Datacenter PUE Cooling Efficiency Simulator — hyperscale
A tool that lets you pseudo-design hyperscale datacenters of the PUE 1.1 class achieved by Google and Meta, along seven axes: IT load, cooling type, climate, inlet temperature and PDU efficiency. Switching from air to immersion cooling moves annual energy, CO₂ and water use in real time, so you can intuitively check how the choice of site and cooling technology drives operating cost.
Parameters
DC type
Default preset (other parameters take precedence in the calculation)
IT load
MW
Cooling type
Dominant factor for PUE. Liquid / immersion deliver large gains.
Outdoor temp T_out
°C
Server inlet temp T_inlet
°C
ASHRAE TC 9.9 recommends 18 to 27 °C. Warm-aisle: 27 to 35 °C.
DC cross-section view — racks, cooling tower, air / liquid loops
Visualizes in real time the flow from rack heat to cooling equipment to outdoor rejection, plus the PUE bar. Colors indicate the cooling type (blue = liquid, orange = air).
PUE sensitivity — IT-load scale vs total PUE
PUE comparison across cooling types (same conditions)
The theoretical PUE floor is 1.0; Google's 12-month average is 1.10. Cooling and distribution losses push it higher, and the hyperscale standard is to drive it down to 1.1–1.2 with liquid cooling, evaporative cooling and temperate siting.
The approximation used by this tool. PUE_cool is the baseline from the cooling type, Δ_climate is the climate-zone correction, a higher inlet temperature contributes negatively (less cooling demand), and lower PDU efficiency contributes positively.
Datacenter PUE Cooling Efficiency — Google/Meta hyperscale
🙋
PUE is basically the "fuel economy" of a datacenter, right? I've heard 1.0 is the theoretical limit — how impressive is Google's 1.10 actually?
🎓
Nice analogy. PUE = total datacenter power ÷ IT-equipment power, and 1.0 would be the ideal where literally no electricity goes to anything other than the servers. Google is at 1.10 on a 12-month average, and the Saint-Ghislain site in Belgium has reached 1.07. A typical Japanese enterprise DC sits at 1.5 to 1.8, so for the same servers Google needs 30 to 60 % less total power. For a 10 MW IT load that's 10 to 15 GWh per year — on the order of hundreds of millions of yen in electricity bills.
🙋
Wait, hundreds of millions of yen per year?! How does that gap open up — the servers themselves are the same, right?
🎓
The biggest factors are cooling and siting. Enterprises typically run "server room + large chiller (circulating 4 °C chilled water)", which alone adds 0.4 to 0.5 to PUE. Google, on the other hand, uses evaporative cooling that draws on outdoor air directly, free cooling in cold climates, and liquid cooling for AI racks. Try switching the "Cooling type" on the left from air-cooled to direct liquid or immersion — PUE drops from 1.5 to about 1.1 right away.
🙋
You're right — immersion brought it down to 1.03. That's the technique where servers are submerged in liquid, isn't it? Is it actually reliable?
🎓
It uses single-phase dielectric oil or fluids like 3M Fluorinert and fully submerges the hardware. Microsoft, Meta, and in Japan Sakura Internet, are already running it in production. Once you get to NVIDIA H100 at 700 W per card and 30 to 50 kW per AI rack, air cooling simply cannot keep up physically. Project Natick (Microsoft's submerged DC) takes it further — "cool with seawater 24/7" — and combines stable low temperatures with natural convection to hit PUE 1.07. Siting and cooling are the two wheels of the same cart.
🙋
It's interesting that raising the server inlet temperature also saves energy. PUE dropped when I went from 27 °C to 32 °C. Why is that?
🎓
Raising the inlet temperature by 1 °C lets you also raise the chiller setpoint, so the period where the indoor-outdoor delta is small grows and free cooling kicks in more often. ASHRAE TC 9.9 specifies 18 to 27 °C recommended and 15 to 32 °C allowable, and hyperscalers commonly run right at the edge of the allowable band at 27 to 32 °C. Continuous high-temperature operation does shorten the life of HDDs, SSDs and electrolytic capacitors though, so it's a trade-off against equipment MTBF. Recently, AI has been controlling the inlet temperature dynamically — Google famously cut cooling power consumption by 40 % in 2014 when they deployed DeepMind.
