Probably cerca 2014 as that's when AlexNet was released, demonstrating that neural networks could beat traditional ML models at image recognition tasks. I recall the researchers used Cuda to optimize their training setup.

AlexNet kicked off a new wave of research around neural networks by demonstrating they could be scaled well and trained on GPUs.

Yeah - definitely by 2014 they were well entrenched within academia with their CUDA offerings. By that point it wasn't "luck".