The difference is that each additional iPhone or ML accelerator or cloud server CPU provides incrementally more utility. More ML accelerators means more ML processing.
Proof-of-work mining is the only system designed to get less efficient with each additional unit of compute power. Add another GPU to the mining network and the network will automatically become less efficient. More GPUs doesn’t mean more transactions, but the system pays people to essentially burn these GPUs.
That’s the complaint. Not that chips are being used, but that the marginal output of each additional GPU is negative.