Show HN: Phobos – A tiny scale-free kernel language with tile-DAG support

I have a question.. As all these GPU cores are just used as computational acceleration for a CPU code, how the fuck they claim to have and control thousands of GPUs running one model inside a datacenter? Are they just using too big arrays :) - it makes no sense to me. How many GPUs one CPU code may need? How many CPUs they have and How then these CPUs are running in parallel? Do they at all?????