Depends on the use case. Dedicated datacenters for ML training can trade off power reliability vs. other factors like cost or carbon emissions.
Datacenters and especially ML training hardware is highly capital intensive and depreciates at basically constant rate regardless of utilization.
I see currently no scenario where you wwould be willing to idle this expensive infrastructure just to save pennies on the dollar on a grid connection; carbon credits would have to be nonsensically expensive for this to happen.
See e.g. this post from Urs Hölze, one of the fathers of hyperscale computing: https://www.linkedin.com/posts/urs-h%C3%B6lzle_rethinking-lo...
Assuming ~$10M of capex (to buy the datacenter) per MW of electrical power (required by the datacenter), and hardware that is obsolete after 5 years (or even 10!), turning that datacenter off for an hour just to save like ~50$/MWh (or whatever spot price is) seems extremely counterproductive, because your hardware running for that additional hour is worth multiple times that (you spent like >$100 per operating hour on the hardware alone assuming 10 year lifetime).
It seems much more attractive (and credible) to just install more batteries (or even a gas turbine), instead of chasing demand-side-regulation pretensions.
edit: thx for the link though, that is a very interesting study/data even if I disagree with that conclusion!