A Practical Sustainability Screen for AI Infrastructure
Review the workload, facility, power source, cooling design, and supporting records before accepting an environmental claim.
A label such as “green AI” does not identify what was measured. Start with a defined workload and system boundary. Then keep each input, estimate, and source date visible.
Define the workload
- State whether the measurement covers training, inference, data preparation, or all three.
- Record the model version, hardware type, run time, utilization, and number of runs.
- Separate measured electricity from estimated electricity.
Check the facility inputs
- Electricity: Record the facility meter boundary and measurement period.
- Cooling: Identify the cooling method and the energy included outside the computing equipment.
- Water: Distinguish withdrawal from consumption and state the local water source.
- Backup power: Include test runs and operating hours when they are material to the review.
Describe the power claim precisely
A renewable energy certificate, a power purchase agreement, and direct on-site generation are different instruments. Record the contract type, region, time period, and matching method. Do not describe annual matching as hourly matching.
Include hardware use and replacement
Operational electricity is only one part of the decision. Review expected utilization, useful life, repair options, and the plan for equipment that leaves service. A lower-power server does not improve the full system result if poor utilization requires more hardware.
Keep the comparison reproducible
Compare alternatives with the same workload, quality threshold, measurement period, and system boundary. Record any excluded energy or water use. If a vendor supplies an estimate, retain the vendor method and date with the estimate.
What a review record should contain
- The decision being reviewed.
- The workload and facility boundary.
- The measured values and source records.
- The assumptions used for missing values.
- The person who reviewed the result and the review date.
This record does not prove that one design is sustainable. It makes the claim specific enough for another reviewer to test.
Related articles
How to Review Water Risk for an AI Data Center
A site-diligence checklist for water sources, cooling design, permits, consumption, and operating limits.
When Carbon-Aware Scheduling Helps, and When It Does Not
How to review movable workloads, grid data, deadlines, baselines, and reporting before claiming a saving.
Why AI Analysis Still Needs Explicit Human Approval
How specialist analysis, review records, approval authority, and external action stay separate.