AI's Compute Crunch Has a Power Problem
By Sustainability Awakening • Dr. Ahmad Mahmood
Sold out
AWS: booked into 2028
AWS capacity is sold out for 2026, largely for 2027 and ‘a good chunk’ of 2028, Andy Jassy said on 23 Oct.
Source: Reuters
Rental prices
GPUs cost more to rent
+21%
Average cloud GPU rental price change over 12 months, 21 models.
Source: GetDeploying tracker, Oct 2026
Power per chip
Each GPU is a space heater
Nvidia H100
700 W
Nvidia B200
~1,000 W
Source: Nvidia specifications
Do the maths
10,000 B200s
Facility load
~18 MW
Energy a year
~160 GWh
Source: our estimate, PUE 1.3
Carbon
~56,000 t CO2 a year
That cluster on the average US grid emits as much as about 12,000 cars.
Source: our estimate; EPA eGRID
Water
~15M gallons on site
Evaporative cooling at average US rates, before water used by power plants.
Source: our estimate; LBNL
The bottleneck
Power, land, permits
Builds wait on grid connections. US data centers could use 6.7–12% of power by 2028.
Source: LBNL 2024
The risk
Stranded and idle chips
Chips can arrive before power does, and idle GPUs still carry their embodied carbon.
What helps
Use every watt
Liquid cooling
PUE <1.2
Clean PPAs
Lower Scope 2
High utilisation
Less waste
Your next step
Model an AI campus
Read the full analysis →
AI Data Center Energy Modeler
Read the full AI compute analysis