Deep learning is too expensive for most developers

It isn't that it is too expensive, it's the the value isn't there.

Of course it is expensive to run Hinton's digits over and over again because there is no value in doing so.

Frankly I wonder if it would work if you ran anything other than the NIST digits because (1) nobody has tried, and (2) getting training sets is expensive... MUCH more expensive than running the model.