Hacker News Reader

They can still divide them up for the smaller sizes preemptible and destroy those once they get a request in. That way utilisation should be quite good. Yield probably not as high (since hardware costs will be higher than for multiple small machines) but the yield of renting out the highest variant sometimes will be worth it.

Yes, I configured a 64 cpu machine in a matter of minutes, ran a test on it for 24 hours, then shut it down and deleted the instance. Total cost was around $200.

*vCPU.

But yes, they probably strive to keep below 10% vacancy on all hardware.

GCP has live migration. Maybe they make space if it doesn't exist.

Hard to make hardware appear out of nowhere.

They only use live migration to do host maintenance iirc.

There are lots of reasons to shift stuff around, and there’s lots of hardware :) -googler

Idle, or running their own internal computations, or running lower preemptiable stuff.

Preemptible VMs exist in part to avoid leaving capacity sitting idle.

We hate idle resources. :)