The Consumer Incidence of the AI Buildout: A Reproducible Ledger

This study assembles a reproducible consumer-incidence ledger: how the AI data-centre buildout reaches a single household through more than one channel, with each channel's AI-attributable share explicitly bounded and the contested points given equal weight. It continues the AI Cost Watch series on the unit economics of the AI buildout, as a standalone study. Five channels are priced and graded into three tiers of certainty - direct household cash (device prices, involuntary AI in consumer subscriptions, electricity), a public-budget shift (foregone state tax revenue, shown per taxpayer), and contingent retirement exposure (a stock at risk, never summed) - while a sixth channel, water and the local environmental costs around it, is held outside the ledger because an honest national dollar figure is unattainable (double-counting with electricity, agriculture-dominated base rates, and largely non-monetary scarcity costs). Incidence is bimodal, so two archetype households are shown side by side: national-average and data-centre-corridor. Leading with the low bound, the direct annual cash incidence is about 31 to 223 dollars for a national-average household and 77 to 409 dollars in a data-centre region; the public-budget cost is 1 to 10 dollars per taxpayer nationally and 80 to 280 dollars in a host state; AI-exposed retirement equity is 7,900 to 29,700 dollars of the median balance, reported as risk and not as a charge. The energy channel is the contested core and is presented with both sides at full weight - PJM's independent market monitor attributing 63 per cent of the capacity-auction increase to data centres, against SemiAnalysis's market-design argument and the flat ERCOT comparison - and reframed as a deferred rate shock still moving through utility-commission dockets. Every coefficient is a named, tunable input in an accompanying standard-library script, mapped to a primary or named source with a motive tier; the energy channel was put through a self-red-team and an independent, non-author review before release. Independent analysis, not investment advice. Produced with assistance from Claude Opus 4.8, an Anthropic model. Anthropic is a direct participant in the buildout costed in this study (data-centre operation and contracting, a public grid-upgrade pledge, and the Claude Code product named in the corporate-cost reporting referenced), so any reading touching Anthropic is non-neutral and primary coverage should be preferred. The reproducible script is provided so the analysis does not rest on the author's or the model's good faith. https://nmairesearch.github.io/