How energy metrics get gamed
Five common ways energy figures mislead, how to spot each one, and the rule in the Joule Ledger that prevents it. The patterns are generic; none describes a specific company.
Shifting the boundary
What it looks like. An efficiency figure measured at the chip is set beside a competitor's figure measured at the rack or the facility.
How to check. Is the meter point stated, and is it the same on both sides?
The Joule Ledger rule. Every result names its meter point; cross-boundary comparisons are not allowed.
Leaving out cooling
What it looks like. A "system" figure quietly excludes cooling and power conversion, so it looks like a facility figure but is not.
How to check. Does the figure say whether cooling and conversion losses are included?
The Joule Ledger rule. Facility claims use M4, which includes cooling and conversion; a partial ratio must be labeled partial.
Loosening the quality bar
What it looks like. Decisions per joule rises because more low-quality answers are counted.
How to check. Is accuracy reported beside the efficiency figure, against a stated quality gate?
The Joule Ledger rule. Only tasks that pass a stated quality gate count, and accuracy is reported beside Decision Yield.
Presenting estimates as measurements
What it looks like. A modeled or projected figure is reported as if it were metered.
How to check. Does the figure say Measured, Estimated or Projected?
The Joule Ledger rule. Every result carries a label; telemetry counts as Measured only after validation against an analyzer.
Picking the best months
What it looks like. An annual PUE is quoted from the coolest months of the year.
How to check. Does the figure cover a full 12 months of energy?
The Joule Ledger rule. Annual Overhead Ratio claims need at least 12 months of integrated energy.
A one-line disclosure anyone can use
When you publish an energy figure, state it like this: "Energy measured at the facility input (M4), over 12 consecutive months, labeled Measured, graded Declared." That one line answers most of the checks above.
The underlying rule
Most misleading figures fail one of two tests: the boundary is not stated, or the numbers would not balance. The meter points and the closure rule exist for exactly those two failures.
Questions
Is a misleading figure always deliberate?
No. Most come from honest shortcuts: convenient meters, short test periods or estimates that were never labeled.
Keep reading
Take the next step.
Five ways AI and data-center energy figures mislead: shifted boundaries, missing cooling, loose quality bars, estimates as data, picked months.