How much power does a local AI node use?
A grounded look at idle draw, inference bursts and sustained load across compact, workstation-class and rack-mounted local AI hardware.
Power figures for AI hardware are usually quoted as peak board power, which tells you almost nothing about what a node will actually consume over a week of real use.
Three numbers matter more. Idle draw sets your floor and, for an always-on node, often dominates annual consumption. Burst draw sets the requirement your supply path and protection must tolerate. Sustained load under a realistic workload mix tells you what your energy plan needs to cover.
Compact nodes typically sit in the tens of watts at idle and low hundreds under load. Workstation-class systems with a discrete accelerator can idle in the low hundreds of watts and peak well beyond that. Multi-accelerator systems belong in a properly assessed space with attention to ventilation, noise and circuit capacity.
Measure at the socket, not from software counters alone. Software reporting misses power supply losses, drives, networking and cooling.
Once you have real figures, feed them into the energy-to-compute planner rather than relying on the datasheet number.
- hardware
- power
- measurement