Account for the Cost Across the Whole Life of the System
Do we know what our AI costs the environment, from training to retirement?
Advanced: 22 controls
Who it applies to
Enter for every AI system whose letter is A (this area has only Advanced controls). Some of its controls apply only where you train or adapt the model (B1). Go as deep as your system letter, and implement in your mode.
Why it exists
Measure and account for the environmental cost of AI across its whole life — training, running, and retirement — because what you do not measure, you cannot manage or improve. The largest costs are often invisible: the intense energy of training or adapting a model, and the steady cost of running it at scale over years. Bring these into view, incorporate them into how you evaluate AI initiatives, and you turn sustainability from an aspiration into a managed number that sits alongside cost and risk in the decision.
What to actually do
Estimate and track the environmental cost from training through operation to retirement. Bring the energy and carbon of building, running, and eventually decommissioning each significant system into a figure you track, and fold it into the value assessment of the Governance strategy dimension, so environmental cost is weighed alongside financial cost and risk rather than ignored until it is regulated.
Controls
Advanced2 controls
Control 218
Account for lifecycle environmental cost of significant AI systems.
AdvancedDetectiveall rolesper system
What it is
Training, operation, and retirement, tracked and folded into initiative evaluation.
How to implement it · Minimum content
Scope: 'significant' systems, defined by a set threshold of spend or request volume
Training or adaptation: compute hours, energy and emissions of own runs; provider figures for bought models where disclosed
Operation: monthly requests × energy per request, or the provider's usage-based reports
Retirement: decommissioning effort and the storage that remains (archived model, retained data)
Method and data source for each figure, with its uncertainty noted
Annual total and trend, cited in the initiative's value assessment beside cost and risk
How to start
Take the most-used AI system and produce a first estimate of its monthly operating energy from request volumes and provider data.
Evidence
The lifecycle estimates per significant system and the value assessments that cite them.
Mode: light / standard / full
light
One person keeps provider-based estimates in a spreadsheet and adds the total to each business case.
standard
System owners update estimates quarterly; finance includes them in appraisals.
full
The sustainability function collects metered data and reports per system to the investment committee.
Control 219
Measure and disclose training compute and energy for large models you train.
AdvancedDetectivedeveloperper system
What it is
Compute, energy and, where measurable, emissions of each training run, recorded and disclosed.
How to implement it · Minimum content
Per training run: model, purpose, dates, hardware type and number of accelerators
Compute: accelerator-hours, and total operations where available
Energy: metered from hardware or facility, or estimated as power draw × hours × facility overhead
Emissions: energy × carbon intensity of the location, where measurable
Experimental and failed runs included, not only the final one
Disclosure: the figures and method published in the model's documentation
How to start
Before the next training run, switch on power and utilisation logging on the training hardware, or ask the compute provider for per-job energy reports.
Evidence
The training-run records and the published disclosure in the model documentation.
Mode: light / standard / full
light
The developer copies per-job figures from the compute provider into a log and the model documentation.
standard
The model team records each run in its experiment tracking and compiles the disclosure at release.
full
Energy is metered per job automatically and disclosed in a standard section of every release.
Evidence
The lifecycle environmental estimates; their inclusion in initiative value assessment. Training compute and energy measurements and disclosures.