China AI Data Centre Visits: Question the Energy Evidence
Energy claims need a boundary, denominator and time period. Use a facility-versus-workload worksheet before comparing AI data centre efficiency.
Direct answer. Ask what equipment the energy measurement includes and what useful work the denominator represents. Facility efficiency and energy per successful AI task are different claims; neither can be inferred from the other without additional evidence.
Name the claim before comparing numbers
A China AI data centre visit may include cooling exhibits, power dashboards and model-serving demonstrations. Treat each as a different evidence source. A cooling-system demonstration does not establish annual site performance, and a model benchmark may exclude facility overhead. Ask whether figures are observed, modeled or contractual targets.
Use permitted aggregate data rather than requesting sensitive infrastructure drawings. Record the measurement period, meter boundary, operating load and equipment included. A short low-load interval is not interchangeable with a full operating year. If the presenter cannot disclose the boundary, keep the comparison qualitative.
Keep two worksheets separate
| Facility worksheet | Workload worksheet |
|---|---|
| Metered equipment and cooling boundary | Model, input size and serving configuration |
| Measurement period and load profile | Completed tasks, retries and failures |
| Measured versus projected result | Energy associated with the measured workload |
| Excluded infrastructure | Quality threshold defining a useful completion |
For workload comparison, ask what counts as a successful task. Reducing energy while reducing answer quality may not improve the business outcome. Record retries separately so that an efficient individual request does not conceal a costly recovery loop. These are proposed comparison questions, not a claim that any particular host reports them.
NIST AI RMF Core includes environmental-impact assessment in MEASURE 2.12. It provides a voluntary risk-management rationale, not a methodology for calculating a facility efficiency ratio or proof of a site's sustainability.
Leave with a bounded evidence request
Request a redacted measurement-method note and the definition of the operating period behind a headline number. If the business case depends on electricity cost, have your engineering and finance teams examine tariffs, utilization and service configuration separately; this article does not supply an investment return estimate.
Use the deployment evidence guide to distinguish observed results from projections. Arrange a focused visit when you can name the workload and infrastructure questions your team needs answered.