Performance per watt surfaces as a competitive axis — AMD claims its latest AI systems are 4x more efficient than two years ago, as reported by The Register

The leading-indicator reading lies in the pitch for AI hardware shifting from absolute performance toward performance per watt. What a vendor chooses to boast about reflects the constraint on the buyer's side more than the number itself does. Where new data center build is rate-limited by the interconnection queue, the buyer's question stops being "how fast is it" and becomes "how much more can I run on the same feed." Even as an unverified claim, the fact that this pitch is being pushed to the front is itself an observation point for that shift.

The reading splits between (a) power efficiency being written into actual procurement criteria, showing up in contract terms and tender specifications, and (b) it remaining a marketing frame around the power constraint while purchasing decisions still turn on absolute performance and available supply. Under (a), pressure builds for common measurement conditions and third-party benchmarks get built out. Under (b), self-reported vendor figures accumulate side by side and stay mutually incomparable.

What to watch next: (1) whether third-party benchmarks such as MLPerf accumulate results with power measurement attached; (2) whether per-watt metrics appear in hyperscaler procurement terms or disclosures; (3) whether competitors begin publishing figures on the same footing — followers would mark a genuine shift of axis, silence would leave this a solo pitch.