Energy layer

Datacenter Power Calculator

Estimate MW capacity required for an AI training or inference cluster of a given size.

The engineer question
How much power does an AI cluster of N GPUs need?

Where the data lives today

The current quarter snapshot is generated by the V17 Phase D Product Landscape module and lives on the parent topic page. The interactive comparator below ships once a structured spec datasource is wired.

View Energy product landscape →

Inputs

Total accelerators in the cluster (training + inference combined).

Hyperscale 2026: 1.10–1.25 (liquid). Air-cooled colo: 1.4–1.6.

Result

IT load (accelerators only)
7,000 kW
IT load (incl. server overhead)+30% for CPU / memory / NIC / storage
9,100 kW
Facility load (IT × PUE 1.30)
11.83 MW
Sized capacity (N+1)What the substation contract has to deliver
13.01 MW
Annual energy
103.6 GWh
Annual electricity cost@ $0.08/kWh
$8.3 M

Recommendation

13 MW is a mid-size colo fit. Most Tier-3 colocation providers can deliver this in 6–12 months on existing campuses.

Assumptions

  • · NVIDIA H100 SXM5: 700 W TDP per accelerator (sustained training draw can be 5–15% higher than nameplate)
  • · Server overhead factor: 1.3× (CPU + memory + storage + NIC, typical for an 8-GPU server)
  • · PUE: 1.30 (1.10–1.25 for liquid-cooled hyperscale 2026; 1.4–1.6 for air-cooled colo)
  • · Electricity cost: $0.08/kWh, US industrial blended average; varies $0.04–$0.18 by state
  • · Redundancy factor (N+1): 1.1× applied to nameplate capacity, not utilisation
  • · Excludes: future training intensity ramps, idle / between-job consumption (subtract ~20%), spot pricing arbitrage

Worked example (default inputs)

Result

IT load (accelerators only)
7,000 kW
IT load (incl. server overhead)+30% for CPU / memory / NIC / storage
9,100 kW
Facility load (IT × PUE 1.30)
11.83 MW
Sized capacity (N+1)What the substation contract has to deliver
13.01 MW
Annual energy
103.6 GWh
Annual electricity cost@ $0.08/kWh
$8.3 M

Recommendation

13 MW is a mid-size colo fit. Most Tier-3 colocation providers can deliver this in 6–12 months on existing campuses.

Assumptions

  • · NVIDIA H100 SXM5: 700 W TDP per accelerator (sustained training draw can be 5–15% higher than nameplate)
  • · Server overhead factor: 1.3× (CPU + memory + storage + NIC, typical for an 8-GPU server)
  • · PUE: 1.30 (1.10–1.25 for liquid-cooled hyperscale 2026; 1.4–1.6 for air-cooled colo)
  • · Electricity cost: $0.08/kWh, US industrial blended average; varies $0.04–$0.18 by state
  • · Redundancy factor (N+1): 1.1× applied to nameplate capacity, not utilisation
  • · Excludes: future training intensity ramps, idle / between-job consumption (subtract ~20%), spot pricing arbitrage

Related tools in the Energy layer

Get notified when Datacenter Power Calculator numbers update

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

→ Get this data as JSON