PUE (Power Usage Effectiveness)

Overview#

Power Usage Effectiveness is the dominant operational efficiency metric in the data centre industry. The Green Grid published the first PUE white paper in 2007, the metric was formalised as ISO/IEC 30134-2 in 2016, and today every credible operator — hyperscale, colocation, neocloud, or enterprise — publishes a PUE number against their facilities. The arithmetic is deliberately simple: PUE = total facility energy / IT equipment energy. A PUE of 2.0 means that for every kilowatt delivered to IT, the facility burns another kilowatt on cooling, power conversion, lighting, security, and ancillaries. A PUE of 1.10 means just 100 watts of overhead per kilowatt of IT. At a 20 MW IT load, the difference between PUE 1.5 and PUE 1.15 is approximately 7 MW of continuous facility draw — roughly $6-12 million per year at typical commercial electricity prices, before any carbon accounting.

PUE matters for AI infrastructure even more than for general-purpose compute because the absolute energy numbers are larger and growing faster. A 100 kW HGX-B200 rack costs roughly $90,000-$140,000 per year in raw electricity alone at PUE 1.0; the facility overhead at PUE 1.5 adds another $45,000-$70,000 per rack per year, while PUE 1.15 limits the overhead to $13,500-$21,000. Across a 50-rack cluster, choosing a 1.15-PUE site over a 1.5-PUE site recovers $1.5-2.5 million per year — enough to fund several additional GPU procurements over a five-year refresh cycle.

This entry helps you read a published PUE number critically, decide what target to set for your own deployment, and connect PUE to the broader efficiency stack (WUE, CUE, free cooling, DLC, immersion). Yobitel NeoCloud's published per-region PUE figures, surfaced through the Omniscient Compute capacity feed, are designed to be compared on an apples-to-apples Category 2/3 basis — the comparator a serious AI procurement team should be running.

The Formula and What It Includes#

PUE is dimensionless — a ratio of energies, both measured in kWh. The numerator is the total energy crossing the data centre's utility boundary over the reporting period (typically 12 rolling months). The denominator is the IT energy measured at a defined point within the facility, which is where the three measurement categories diverge.

• Numerator (Total Facility Energy): utility-supplied electricity, on-site generation supplied to the facility, on-site fuel imported and consumed (diesel for testing, natural gas for CHP). Renewable purchase agreements are accounted at the contractual delivery point, not netted against on-site consumption for PUE purposes.
• Denominator (IT Equipment Energy): servers, storage, network gear — the equipment that does work. Where exactly this is measured (UPS output / PDU output / rack inlet) defines the PUE Category.
• Excluded from the boundary: tenant offices, training rooms, NOC space if it serves multiple sites, and anything not physically inside the operational data centre envelope.
• Included in the boundary but often debated: external lighting, perimeter security, water-treatment plant if dedicated to the data hall, on-site administrative loads. Disclose what is in or out.

PUE arithmetic — single facility, single month
================================================

  Total facility energy   = Utility kWh
                          + On-site generation kWh
                          + Imported fuel kWh (converted)

  IT equipment energy     = sum( PDU output kWh )   -- Category 2
                          = sum( Rack inlet kWh )   -- Category 3

  PUE                     = Total facility / IT equipment
  DCiE (deprecated)       = 1 / PUE        (legacy metric)

  Worked example (UK NeoCloud region, monthly):
    IT load (PDU)         = 14,800,000 kWh
    Cooling (DLC + dry)   =  1,480,000 kWh
    Electrical losses     =    370,000 kWh
    Lighting + BMS + sec  =     74,000 kWh
    Total facility        = 16,724,000 kWh

    PUE                   = 16,724,000 / 14,800,000
                          = 1.130

Measurement Categories (ISO/IEC 30134-2)#

ISO/IEC 30134-2 defines three measurement categories. Higher numbers represent more rigorous, more frequent measurement closer to the IT load. The category is part of any defensible PUE claim — a Category 1 number from a colocation marketing sheet is not comparable to a Category 3 number from a hyperscaler's sustainability report.

Instantaneous vs Annualised vs Design PUE#

PUE is reported in three different time horizons and the difference matters. Conflating them is the single most common source of misleading PUE claims.

• In a UK temperate climate, an annualised 1.15 facility might run instantaneous 1.05 in February and instantaneous 1.30 on the hottest July afternoon. The annualised average is what matters operationally; the peak matters for sizing.
• In a tropical climate (Singapore, Mumbai, Dubai), the seasonal swing is much smaller but the absolute PUE is materially higher because free cooling is harder. A 1.30 annualised PUE in Singapore can represent better engineering than a 1.20 annualised PUE in Stockholm.
• A facility at 30 % IT load typically runs 0.10-0.20 worse PUE than the same facility at 80 % load because fixed cooling and electrical overheads dominate. Design PUE figures are usually quoted at the design IT load — confirm utilisation before believing the number.

Industry Benchmarks by Facility Type#

Published PUE figures vary widely by facility type, climate, and vintage. The table below summarises 2024-2026 industry-typical ranges from Uptime Institute Global Data Center Surveys, EU CoC Data Centres, and operator sustainability reports.

