Axis III · D8
Industrial AI economics
Thesis. Industrial AI has a cost in energy, infrastructure, human time and organizational capital, most of which remains invisible to classical TCO models. As long as a project does not provision its governance and revalidation cost, it significantly understates its real budget.
The distinction that cuts
Training cost vs governance cost. Industry has learned to account for the first (FLOPs, GPU-hours, electricity). It largely ignores the second (audit, revalidation, training, human oversight, regulatory management), which dominates ten-year TCO.
Typical market error
Calibrating ROI on the POC phase, where costs concentrate on compute and data, then projecting linearly, when production introduces a qualitatively different cost structure: support contracts, regulatory version uplift, model maintenance, periodic retraining, continuous operator training, legal charges in case of incident. The POC's financial model is wrong for production.
Failure signals
No 5- or 10-year TCO including compliance. No provisioning for post-drift revalidation. Energy treated as an externality, when AI data centre consumption approaches thresholds where electricity becomes the limiting factor (cf. IEA Electricity 2024 and 2025). Confusion between business ROI (measurable process gain) and project ROI (profitability of the isolated AI investment). No sensitivity analysis to regulatory tightening (EU AI Act, future codes of practice).
References
IEA Electricity 2024 and 2025; Patterson et al. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink (IEEE Computer, 2022); Stanford HAI AI Index Report (2025 edition) on training costs; for health, HAS methodology for medico-economic evaluation; ISO 14040 (LCA) where environmental assessment is required.
Ground of implementation
PREDICARE is a territorial programme framed at €3.5 M for 200 patients. It explicitly integrates GHT governance cost, periodic score revalidation, clinician training and ingestion pipeline maintenance. The model cost to programme cost ratio is inverted compared to the cliché. The model represents a minority fraction of the budget; the organization absorbs the rest. The instance illustrates a cost structure in a regulated territorial setting; it does not prove this ratio transposes outside public health.
Articulation
Conditions D9, since organizational absorption cost is generally the most underestimated line item. Intersects D5, since the cost of effective oversight is rarely provisioned.