A single NVIDIA GB300 NVL72 rack holds 72 GPUs and draws around 135 kW in normal operation, up to 142 kW at peak. Run it for a year and it uses about 1.18 million kWh. A typical UK home uses 2,500 kWh a year, so one rack draws as much as roughly 470 homes. That’s before cooling, which can add another 20–40%.
Compare that with a traditional enterprise server rack, which draws a few kilowatts to around 15 kW. AI racks are ten times denser, which is why they need liquid cooling and why data centres are now designed around power first.
The maths: 135 kW × 8,760 hours = 1,182,600 kWh ÷ 2,500 kWh ≈ 470 homes.
What it means: every GPU deployment is also a power project. The first question on any AI build is where the megawatts come from and when they’ll be energised.