That's 50k if you already know the exact architecture you're going to use, and the exact training data you're going to use, and nail it all on the first try. Mu

That's 50k if you already know the exact architecture you're going to use, and the exact training data you're going to use, and nail it all on the first try. Multiply that by however many failed attempts it takes while experimenting with new ideas, especially when the corporate players stop being so eager to publish the finer details of their research, and that 50k could very quickly turn into millions.