Frontier models cost the most to train and require the largest compute clusters, so only a few labs can build them. The race to lead the frontier drives much of the industry's chip and power demand.