseline |
| Retrofit | Ease of upgrading an existing facility | Standard — no special infrastructure | Moderate (RDHx) to complex (full DLC or immersion) |
| Acoustic noise | Typical in-row noise level (dBA) | 70–85 dBA at rack level | 40–55 dBA (server fans reduced or eliminated) |
| AI/HPC readiness | Ability to support 40+ kW GPU racks | No — physical limit ~20 kW/rack with air | Yes — designed for 40–200+ kW racks |
03 Methodology — comparison metrics & physics
Each comparison row rests on a physical or financial argument. This section documents those arguments and their mathematical basis so the cited ranges can be evaluated critically.
Heat-transfer physics: why water beats air
Cp_water = 4.186 kJ/(kg·K) (specific heat capacity at 20 °C)
Cp_air = 1.007 kJ/(kg·K) (dry air at standard conditions)
ratio_mass = Cp_water / Cp_air ≈ 4.2× (per unit mass)
ρ_water ≈ 998 kg/m³ ρ_air ≈ 1.20 kg/m³
ratio_vol = (Cp_water × ρ_water) / (Cp_air × ρ_air) ≈ 3 460×
Volumetric heat capacity of water vs air: ~3,500× higher. A 25 mm pipe flowing water at 0.5 L/s removes the same heat as a 600 mm duct at 5 m/s. This is the physical root of every PUE and density advantage cited.CITED-RANGE · Engineering thermodynamics
PUE range basis
PUE = Total_Facility_Power / IT_Equipment_Power (Uptime/ASHRAE definition)
Air-cooled PUE range: 1.30 – 1.60 [CITED-RANGE — Uptime Institute Global Survey]
DLC PUE range: 1.02 – 1.15 [CITED-RANGE — ASHRAE TC 9.9 DLC guidelines]
PUE_delta_low = 1.30 − 1.15 = 0.15 (conservative gap)
PUE_delta_high = 1.60 − 1.02 = 0.58 (maximum gap)
PUE improvement arises from (a) eliminating CRAH fan power (~5–15% of total facility power in air-cooled facilities), (b) enabling higher chilled-water supply temperatures (40–45 °C vs 12–15 °C), which allows free cooling or dry coolers to work more hours per year, and (c) reducing cold-aisle bypass and hot-aisle recirculation losses. The ranges are medians and percentiles from annual industry surveys — individual facilities vary significantly.CITED-RANGE · Uptime Institute, ASHRAE TC 9.9
Rack density ceiling — air vs liquid
Air density ceiling (with containment): 15–20 kW/rack [CITED-RANGE — ASHRAE A5.1-2021]
RDHx (rear-door HX) ceiling: 30–40 kW/rack [CITED-RANGE — vendor specifications]
Direct-to-chip (cold plate) ceiling: 40–120+ kW/rack [CITED-RANGE — NVIDIA DGX, OCP specs]
Immersion (single/two-phase) ceiling: 100–200+ kW/tank [CITED-RANGE — immersion vendor data]
The air ceiling is not a design choice — it is a physical limit set by the maximum airflow a rack can deliver without unacceptable backpressure. NVIDIA DGX H100 (10.2 kW/node × 4 nodes = ~40 kW + networking) and GB200 NVL72 (cited >120 kW) demonstrate that AI GPU racks already exceed the air-cooling ceiling at moderate cluster scale.CITED-RANGE · ASHRAE A5.1-2021, NVIDIA DGX H100/GB200 specs
PUE payback methodology
annual_opex_savings = IT_kW × 8760 h/yr × PUE_delta × energy_rate_$/kWh
payback_years = capex_premium_$ / annual_opex_savings
capex_premium = liquid_capex − air_capex. PUE_delta and capex figures are each cited ranges, so payback is itself a range. The editorial page cites 3–5 years for high-density AI/HPC deployments — this assumes PUE_delta ~0.30, IT load 1 MW, energy rate $0.08/kWh, and a capex premium of ~$2 M: savings = 1000 × 8760 × 0.30 × 0.08 ≈ $210 K/yr → payback ≈ $2 M / $210 K ≈ 9.5 yr at the low end; or $500 K/yr at PUE_delta 0.50 → ~4 yr. The published 3–5 yr figure typically reflects higher density (higher PUE_delta) and lower energy rates.CITED-RANGE · Illustrative range; not a guarantee
Water consumption delta
Air-cooled (evaporative tower): WUE ≈ 1.5 – 3.0 L/kWh IT [CITED-RANGE — IEA / Uptime]
10 MW IT load, 8760 h/yr: water = 10 000 × 8760 × 2.0 L/kWh ≈ 175 ML/yr
DLC with dry cooler (closed loop): WUE ≈ 0.0 – 0.5 L/kWh IT [CITED-RANGE — ASHRAE TC 9.9]
Air cooling with evaporative towers consumes make-up water to replace evaporation losses (typically 1.5–3.0% of circulated volume). DLC with dry coolers (air-blast heat rejection) operates in a closed loop and can achieve near-zero water consumption. In water-stressed regions, WUE is a permit and regulatory constraint — liquid cooling enables compliance that air cooling with towers cannot achieve.CITED-RANGE · IEA Water Stress Index, ASHRAE TC 9.9
04 Constants & sources
Every cited range in the comparison page maps to one of the following source groups. The numeric range boundaries are the published medians or percentile bands from those sources — they are not derived from a proprietary model.
