Power is the part of a local AI build people skip - right up until the board brown-outs mid-inference or the electricity bill lands. Here is the math we use.
A Pi 5 idles around 2.7-3.5 W with a minimal OS. Sustained LLM inference on all four cores pushes it to roughly 7-9 W, more with an NVMe hat or USB SSD. Add an active cooler (0.5-1 W) and any AI accelerator hat and budget headroom on top. The official 27 W (5V/5A) USB-C PSU exists for a reason: undervoltage during a long generation run shows up as silent errors and throttled clocks before it shows up as a reboot.
At a sustained 8 W average, 24/7 operation is about 5.8 kWh per month. At typical US or EU residential rates that is roughly $1-2 per month - cheap, but not zero, and it triples if you add a second board or an accelerator. The number that matters more is cost per query: divide your monthly power + hardware amortization by your monthly query count. For most hobby setups the honest answer is that local beats cloud only above a certain usage level - knowing yours is the whole game.
We built two tools for exactly this. The Local AI Power Budget Planner (€4) walks PSU sizing, cooling draw, and runtime budgets in one sheet. The Total Cost Calculator (€4) does the hardware-plus-power-plus-time math and gives you cost per month and per query against cloud API prices. Both are instant downloads.