“By 2027, AI training clusters are expected to be approximately 3 orders of magnitude (OOMs) larger than today, reaching over 10 million A100-equivalent GPUs.”