What PUE Rewards (and What It Doesn't)#

PUE is a useful efficiency metric, but it measures one thing only: the overhead ratio. It is silent on every other dimension that matters for a complete sustainability and operations picture. Treating PUE as the only number — rather than the first of several — is the single biggest reporting failure pattern.

• Rewards: cooling efficiency, electrical efficiency, UPS topology, free cooling, hot-aisle containment, EC fans, warmer setpoints, busbar over cabling, eco-mode UPS.
• Does not reward: water consumption (use WUE), carbon intensity of grid mix (use CUE or renewable percentage), grid-services participation, IT utilisation (use TUE / ITUE), embodied carbon of construction, or how much useful compute the watts actually produce.
• Can be gamed: very low IT load makes the ratio look bad without indicating anything wrong; aggressive setpoints (raising data hall supply temperature past ASHRAE A1/A2) improve PUE at the cost of equipment lifetime and hot-spot risk; selectively excluding ancillary loads from the boundary; reporting best-month instead of annualised.
• PUE companion metrics that close the gap: WUE (water L/kWh), CUE (kgCO2e/kWh), TUE (combines PUE × ITUE to credit efficient IT itself), REF (renewable energy factor), ERF (energy reuse factor — credits waste-heat reuse).

Improvement Levers — How to Move PUE Down#

Most PUE improvement comes from cooling architecture. The pareto chart of efficiency capex is dominated by getting the cooling right; everything else is rounding error by comparison. The bullets below order levers from highest-impact to lowest, with the typical PUE delta each contributes.

• Cooling architecture (delta 0.20-0.40): adopt warm-water direct-to-chip, immersion, or hybrid DLC + RDHx. Mechanical compression cooling (chillers) is the single largest line item on a typical PUE breakdown — eliminating it through warm-water operation is the biggest single move available.
• Free cooling (delta 0.05-0.15): in temperate climates, dry coolers or adiabatic-assist coolers can reject heat year-round without a chiller. The UK, most of northern Europe, and parts of the US north-east and Pacific north-west are practical free-cooling climates for warm-water IT loops.
• Setpoints (delta 0.05-0.10): raising supply-air temperature from 22 °C to 27 °C cuts chiller hours by 30-50 % in temperate climates with no impact on modern servers (ASHRAE class A1/A2 allows continuous operation up to 32 °C inlet).
• Containment (delta 0.03-0.08): hot-aisle or cold-aisle containment eliminates bypass airflow and lets the cooling system run only as hard as it needs to. The cheapest single intervention available to a legacy site.
• Fan upgrades (delta 0.02-0.05): EC fans with variable-speed control replace constant-speed AC fans for 30-50 % fan-power savings.
• Electrical topology (delta 0.02-0.05): high-efficiency UPS (eco-mode, lithium-ion), 415/240 V distribution closer to the rack, and busbar over cabling all shave fractions off the denominator's losses.
• Waste-heat reuse (delta to ERF, not PUE directly): feed return water at 45-55 °C into district heating, greenhouse heating, or process pre-heat. Credited via the Energy Reuse Factor — ERF — rather than PUE, but materially improves the overall sustainability story.
• Lighting and ancillaries (delta 0.005-0.015): small in absolute terms but easy wins — LED, occupancy sensors, BMS scheduling.

Pair PUE with WUE, CUE, and Other Companion Metrics#

PUE is one of eight ISO/IEC 30134 series metrics that together describe data centre resource efficiency. Reporting PUE in isolation is the equivalent of reporting only revenue and ignoring profit — necessary but radically insufficient. The companion metrics below close the gaps PUE leaves open.

Sizing PUE Headroom for an AI Cluster#

PUE is not just a backwards-looking metric — it drives forward-looking capacity planning. The denominator of any data centre sizing exercise is the IT load you intend to land; the numerator is the facility envelope you have to budget for. PUE sets the multiplier between them. Worked examples below walk a 5 MW IT-load AI cluster through three PUE scenarios.

• 5 MW IT load at PUE 1.10 → 5.5 MW utility commit, 0.5 MW overhead. NeoCloud-class warm-water DLC + dry cooling, temperate UK/EU climate.
• 5 MW IT load at PUE 1.30 → 6.5 MW utility commit, 1.5 MW overhead. Modern AI-ready wholesale colocation, hybrid air + RDHx.
• 5 MW IT load at PUE 1.60 → 8.0 MW utility commit, 3.0 MW overhead. Legacy enterprise or retail colocation pressed into AI service.
• Annual energy delta between PUE 1.10 and PUE 1.60 for a 5 MW IT load at $0.20/kWh commercial: ~$22 million utility, ~$28 million utility, ~$36 million utility respectively. The PUE 1.10 site saves ~$14 million per year on facility overhead alone — and that's before carbon accounting.
• Utility-commitment headroom: most utility tariffs charge a demand component for the contracted peak even if it isn't used. A site with high PUE pays this twice: once on the IT load, once on the overhead. Sizing the utility commit to design PUE × design IT load is the prudent baseline.