🙋
Water consumption (WUE) is shown too. When I switch to evaporative cooling it uses water on the order of 100 ML. Is that a problem?
🎓
A big one. A single site can consume 1 to 5 million m³ per year (200 Olympic swimming pools), and in Arizona and other arid parts of the western U.S., water-rights disputes with local municipalities are real. That's why recent hyperscale capacity is concentrating in the Nordics — "cheap electricity and abundant water" places like Verne Global in Iceland or Lefdal Mine in Norway. Closed-loop liquid cooling, on the other hand, has nearly zero WUE. We're entering an era where PUE, WUE and CUE (carbon) all have to be optimized in parallel.
Frequently asked questions
PUE (Power Usage Effectiveness) = total datacenter power / IT-equipment power. The metric was proposed by The Green Grid in 2007. The theoretical minimum is 1.0 (zero cooling and distribution losses). Among modern hyperscalers, Google reports a 12-month average of 1.10, with its Saint-Ghislain site in Belgium at 1.07. Meta is 1.09, AWS roughly 1.20, while enterprise on-prem datacenters average 1.5 to 1.8. This tool estimates PUE from cooling type, climate, inlet temperature and PDU efficiency to support early-stage design sizing.
In this tool's model, the cooling PUE contribution is about 1.50 for legacy chilled-air, 1.15 for evaporative, 1.20 for adiabatic, 1.10 for direct-to-chip cold-plate liquid, and 1.03 for immersion. NVIDIA H100-based AI racks are 30 to 50 kW per rack — too dense to be cooled by air, so liquid or immersion cooling is essentially mandatory. However liquid cooling carries high CAPEX and maintenance overhead, so phased adoption is realistic for retrofits of existing datacenters.
Yes. ASHRAE TC 9.9 defines a recommended range of 18 to 27 °C and an allowable range of 15 to 32 °C, and modern hyperscalers commonly operate in a 27 to 35 °C warm-aisle regime. Raising the inlet temperature by 1 °C improves cooling-equipment COP, reduces chiller runtime, and tends to lower PUE by roughly 0.005 to 0.01. The expression inletTempAdj = (27 − T_inlet) × 0.005 used in this tool is a linear approximation of that empirical trend. There is, of course, a trade-off with hardware lifetime and failure rate, so operating above 35 °C is not recommended.
WUE = annual cooling water / IT-equipment energy (L/kWh) and is a sustainability metric on par with PUE. Evaporative cooling has a relatively high WUE of 1.0 to 1.5 L/kWh and conflicts with water resources in arid regions such as the U.S. Southwest. Air-cooled chillers and closed-loop liquid systems have WUE near 0 but pay for it in extra electricity (higher PUE). This tool approximates WUE as evaporative = 1.2, adiabatic = 0.5, others = 0, and the optimum is chosen jointly from site climate, water cost and electricity source (renewable / fossil).
Real-world applications
Hyperscale cloud (Google / Meta / Microsoft / AWS): Each company's annual sustainability report publishes fleet-wide PUE and WUE. The most recent disclosed figures are Google 1.10 (12-month average), Meta 1.09, Microsoft 1.18 and AWS 1.20. New-datacenter siting decisions hinge on three factors — climate zone (number of free-cooling days), renewable share of the local grid, and water cost — and the climate-zone, cooling-type and WUE parameters of this tool form a miniature version of that decision.
AI / HPC dedicated datacenters: Stacking 4–8 DGX servers per rack, each with eight NVIDIA H100 (700 W) or B200 (1 kW) cards, easily yields 30 to 50 kW per rack, and AI training pods at 100 kW/rack are not unusual. That exceeds the physical limit of air-cooled chillers (~25 kW/rack), making cold-plate liquid or immersion cooling mandatory. The 1.03 to 1.10 PUE you see in this tool for liquid or immersion reflects that reality.
Enterprise on-prem and colocation: Captive datacenters for finance and manufacturing still typically run at PUE 1.5 to 1.8, with CRAC (Computer Room Air Conditioner)-style cooling. Colocation operators (Equinix, NTT, KDDI) aim for 1.4 to 1.6 by adopting hot-aisle containment and outdoor-air intake to separate hot and cold airflows. Setting this tool to evaporative + temperate + 97 % PDU brings it down to 1.17, useful for sketching out a retrofit roadmap.
