If you have a problem that requires and hour or a day to compute, then spending fifteen minutes for data transfer up and down (particularly in the face of lowered costs) is often a profitable trade. Movie studio render farms are a classic example of such compute jobs. Weather or geological resource prediction could be another. There are many such high-compute jobs in practice.
>>hollas+(OP)
Not only are those jobs relatively rare, but they also require a lot of hardware and a lot of power. Both of those are going to be in short supply. (Solar power requires sunlight, and the target locations are deliberately in shadow.)