Compute is scarce and expensive enough that access to it functions as a competitive moat: labs sign multi-year, multi-billion-dollar deals for GPU or TPU capacity, and 'compute-constrained' is a common explanation for why a lab trains fewer or smaller models than its research roadmap would otherwise support. Regulatory frameworks like the EU AI Act use compute thresholds (measured in total training FLOPs) as one trigger for extra obligations on 'systemic risk' models, treating raw compute spent on training as a rough proxy for capability and risk.