The first comment on the article (not Devac, the actual article page) was where my mind went, except I'm not gonna use GPUs as a heatsink, I'm going to use the GPUs to do computations. But of course, that'd be institutionalized, and research might get held up until energy grid conditions were favorable. Not ideal.
Let's say you pay $.01 per CPU per hour for my Beowulf cluster if scheduled but $.005 per CPU per hour opportunistically. Your protein folding analysis is gonna take 2500 CPUs about 30 hours to run your calcs. It's gonna take you a week to get scheduled and $750 and more than a day once you're on deck. Or, It's gonna take a week to get scheduled and maybe another week (and maybe three days to run) if we're doing it when the power is cheap. You're looking at a couple weeks anyway; do you want it at the beginning of the week for $750 or the end of the week (ish) for $375? Every musician of the '70s, '80s and '90s knows the value of "cheap studio time." You can get your goofy garage band in between midnight and 4am because Fleetwood Mac has booked the whole studio for a month but they usually knock off at 4 to go blow coke up Stevie Nicks' ass so they'll never even know we were there. The cluster equivalent of "off peak pricing" will find plenty of use.