AI datacenter power and cooling
What an AI datacenter's power and cooling add up to: PUE, WUE, CUE and ERE over a year; I²R busbar loss at 12, 48 and 800 V; rack → Scalable Unit → hall in MW; the cooling ladder and die heat flux; coolant flow from Q = ṁ·cp·ΔT with pump pressure ∝ Q²; central-UPS vs 48 V power trains; N+1 and 2N backup; and what a better PUE saves.
Added
How to use it
Eight tabs, each a first-principles calculator for one decision in an AI facility. Every input is editable and every result is an estimate; the hardware figures it quotes are in one table on the Cooling & heat flux tab, each with its source.
-
PUE · WUE · CUE · ERE — an IT load in MW and its utilisation become a year of IT and facility energy (× PUE), litres of water (WUE × IT kWh — PUE does not enter, because WUE is defined per IT kWh), tonnes of CO₂e (facility kWh × grid intensity), CUE = PUE × grid intensity, and ERE = (facility − reused heat) ÷ IT, which falls below 1 when a site exports more heat than its overhead.
-
I²R & voltage — the same rack power delivered at several voltages down the same copper: current I = P/V, loss I²R, voltage drop. 12 → 48 V is a quarter of the current and 1/16 of the loss.
-
Rack → hall — GPUs and kW per rack, racks per Scalable Unit, SUs per hall, an allowance for networking and storage racks, and PUE: GPUs, IT MW, MW at the meter, and MWh a year.
-
Cooling & heat flux — where a rack’s kW sits on the cooling ladder (air, direct-to-chip liquid, immersion), and a chip’s average heat flux, power ÷ die area, against the ≈50 W/cm² air limit.
-
Coolant flow — Q = ṁ·cp·ΔT for water or PG25 gives the flow a heat load needs; the loop’s pressure drop scales as Q² from a reference point, and the pump power follows. The Lenovo GB300 NVL72 flows at 25–45 °C supply water are plotted.
-
Power train — utility to die, stage by stage, for a legacy central double-conversion UPS, a hyperscale 48 V rack bus with in-rack batteries, and 800 V DC; which losses PUE sees and which it cannot (they sit inside the IT load).
-
Backup — N, N+1, N+2, 2N and 2N+1 for a load and module size: modules, installed capacity, loading, and an idealised availability; battery energy to bridge a generator start; Uptime’s Tiers I–IV.
-
PUE compared — two PUEs for the same IT load: energy saved (1 − PUEa ÷ PUEb), the extra the worse site uses, the ratio of overheads, dollars and carbon.
Related: thermal design (junction temperature, heatsinks, airflow), fluid flow and pipe pressure drop (sizing the coolant pipes), circuit calculators (Ohm’s law and power), solar ROI, AI hardware calculator (what the GPUs do with the power), and compute in orbit (the same heat with no air or water).
Companion to power delivery: why AI racks went to 48 volts and the physical hardware hierarchy (Part X), power and power generation, the cooling wall and beyond PUE: WUE, CUE and ERE (Part XIII), datacenter design (Part XIV, including direct-to-chip liquid cooling) and water for cooling of _AI Acceleration: From Transistors to Trillion-Parameter Models.
Limits
-
Annual figures assume a constant average load; real sites vary hour by hour and by season (WUE especially). Grid intensity is an average, not a marginal or hourly-matched figure.
-
The busbar model is a uniform copper conductor at 20 °C, out and back; real busbars run warmer (higher resistance) and connectors add resistance of their own.
-
The cooling-ladder ranges are guidance from the book and vendor material, not hard limits. Heat flux is averaged over the die; hotspots are higher.
-
Coolant properties are fixed near 30 °C. ΔP ∝ Q² assumes turbulent flow through a fixed loop.
-
Power-train stage efficiencies are typical full-load values; they fall at light load.
-
Availability treats modules as failing independently and ignores distribution paths and common-mode failures — which is what Tier III and IV design is really about.
Sources
-
The Green Grid, PUE, WUE (WP#35) and CUE (WP#32); ISO/IEC 30134 (PUE, ERE, WUE, CUE).
-
Uptime Institute, Tier Classification System; Global Data Center Survey 2025 (industry PUE ≈1.54).
-
NVIDIA, GB200 NVL72; DGX SuperPOD reference architecture (Scalable Units); H100, A100, DGX B200 datasheets and architecture whitepapers (full list in the tool’s hardware table).
-
Open Compute Project, Open Rack v3 (48 V busbar, power shelf, BBU).
-
Pantheon, GB200 and GB300 NVL72 power and cooling requirements (Lenovo GB300 flows and pressure drops).
-
CoolIT Systems, CHx1000 CDU (1.5 L/min per kW).
-
Meta, Prineville (PUE 1.07 at launch); Google and Microsoft environmental reports (PUE, WUE).
-
Coolant properties: ASHRAE Handbook — Fundamentals (water); Dynalene PG-25 data (propylene glycol 25 %).
