Review of 103 studies finds AI's water and supply-chain footprint is barely measured

The review reports a Gini coefficient of 0.74 for lifecycle terms and an Intervention Bias Ratio of 3.6, which the authors read as a persistent overemphasis on compute-focused fixes. Mean coverage reached up to 39.5% for training and inference phases, against 17.1% for green end-of-life and circularity.

Measurement is a second weak point. Cloud-provider emissions dashboards use differing methods, often leave out water and Scope 3 coverage, and block cross-system comparison. The authors propose a five-phase lifecycle mapped to Life Cycle Assessment stages and a hybrid approach that calibrates estimator tools against direct power metering. They describe these constructs as candidate tools until validated.

The search began in October 2024 and finished data extraction in August 2025. It covered Web of Science (1,035 records), ScienceDirect (318), IEEE Xplore (326) and 16 other sources, with two independent assessors screening 757 de-duplicated records and excluding 659 at title and abstract stage. Two more were dropped at full text, leaving 96, and 7 added through citation chaining gave 103.

The term analysis showed head terms such as model training, deployment, inference and power consumption dominating. Long-tail topics, including carbon-aware scheduling, power capping, conflict minerals, remanufacturing and e-waste pathways, appeared rarely but with visible growth. Recent articles show broader coverage than pre-2020 work, yet the authors say the gains do not close the upstream and downstream gap.

On measurement, the review argues indirect estimators must be calibrated against direct readings such as RAPL, rack-level meters and facility telemetry. Unresolved choices include boundary conventions, emission-factor provenance and amortisation of embodied impacts, which the authors say make cross-study comparisons fragile.

The authors say hardware strategies such as accelerator choice, placement across the edge-cloud continuum, power capping and carbon-aware scheduling can raise performance per watt. They propose governance through Plan-Do-Check-Act cycles with phase completion criteria and performance-environmental thresholds.