Why? Unlike loads involving a real physical process there is absolutely no need for AI-training to be constant.

everfrustrated an hour ago | parent | next [-]

You are correct in the sense that they can stop work in a way many generic server use cases can't (which is seen in lowering power supply reliability requirements as the article mentions), but running expensive servers at 50% utilization would dramatically affect the revenue generated per capital invested - IE you couldn't afford to buy the servers.

	▲	ZeroGravitas 8 minutes ago \| parent [-]
		If it just costs more, similar to how it's possible to have a low water use datacenter, then in a perfect spherical cow economic universe then the lower cost option is better. In the real world you'd need to add some cost to account for the externalities on water and GHG. Do the numbers still work out for the gas powered plant? If not then you're just exploiting unaccounted for externalities.

▲

pu_pe an hour ago | parent | prev [-]

If you invest into chips that deprecate in value really fast, not utilizing them to their full capacity because of power constraints would be counter productive.