| Metric | Air (cited range) | Liquid (cited range) | Primary source |
|---|---|---|---|
| PUE — global average | ~1.58 | — | Uptime Institute Global Data Center Survey (annual) |
| PUE — hyperscaler best-in-class (air + free cooling) | ~1.20 | — | Uptime Institute / Google/Microsoft published PUE |
| PUE — DLC (direct liquid cooling) | — | 1.02–1.15 | ASHRAE TC 9.9, Liquid Cooling Guidelines (2021) |
| Rack density — air (with containment) | 15–20 kW/rack | — | ASHRAE A5.1-2021 § Rack Cooling Design |
| Rack density — cold plate / DLC | — | 40–120+ kW/rack | NVIDIA DGX H100 spec (10.2 kW/node); GB200 NVL72 (>120 kW); OCP Liquid Cooling spec |
| Rack density — immersion | — | 100–200+ kW/tank | Submer, GRC, LiquidStack vendor data sheets |
| CAPEX — air cooling | $2–4 M/MW | — | Turner Construction DC Cost Index; Uptime Institute cost benchmarks |
| CAPEX — liquid cooling | — | $3–6 M/MW | Turner Construction DC Cost Index (liquid DLC premium cited range) |
| OPEX reduction — liquid vs air | baseline | 30–50% cooling energy reduction | ASHRAE TC 9.9; vendor case studies (Dell, HPE, Vertiv) |
| WUE — evaporative air cooling | 1.5–3.0 L/kWh | — | IEA "Water Consumption of Data Centers" (2022) |
| WUE — DLC closed-loop | — | 0.0–0.5 L/kWh | ASHRAE TC 9.9; Green Grid WUE metric definition |
| Acoustic noise — air-cooled rack | 70–85 dBA | — | ASHRAE equipment noise class ratings |
| Acoustic noise — liquid-cooled (reduced fans) | — | 40–55 dBA | Vendor measured data (Supermicro, HPE DL380 Gen11 with DLC) |
| Specific heat capacity — water | 4.186 kJ/(kg·K) at 20 °C | NIST Chemistry WebBook; engineering thermodynamics tables | |
| Specific heat capacity — air (dry) | 1.007 kJ/(kg·K) at 20 °C, 1 atm | NIST Chemistry WebBook; ASHRAE Handbook of Fundamentals | |
05 Outputs — comparison table rows
The comparison page outputs seven rows. The table below maps each row to its comparison basis, the cited range on each side, and the directional verdict (which architecture wins on that metric for high-density AI/HPC deployments).
| Row | Air cooling | Liquid cooling | Directional winner | Basis |
|---|---|---|---|---|
| Max rack density | 15–20 kW/rack | 40–120+ kW/rack | Liquid | Physical heat-removal capacity limit |
| PUE impact | 1.30–1.60 | 1.02–1.15 | Liquid | Fan power elimination + higher ΔT return water |
| CAPEX | $2–4 M/MW | $3–6 M/MW | Air (lower first cost) | Infrastructure procurement benchmarks |
| OPEX | Higher baseline | 30–50% reduction | Liquid | Fan energy + overcooling elimination |
| Retrofit ease | Standard — no special infra | Moderate–complex (piping, CDUs, floor load) | Air (simpler) | Facility modification scope assessment |
| Acoustic noise | 70–85 dBA | 40–55 dBA | Liquid | Server fan reduction / elimination in DLC |
| AI/HPC readiness | No (ceiling ~20 kW) | Yes (40–200+ kW) | Liquid | NVIDIA DGX density exceeds air ceiling |
Reading these ranges: "winner" reflects the dominant outcome for the metric in isolation. A deployment with rack densities below 15 kW may reasonably prefer air cooling on CAPEX and simplicity. The comparison is not a universal verdict — it is a structured summary of the tradeoffs at the density levels relevant to AI/HPC infrastructure (40–120+ kW/rack).
06 Worked example — density-transition & PUE-payback illustration
A 1 MW IT load data hall contemplating a technology transition from air to direct liquid cooling. All figures use the midpoints of the cited ranges above. This is an illustrative exercise using cited-range midpoints — not a site-specific financial model.
