Each new generation of AI models requires roughly a four-times multiplier in compute resources