Linking to an example
Links can be written by hand. Add the tab after #: metrics, busbar, hall, cooling,
flow, train, backup or pue. Opening a link replaces this tool’s saved settings; any
setting the link does not give returns to its default. All values are plain numbers (1e3 works;
if you write a `, encode it as `%2B`, since ` in a URL means a space).
| Parameter | Meaning | Example |
|---|---|---|
|
IT load in MW (the metrics, backup and PUE-compared tabs share it) |
|
|
Average utilisation of that load, % |
|
|
Facility PUE (metrics and hall tabs) |
|
|
Water usage effectiveness, litres per IT kWh |
|
|
Grid carbon intensity, kg CO₂e per kWh |
|
|
Heat exported for reuse, MWh per year |
|
|
Rack power for the I²R tab, kW |
|
|
Voltages to compare, comma-separated; the first is the baseline |
|
|
Busbar length one way, m |
|
|
Copper cross-section, mm² |
|
|
GPUs per rack; kW per rack (also the cooling ladder’s rack) |
|
|
Racks per Scalable Unit; number of SUs; networking and storage overhead % |
|
|
Chip for the heat-flux calculation: |
|
|
Custom chip power (W) and die area (mm²) |
|
|
Heat load (kW) and coolant temperature rise (K) |
|
|
|
|
|
Pump reference flow (L/min) and pressure drop (psi), pump efficiency % |
|
|
Power-train tab: load at the chips, MW |
|
|
Stage efficiencies, %: central UPS, server PSU, 48 V power shelf, VRM |
|
|
Backup: module size (MW); |
|
|
PUE compared: the better and worse PUE, electricity $/MWh |
|
|
One of the presets below; parameters given with it override the preset’s |
|
Presets (the part number is the book’s):
example= |
Shows |
|---|---|
|
500 MW IT at PUE 1.15 and WUE 1.4 L/kWh: ≈6.1 GL of water a year (WUE is per IT kWh, so PUE does not enter) (open) |
|
A 100 MW site at Microsoft’s hot-dry Arizona WUE of 1.63 L/kWh (open) |
|
The same site at Microsoft’s Danish WUE of 0.01 L/kWh (open) |
|
An evaporative site at the ≈1.8 L/kWh industry average (open) |
|
CUE = PUE × grid intensity: a fossil-heavy grid (≈0.45 kg/kWh) against a near-zero one — change ci to compare (open) |
|
ERE below 1: a 50 MW site exporting ≈100,000 MWh/yr of heat, as Meta’s Odense campus does (site size illustrative) (open) |
|
A 120 kW rack at 12 V vs 48 V on the same copper: a quarter of the current, 1/16 of the loss (open) |
|
An Open Rack v3 busbar’s ≈18 kW at 12 V and 48 V (open) |
|
A 600 kW Kyber-class rack at 48 V vs 800 V DC (open) |
|
One GB200 NVL72 rack: 72 GPUs at ≈132 kW (open) |
|
NVIDIA’s Scalable Unit: 8 NVL72 racks, 576 GPUs, ≈1.06 MW of compute racks (open) |
|
A 16-SU hall: 128 racks, 9,216 GPUs, ≈17 MW of compute racks (open) |
|
H100: 700 W over 814 mm² = 86 W/cm², past the ≈50 W/cm² air limit; a 12 kW enterprise rack (open) |
|
B200 at 1,000 W over two dies (area estimated) (open) |
|
A 132 kW GB200 NVL72 rack on the cooling ladder: direct-to-chip liquid (open) |
|
A projected 600 kW Rubin Ultra Kyber rack: far past air (open) |
|
Meta/Rittal High Power Rack at 90 kW (open) |
|
120 kW of PG25 at ΔT 10 K: ≈180 L/min (≈225 L/min at 8 K) (open) |
|
Lenovo GB300 NVL72 flows 59–177 L/min from 25 to 45 °C supply, ΔP rising as Q² (open) |
|
A 1 MW CDU at 1.5 L/min per kW (CoolIT CHx1000): ΔT ≈ 10 K (open) |
|
Central double-conversion UPS vs 48 V rack bus vs 800 V DC, utility to die (open) |
|
The same with a 98 % (eco-mode) UPS: most of the gap closes (open) |
|
A 20 MW load on 2.5 MW modules: N, N+1, 2N and 2N+1 compared (open) |
|
Battery to bridge a 30 s generator start for a 132 kW rack (open) |
|
Prineville PUE 1.07 vs 1.5: 29 % less energy (the old site uses 40 % more) (open) |
|
Google’s 1.09 fleet vs the Uptime ≈1.56 industry average (open) |
|
Hyperscaler ≈1.10 vs industry ≈1.5 (open) |
Feedback is a gift. Something to improve, a feature to add, or a bug? Tell me.
Runs entirely in your browser; nothing you type leaves this page.