Racked servers pre-GPU used about 1kW each at high load. 5-10kW for a rack (20 to 40 servers) was common, much more than this would be a special thing.

AI use has driven this to 20, 50, 100kW per rack or more. Supposedly 1000kW for the hyperscalers.

Extrapolate to the whole building and you have a lot of waste heat that has to go somewhere.

Instead of being similar to light industry, the hyperscale AI datacentre is using as much energy as a large steelworks.

I have a small home server setup. It’s stores all my data, and i can transcode video and stuff on it no problem.

I put in a GPU to run some local LLM stuff. Now the whole thing generates enough heat and noise that I notice it if i sit near the cabinet it’s in.

Scale matters.