Research and university supercomputers: Fugaku in Kobe is reported at PUE 1.13 with a water + air hybrid, and NVIDIA Selene at 1.07. In HPC, power efficiency (GFLOPS/W) has overtaken raw precision as the competitive axis, and the Green500 ranking is closely tied to PUE. In HPC mode (90 %+ CPU utilization, long sustained runs), average IT load is also high, and dynamic cooling response is heavily exercised.
Common misconceptions and pitfalls
The biggest misconception is that "low PUE = green" is not necessarily true. PUE is purely "ancillary power relative to IT power" and does not reflect whether the electricity comes from renewables or coal. A PUE of 1.05 in a region with 90 % coal-fired power can emit more CO₂ than a PUE of 1.40 site running on 80 % renewables. That is exactly why Google and Microsoft aim for 24/7 carbon-free energy (CFE), and PUE should ultimately be evaluated jointly with the CO₂ intensity of the electricity source (CUE = Carbon Usage Effectiveness). This tool approximates with a flat 0.4 kg-CO₂/kWh, but in practice you should apply region-specific grid factors.
Next: "evaluate PUE as an annual average, not an instantaneous value." Winter free-cooling yields PUE 1.05 and summer peaks hit 1.35 — seasonal swings are large. Many sites flaunt their best instantaneous figure of 1.05, but ASHRAE and The Green Grid recommend reporting "12-month weighted average". If you sweep outdoor temperature between −10 °C and 40 °C in this tool, PUE moves by more than 0.1 for the same DC. In facility design, you need both the summer peak (for capacity sizing) and the annual average (for operating cost) — a two-stage view.
Finally: "liquid cooling does not always win." Liquid and immersion cooling do beat air on both PUE and cooling efficiency, but (1) capital cost is 2 to 4 times higher than air, (2) leakage risk and maintenance effort, and (3) server vendor warranties are limited. For commodity servers below 10 kW/rack, evaporative cooling plus hot-aisle containment that delivers PUE 1.15 while keeping CAPEX low often produces better ROI. The right approach is to fix the IT-load density (kW/rack) first, then choose the cooling type.
How to Use
Set IT Load (MW) between 5–50 MW to define total computational power consumption in your hyperscale facility.
Input outdoor ambient temperature (°C) typical for your datacenter location; use 15°C for temperate climates, 35°C for hot regions.
Configure server inlet temperature setpoint (18–27°C); lower values improve chip performance but increase cooling energy; Google's 1.1 PUE operates at ~24°C inlet.
Adjust power distribution efficiency (92–98%); account for UPS, PDU, and transformer losses; 95% is typical for modern hyperscale with 48V DC distribution.
Run simulation to view Total PUE, cooling load (kW), annual energy consumption (GWh/y), and water usage (ML/y).
Worked Example
A 30 MW hyperscale facility in Dublin with 15°C ambient, 24°C server inlet, 96% distribution efficiency: IT load = 30,000 kW. Cooling load ≈ 8,500 kW (with indirect free-air cooling active). Total facility power = 31,250 kW (30,000 ÷ 0.96). PUE = 31,250 ÷ 30,000 = 1.042. Annual energy = 274 GWh/y. With Irish grid carbon intensity 350 g CO₂/kWh, annual emissions ≈ 96 kt CO₂. Water evaporation loss (adiabatic coolers) ≈ 185 ML/y at full load, seasonal average ≈ 110 ML/y.
Practical Notes
Free-air and evaporative cooling activate when outdoor temperature drops below server inlet setpoint; disable them manually to see impact of mechanical chiller-only scenarios on PUE degradation (typically 1.3–1.5 in warm climates without economization).
Hyperscale operators in water-constrained regions (Middle East, Southwest USA) prioritize dry-cooler designs; expect PUE penalty of 0.08–0.15 versus wet cooling.
48V DC distribution and immersion cooling substrates reduce power distribution losses to 2–4%, shifting PUE gains to infrastructure efficiency rather than cooling alone; validate against your facility's actual PDU topology.
Annual energy and CO₂ assume constant 24/7 operation; apply seasonal temperature variation and workload diversity factors (typically 0.75–0.85 utilization) for real-world estimates.