Every colocation gives you a different power budget per rack.
So you don’t fill a rack with servers. You fill it with watts. Each rack becomes a bin-packing problem: which mix of machines fits under the power ceiling. Often the answer leaves slots empty, which means space you’re paying for and can’t use.
Now do that across dozens of racks, where a single GPU node can draw as much as ten CPU servers. Then sequence it across three parties who are each waiting on the other two: lenders, server manufacturers, and colo providers.
Getting the servers financed, powered, and racked at the same time is quite a job.
もっと見る