- Current state (air cooling): 1 000 kW IT load at PUE midpoint
1.45→ total facility power = 1 450 kW; cooling overhead = 450 kW. - Target state (DLC): same IT load at PUE midpoint
1.08→ total facility power = 1 080 kW; cooling overhead = 80 kW. - Annual energy saving:
(1 450 − 1 080) kW × 8 760 h/yr = 3 241 200 kWh/yr→ at $0.08/kWh → ~$259 K/yr saved. - CAPEX premium (cited range midpoint): liquid capex midpoint $4.5 M/MW − air capex midpoint $3.0 M/MW = $1.5 M premium for this 1 MW build.
- Simple payback:
$1.5 M / $259 K/yr =~5.8 years at cited-range midpoints. - Density unlock: air ceiling ~18 kW/rack (midpoint) → at 1 MW IT load:
1 000 kW / 18 kW =~56 racks needed. DLC at 80 kW/rack:1 000 kW / 80 kW =~13 racks — a 4.3× floor-space reduction. - Water consumption comparison: air (evaporative tower, WUE 2.25 midpoint):
1 000 kW × 8 760 h/yr × 2.25 L/kWh =19.7 ML/yr. DLC (closed loop, WUE ~0.1): ~0.9 ML/yr → 95% reduction.
Illustrative note: all values are midpoints of cited ranges — actual outcomes depend on local energy cost, climate (free-cooling hours per year), utilisation factor, lease vs own, and workload profile. The payback range using the full cited-range boundaries spans approximately 2.5–12 years; a formal financial model with site-specific inputs is required for a capital decision.
07 References & standards
- ASHRAE TC 9.9 (2021) — Liquid Cooling Guidelines for Datacom Equipment Centers: PUE ranges for DLC facilities, rack density ceilings for cold-plate and immersion cooling, WUE targets for closed-loop systems.
- ASHRAE A5.1-2021 — Thermal Guidelines for Data Processing Environments (5th edition): air-cooling rack density ceiling at 15–20 kW/rack with containment.
- ASHRAE Handbook of Fundamentals — Specific heat capacity and density values for air and water at standard conditions.
- Uptime Institute Global Data Center Survey (annual) — Global average PUE (~1.58), hyperscaler PUE distribution, CAPEX/OPEX benchmark data.
- NVIDIA DGX H100 Product Specification — 10.2 kW per DGX H100 node (8× H100 SXM5 GPUs); rack-level density 45–50 kW for a 4-node configuration with networking.
- NVIDIA GB200 NVL72 System Specification — Rack thermal design power exceeding 120 kW; requires direct liquid cooling.
- IEA "Data Centres and Data Transmission Networks" (2022) — Water consumption of data centers: WUE ranges for evaporative cooling, water-stress context by region.
- Open Compute Project (OCP) Liquid Cooling Specification — Rack density targets and CDU (Coolant Distribution Unit) interface standards for AI/HPC deployments.
- Turner Construction DC Cost Index (2024) — CAPEX ranges for air-cooled and liquid-cooled data center construction per MW IT load.
- The Green Grid — PUE and WUE metric definitions; industry best-practice benchmarks for both metrics.
08 Assumptions & limitations
The comparison page is an editorial summary of industry-sourced ranges. The following boundaries define what it does and does not represent:
- Cited ranges, not prescriptive values. Every figure is a range from a published source. No single facility is guaranteed to land at any specific point within that range.
- No user inputs. The comparison page does not compute site-specific outcomes. It is a reference page, not a calculator.
- Technology snapshot. Ranges reflect published data as of 2024–2025. Cooling technology is evolving; specific numbers from next-generation GPU platforms (beyond GB200) may shift the density ceiling upward.
- Currency of CAPEX estimates. Cost ranges are USD and reflect Western-market (North America, Western Europe) procurement. APAC and emerging-market costs differ.
- Energy rate assumption. The worked example uses $0.08/kWh as an illustrative rate. Indonesian commercial electricity is approximately Rp 1 100–1 500/kWh; adjust the payback accordingly.
- Not investment or engineering advice. This page is an educational reference. Facility-specific decisions require a qualified mechanical engineer, thermal analysis, and site-survey data.
- Hybrid architectures not separately modelled. Rear-door heat exchangers (RDHx), in-row cooling, and partial DLC retrofits occupy a spectrum between the two poles; their PUE and cost outcomes are not separately enumerated in the comparison rows.
Data declaration: all numeric ranges are cited from the external sources listed in Section 07. No proprietary model, survey instrument, or database operated by ResistanceZero was used to generate these ranges. Where sources disagree, the range boundaries reflect the spread across sources.