What is the utilization of a hyperscaler datacenter?
Posted by banana_zeppelin@reddit | sysadmin | View on Reddit | 9 comments
I was wondering. Does anybody know what the general utilization of a hyperscaler datacenter is? How many headroom does it have to absorb peaks in usage before offloading requests to other datacenters? 50%? 75%? 90%? What are the economics here?
Grand-Travel1665@reddit
from what I've seen, they don’t run them super hot like 90–100% all the time
it’s usually more like 60–80% ish, but it varies a lot. the key thing is they keep buffer, but that “unused” capacity isn’t really idle — they fill it with stuff that can be killed anytime (batch jobs, spot, internal workloads)
also they don’t rely on a single DC to absorb spikes. they just shift load across regions pretty aggressively
so yeah it’s less this DC has 30% free and more a mix of headroom + traffic shifting + killing low priority stuff when things spike
Massive-Reach-1606@reddit
Thats the cover story. This is all about skynet.
banana_zeppelin@reddit (OP)
gimme your compute
SevaraB@reddit
It’s not a single answer. We agree on burst limits so we can predict what scale out operations will look like and then we build that cushion into our data centers, but for a general baseline, we tend to start with an 80/20 built/available rule of thumb. But either way, hyperscaler doesn’t mean “infinite capacity” or that we’ll allow runaway scaling ops.
banana_zeppelin@reddit (OP)
Thanks, sounds reasonable
Jaki_Shell@reddit
69% give or take.
banana_zeppelin@reddit (OP)
Nice
iDestinaTE@reddit
Who gives and who takes more?
Annh1234@reddit
Depends on who's using it. If you look at some big AI data center papers your can find that they need to add battery buffers for the power use, since the workload goes up and down enough to cause an issue with the building power source.
So from that you can infer that it goes from super low to super high % load