AI datacenters/Economics

AI datacenters

Cost, power and speed of AI datacenters

Three ratios fall out of the figures recorded for AI datacenters. Where an investment and a power figure are both stated, the median is $0.0 billion per megawatt across 31 campuses. Where IT power and an accelerator count are both recorded, the median is 2.0 kW of IT load per accelerator across 24 systems. And 81 running AI systems have both an announcement and a first-operation date, with a median of 2 months between them. All three are indicative: the numerator and denominator often cover different scopes, and the sample is the campuses that publish numbers.

Investment per megawatt

Read with care. A stated investment can cover land, power and several phases, while the power figure may be only the first phase or only the IT load, so ratios far from the median usually mean a scope mismatch, not an unusually cheap or dear campus. Investment includes the accelerators when the operator says it does; where it does not, the ratio understates the cost of a full build.

IT power per accelerator

An accelerator's own TDP is 0.7 kW for an H100 and well over 1 kW for a Blackwell GPU. The whole-IT figure also counts CPUs, memory, networking and storage, so it runs higher. Pairs are shown only when both figures are of the same kind (both estimates or both stated), and values far above or below the median are worth checking against the source before relying on them.

Announcement to first light

This compares like with like only loosely: an announcement may be a plan, a contract or a groundbreaking, and "first operational" may be a first hall of many. For the same measurement on supercomputers, see announcement to first light.