The power and cooling shift is the real constraint, not the headline GPU count. A useful capacity claim should expose rack density, liquid-cooling support, PUE under the target load, water or heat-rejection limits, and the interconnect path to storage. Otherwise an AI cluster can be technically installed but operationally throttled.

For teams choosing a region, I would also ask for a dated capacity proof and a failure plan: what happens when the promised rack slot, power tranche or network path slips? Those details turn a data-center announcement into an engineering decision.