Skip to main content

Energy layer

PUE Comparison

Compare PUE benchmarks across hyperscale operators, regions, and cooling approaches.

The engineer question
What PUE is achievable for liquid-cooled AI facilities in 2026?

Inputs

Critical IT power, not total facility power. Used only for the energy-delta estimate.

Result

Typical median PUEindustry-typical, not your facility
1.16
Typical p25–p75 bandspread of published figures
1.06 – 1.28
Cooling-energy overhead1.6 MW at 10 MW IT
16% of IT load
Energy delta vs 1.5 baselinelower than baseline
−3.4 MW overhead
Annualized facility-energy delta100% duty, vs PUE 1.5
−29.9 k MWh/yr

Recommendation

1.16 is a strong, achievable target for direct-to-chip liquid (cold plate). Most of the remaining overhead is fans/pumps and power-distribution loss, not heat rejection. Liquid removes the air-handling penalty but you still pay for CDU pumping and dry-cooler fans in temperate (e.g. oregon, ireland) conditions.

Assumptions

  • · These are INDUSTRY-TYPICAL published ranges, NOT a measurement or audit of your facility. Sources: Uptime Institute Global Data Center Survey (global avg ~1.5–1.56), ASHRAE TC9.9 liquid-cooling guidance, and vendor datasheets (Vertiv / Schneider / Submer / GRC). All figures approximate, ±0.05–0.1 PUE.
  • · Base median for Direct-to-chip liquid (cold plate): 1.22 (p25 1.12 / p75 1.35) before adjustment.
  • · Climate adjustment: Temperate (e.g. Oregon, Ireland) applies a +0.00 PUE offset (free-cooling hours vs heat-rejection penalty).
  • · Operator-tier adjustment: Hyperscale (own design, high density) applies a ×0.95 factor (setpoint discipline + mechanical design maturity). Adjusted PUE is floored at 1.02.
  • · Energy delta = (PUE − 1) × IT load, compared with a PUE 1.5 reference, annualized over 8760 h at 100% duty. Real facilities vary with IT utilization, partial-load chiller efficiency, and seasonal swing.
  • · Excluded: WUE / water consumption, embodied carbon, grid carbon intensity, heat-reuse credits, partial-PUE vs total-PUE boundary definitions, and per-rack density (kW/rack) — all of which materially shift the real number.

Worked example (default inputs)

Result

Typical median PUEindustry-typical, not your facility
1.16
Typical p25–p75 bandspread of published figures
1.06 – 1.28
Cooling-energy overhead1.6 MW at 10 MW IT
16% of IT load
Energy delta vs 1.5 baselinelower than baseline
−3.4 MW overhead
Annualized facility-energy delta100% duty, vs PUE 1.5
−29.9 k MWh/yr

Recommendation

1.16 is a strong, achievable target for direct-to-chip liquid (cold plate). Most of the remaining overhead is fans/pumps and power-distribution loss, not heat rejection. Liquid removes the air-handling penalty but you still pay for CDU pumping and dry-cooler fans in temperate (e.g. oregon, ireland) conditions.

Assumptions

  • · These are INDUSTRY-TYPICAL published ranges, NOT a measurement or audit of your facility. Sources: Uptime Institute Global Data Center Survey (global avg ~1.5–1.56), ASHRAE TC9.9 liquid-cooling guidance, and vendor datasheets (Vertiv / Schneider / Submer / GRC). All figures approximate, ±0.05–0.1 PUE.
  • · Base median for Direct-to-chip liquid (cold plate): 1.22 (p25 1.12 / p75 1.35) before adjustment.
  • · Climate adjustment: Temperate (e.g. Oregon, Ireland) applies a +0.00 PUE offset (free-cooling hours vs heat-rejection penalty).
  • · Operator-tier adjustment: Hyperscale (own design, high density) applies a ×0.95 factor (setpoint discipline + mechanical design maturity). Adjusted PUE is floored at 1.02.
  • · Energy delta = (PUE − 1) × IT load, compared with a PUE 1.5 reference, annualized over 8760 h at 100% duty. Real facilities vary with IT utilization, partial-load chiller efficiency, and seasonal swing.
  • · Excluded: WUE / water consumption, embodied carbon, grid carbon intensity, heat-reuse credits, partial-PUE vs total-PUE boundary definitions, and per-rack density (kW/rack) — all of which materially shift the real number.

Related tools in the Energy layer

Get notified when PUE Comparison numbers update

We refresh the inputs as the market moves. One email when they change.

→ Get this data as JSON