Terrestrial Data Center PUE

What is the PUE for modern liquid-cooled AI data centers?

Answer

Modern liquid-cooled AI data centers achieve PUE values ranging from 1.02 to 1.20, depending on cooling architecture. The central estimate for a new-build, liquid-cooled AI facility is PUE 1.10. Optimistic deployments using full immersion cooling reach 1.03, while retrofitted or hybrid-cooled facilities typically land around 1.20.

The industry is converging on direct-to-chip liquid cooling as the standard for AI racks above 35 kW, with immersion cooling gaining share at the frontier. New-build AI data centers universally adopt liquid cooling; the GB200 NVL72 mandates it at 120 kW/rack.

Analysis

Convergence toward a standard range

The AI data center industry is converging on PUE 1.05-1.15 for new liquid-cooled facilities:

• Direct-to-chip liquid cooling is the dominant approach, commanding ~47% of the liquid cooling market as of late 2025 introl-liquid-cooling.3. It achieves PUE 1.05-1.15 introl-liquid-cooling.1 and is the cooling method mandated by NVIDIA for Blackwell and Rubin platforms introl-nvl72-deployment.1.

• Immersion cooling achieves the lowest PUE (1.02-1.05) but remains a smaller share of deployments due to higher complexity, custom enclosure requirements, and challenges with serviceability.

• Air cooling is effectively obsolete for AI workloads above 35 kW/rack. Since current-generation AI racks (GB200 NVL72) draw 120 kW and next-generation (Vera Rubin NVL72) will draw 180-220 kW, air cooling is not a viable option.

Why PUE 1.10 is the central estimate

The central estimate of 1.10 reflects:

1. Hyperscaler fleet averages of 1.09-1.16 include older air-cooled facilities. New AI-specific builds outperform fleet averages.
2. SemiAnalysis modeling uses 1.15 for a facility with adiabatic cooling assist, which is slightly conservative for pure liquid-cooled deployments.
3. ChinaTalk modeling uses 1.11 as representative for a modern AI data center, which aligns closely with the central estimate.
4. Direct-to-chip systems routinely achieve 1.05-1.15; the midpoint is ~1.10.

PUE overhead components at 1.10

At PUE 1.10, the non-IT overhead is 10% of IT load. No single source provides a component-level breakdown for a liquid-cooled facility at this PUE, but the approximate allocation can be derived from equipment specifications: the GB200 NVL72 power architecture achieves 97% conversion efficiency introl-nvl72-deployment.4, implying ~3% loss for power conversion alone; modern UPS systems add another 1-5% loss depending on mode (ENERGY STAR). The remainder goes to cooling pumps, CDUs, and facility systems. A representative allocation:

• Cooling distribution pumps and CDUs: ~4-5% (editorial estimate)
• Power conversion losses (UPS, PDU, transformers): ~3-4% (consistent with 97% rack-level efficiency introl-nvl72-deployment.4 plus upstream UPS/transformer losses)
• Lighting, security, facility systems: ~1-2% (editorial estimate)

Implications for orbital comparison

PUE 1.10 means a terrestrial facility needs 1.10 kW of total power for every 1.00 kW of IT load. Per the component breakdown above, cooling accounts for roughly 4-5% of IT load, which is 4-5% of total facility power (since at PUE 1.10, IT load is 1.00/1.10 = ~91% of total). This sets a high bar for orbital data centers: eliminating cooling overhead saves only ~4-5% of total power, not the ~40% that air-cooled facilities from a decade ago would have suggested. The case for orbital data centers must rest on power generation cost advantages, not cooling efficiency gains.

Evidence

Cooling technology PUE ranges

Air PUE 1.4-1.8; liquid cooling PUE 1.05-1.15; immersion PUE 1.02-1.03.

Direct-to-chip cooling commands a dominant 47% market share of the liquid cooling market as of late 2025, with Microsoft beginning fleet deployment across Azure campuses.

Hyperscaler reported fleet PUE (2024)

These are fleet averages from the most recent directly-reported company disclosures (sustainability reports, earnings calls) and may differ slightly from values in older third-party compilations. New-build AI-specific facilities achieve lower PUE than these fleet averages.

Hyperscaler average PUEs: AWS 1.15, Google 1.10, Microsoft 1.18, Meta 1.08. PUE of ~1.11 representative for modern AI facility. (Note: these values are from ChinaTalk's publication date and differ slightly from the table above, which uses the most recent directly-reported figures. The discrepancies — e.g., Google 1.10 vs. 1.09, Meta 1.08 vs. 1.09, Microsoft 1.18 vs. 1.16 — reflect different reporting periods and rounding.)

Uses PUE of 1.15 for Colossus 2 modeling (400 MW AI data center in Memphis).

New builds vs. retrofits

New-build AI data centers designed for liquid cooling from the ground up achieve PUE 1.05-1.12. They eliminate the overhead of maintaining parallel air-cooling infrastructure and can optimize facility power distribution for liquid-cooled racks.

Retrofitted facilities face higher PUE (1.15-1.25) due to hybrid cooling architectures, suboptimal airflow management around remaining air-cooled equipment, and legacy power distribution inefficiencies. Retrofitting to support 40 kW racks costs $50K-100K per rack; building new 100 kW infrastructure costs $200K-300K per rack.

NVIDIA mandated cooling specs

NVIDIA mandates liquid cooling for GB200 NVL72: inlet temperature 20-25C, flow rate 80 L/min, pressure drop <1.5 bar. The system generates 120 kW continuously. Deviation triggers automatic throttling that can reduce performance by 60%.