No doubt. Especially for GPU inference at scale. We overclock/overvolt for training and tune way down for inference.

Self-racking lets you rack a bunch of gear you'd never find in VM/dedicated rentals, like consumer parts or older, still very good parts. Overclocking options are available as well if you DIY.

If you need single-threaded performance, colo is really the only way to go anyway.

We have two full racks and we're super happy with them.

> What am I missing?

VRAM capacity given the Cerebras/Groq architecture compared to Nvidia.

In parallel, RAM contracts that Nvidia has negotiated well into the future that other manufacturers have been unable to secure.

Interesting - I'll try some other methods

I generally agree with you. But I put up with it since Notepad++ is good software. It is what it is.