> the most valuable resource is rack space

The limit is power capacity and quite often thermal. Newer DCs might be designed with larger thermal envelopes, however rack space is nearly meaningless once you exhaust thermal capacity of the rack/isle.

Performance within thermal envelope is a very important consideration in datacenters. If a new server offers double performance at double power it is a viable upgrade path only for DCs that have that power reserve in the first place.

Well, Amazon does offer Graviton 4 (quite fast and useful stuff) along side their Epyc machines so there is some utility to them. A 9654 is much faster than a Graviton 4.

EDIT: Haha, I was going off our workloads but hilariously there are some HPC-like workloads where benchmarks show the Graviton 4 smoking a 9654 https://www.phoronix.com/review/graviton4-96-core/4

I suppose ours must have been more like the rest of the benchmarks (which show the 9654 faster than the Epyc).

Cooling takes up rack space, too. There also are workloads that aren’t CPU constrained, but GPU or I/O constrained. On such systems, it’s better to spend your heat budget on other things than CPUs.

> the most valuable resource is rack space

I've always heard it's cooling capacity. I'm also pretty confident that's true

AWS, Google, Hetzner all offer a discount if you use an ARM64 VPS.

Clearly, they want them, because there's demonstrated power savings.

Rack space limits include power limits. E.g. 10kw per rack.