# PUE-driven utility and operating-cost sizing — 5 MW IT cluster

it_load_mw          = 5.0                # IT design load
hours_per_year      = 8760
energy_price_usd_kwh = 0.20              # commercial UK/EU 2026

scenarios = {
    "NeoCloud DLC (PUE 1.10)":      1.10,
    "Modern wholesale (PUE 1.30)":  1.30,
    "Legacy enterprise (PUE 1.60)": 1.60,
}

for label, pue in scenarios.items():
    utility_mw = it_load_mw * pue
    annual_kwh = utility_mw * 1000 * hours_per_year
    annual_usd = annual_kwh * energy_price_usd_kwh
    overhead_mw = utility_mw - it_load_mw
    overhead_pct = (pue - 1) * 100
    print(f"{label:30s} | Utility {utility_mw:.2f} MW | "
          f"Overhead {overhead_mw:.2f} MW ({overhead_pct:.0f}%) | "
          f"${annual_usd/1e6:.1f}M/yr energy")

# Delta NeoCloud vs Legacy
delta_mw_yr = (1.60 - 1.10) * it_load_mw * 1000 * hours_per_year
delta_usd_yr = delta_mw_yr * energy_price_usd_kwh
print(f"\nAnnual savings, PUE 1.10 vs 1.60: ${delta_usd_yr/1e6:.1f}M/yr "
      f"({delta_mw_yr/1e6:.1f} GWh)")

Operational Pitfalls and How to Avoid Them#

Most PUE arguments are arguments about measurement honesty, not engineering. The pitfalls below are the ones that turn up most often in procurement reviews, regulator audits, and customer-facing sustainability disputes.

• Annualised vs instantaneous: a quoted PUE of 1.15 might be the best-month or instantaneous figure; the annualised number in a humid summer climate can be materially worse. Insist on 12-month rolling Category 2 or Category 3 averages and the underlying measurement boundary.
• Boundary disputes: include cooling-tower fan energy, exclude tenant-supplied UPS — the boundary choice can swing the number by 0.10 or more. Specify which ISO category and which boundary line was used.
• Low-load distortion: a half-empty hall has worse PUE than a full one because cooling overheads are largely fixed. New-build halls quote 'design PUE at full IT load', which can be optimistic for years while the hall fills. Ask for current actual PUE alongside design PUE.
• Setpoint chasing: raising temperatures improves PUE but risks reducing equipment lifetime and triggering thermal-throttling events. ASHRAE A1/A2 recommended envelopes (18-27 °C inlet) exist for a reason; allowed envelopes (up to 32 °C) sacrifice headroom for efficiency.
• Renewable accounting confusion: PPAs and unbundled certificates do not change PUE — PUE is a physics ratio, not a contractual one. A site can have PUE 1.50 and CUE near zero (100 % renewable PPA) or PUE 1.05 and CUE 0.4 (coal-grid Cat-3 hyperscaler). Report both; do not conflate.
• Metric saturation: below ~1.10, further PUE improvements come at disproportionate capex. Beyond a point, water (WUE), carbon (CUE), waste-heat reuse (ERF), and embodied energy of construction matter more than the ratio itself.
• Self-certification risk: PUE is operator-reported. Third-party certification (LEED, EU CoC for Data Centres, BREEAM, NCSC for sovereign sites) adds credibility. NeoCloud regions targeting NCSC OFFICIAL alignment publish independently audited PUE as part of the assurance pack.

Where PUE Fits in the Yobitel Stack#

Yobitel NeoCloud is engineered around an annualised Category 3 PUE target of 1.15 in its UK and EU sovereign regions, achieved through warm-water direct-to-chip liquid cooling, dry-cooler heat rejection with adiabatic-assist on hot days, EC pumps, lithium-ion UPS in eco-mode, and — where site partners support it — waste-heat reuse into district heating. The published per-region PUE figure is part of the NeoCloud assurance pack alongside Tier-III/IV certification, NCSC Cloud Security Principles alignment, and ISO 27001 / SOC 2 / ISO 14001 evidence.

Omniscient Compute exposes the rolling 30-day Category 3 PUE per region as a filterable attribute. A customer running a long-horizon training job can route to the lowest-PUE region with capacity; an inference workload bound by latency can see the PUE alongside the latency band and decide accordingly. This turns sustainability from a corporate reporting concern into a per-workload routing input — without the customer having to renegotiate contracts or change endpoints.

InferenceBench, Yobitel's public AI provider benchmark, will surface PUE-class sustainability attributes alongside performance and price-performance leaders in upcoming releases, so customers comparing inference providers see efficiency context next to throughput and latency. Yobitel Managed Operations runs the per-site PUE telemetry pipeline (Category 3 sampling, anomaly detection, monthly attestation) for both NeoCloud and customer-operated sites under its remit.

References#

PUE is well-documented in open standards and in the operator community. The references below are the authoritative starting points for any serious procurement or sustainability evaluation.