Nvidia can't make GPUs fast enough. I doubt 10xing training and/or inference efficiency would result in a decrease in demand. I would be surprised if it didn't instead increase demand. Mind you, Nvidia is pushing hard on TensorRT which optimizes models at inference time and results in major increases in throughput (not 10x though lol).