>I doubt we'll see the big innovation from a garage now, because the cost of entry has become too high for shoestring operators to succeed. > ML training needs fast, expensive hardware. You can experiment with it on a 1080 graphics card, but for industrial applications you need a lot more speed and power.

This isn't the 1990's. AWS exists. Not to mention how hot VC is for AI startups.

Sure, we won't have 15-year olds making the next AI breakthrough but to say that AI is solely in the realm of "big business" is naive.