• 3D graphics - hence GPUs.
• Fluid dynamics simulations (weather, aerodynamics, nuclear, injection molding - what supercomputers do all day.)
• Crypto key testers - from the WWII Bombe to Bitcoin miners
• Machine learning inner loops
That list grows very slowly. Everything on that list was on it by 1970, if you include the original hardware perceptron.
 Its not even a particularly good reason. The only reason we don't have massively parallel general purpose CPUs is because of how specific the problems that can benefit from it are. Even then, modern GPUs are pretty close to being general purpose parralell processors.
I was also thinking about including the work that hard-drive controllers do (like checksumming, and handling 512/4K logical/physical sectors), but the difference there is that, for NIC offload, the kernel has that functionality already, where for hard drives the kernel does not do the work of the hard-drive controller.
Once everyone realizes every practical use of their AI technology is more than adequately met by conventional code, and that there's no grand breakthrough into general artificial intelligence coming anytime soon, the hype cycle will end.
I'm all for skepticism for the current hyperbole, but there's no need to be hyperbolic in the other direction.
There are some applications in which deep learning really does work better than alternatives. The 2018 Gordon bell prize, after all, went to a team that did deep learning on Summit for climate analysis.
There is a nontrivial list of applications that you would have a hard time convincing experts that they would be better off with conventional code.
A better question to ask back then would have been "which large business will depend on the internet in 2015".
Note that the internet in 1995 was pretty bad (in terms of applications and from technical point of view), and the hype led to dot com crash 5 years later. Yet it's hard to overestimate the importance of the internet in today's world.
So if that technology were to catch on, a pedant could argue that most computing workloads really are machine learning.
3D graphics? Check (freebie).
Fluid dynamics? Check - supercomputers increasingly get most of their compute from GPUs.
Cryptography? Check - this is the only one that really got specialized hardware.
Machine learning? Check.
So "large quantities of special purpose hardware" wasn't even used for these. Just large quantities of general purpose parallel processors, known for historical reasons as "graphics processing units."