At current prices, the VM + GPU is $0.57/hr, which doesn’t break Google’s bank and is absurdly cheap for usability testing for a new product.
And the preemtible price is only $0.22 per GPU. Not that the GPUs used here seem to be preemtible, but this might give some hint to the baseline cost.
It's previous gen hardware, so presumably the only cost to google is energy usage and similar. They probably use the newer p100 and v100 cards internally which are significantly faster for deep learning.
Not only that, the hardware is already on, so it's only a matter of the cost difference between an idle GPU and a busy one. It's not inconceivable that if enough paying customers showed up with K80 workloads, these freebie users might have to wait longer.
Presumably, there's not an unlimited number of them.