Technical note

This note exists so that anyone can check the index, reproduce it, or disagree with a specific factor rather than with the whole. Three kinds of figures appear in the release: official series reproduced as published; one estimate built from stated parameters by four multiplications, each parameter a control on the page; and measurements from working systems.

1Official series, reproduced as published

Series	Source	Use in the release
Total public debt outstanding	Treasury, Debt to the Penny, API	Latest daily value. The on-screen figure advances at the mean change per second over the last 30 records. Shown for scale; not attributed to any cost in this index.
Federal contract obligations naming AI	USAspending.gov, API	Prime contract obligations (award types A–D) whose descriptions match any of: `artificial intelligence, machine learning, large language model, generative AI, LLM, AI-enabled, natural language processing`. FY2025 full year; FY2026 to date, by month, by state of performance, by awarding agency and by recipient. A floor by construction: AI work inside larger contracts is usually unlabelled; grants, internal spend and cloud consumption are excluded.
State debt at end of fiscal year	Census Bureau, Annual Survey of State Government Finances, FY2023	Table "Debt at end of fiscal year", in thousands of dollars, converted. Shown for scale on Table 3 and state releases.
Resident population	Census Bureau, Vintage 2024	Per-resident figures on state releases.
Worldwide AI spending, 2026	Gartner, May 2026	$2.59 trillion, +47 % year on year. Base of the world scenario and the first-year growth rate of Chart 1.
Reported AI spending and usage, by sector and organisation (Table 6)	Gartner, IDC, Brookings, U.S. Department of Defense, GSA, CNBC, Bloomberg, Fortune, GeekWire, Dealroom; each row links to its source	Reported figure, what it measures, date and link, reproduced as published. The index applies its reference factors to each row (B1·B2·B3 for a general AI budget, B2·B3 for vendor token revenue, B3 for a coding-agent product) and labels the result a scenario. Rows are added from public disclosures only; the list in data/disclosures.json is public and corrections are welcome.

2Item A: the team estimate (Table 1, Chart 1)

reading bill  = developers × requests per day × working days × tokens read per request × price per million tokens
avoidable     = reading bill × avoidable share
per second    = avoidable / 31,557,600

Parameter	Reference value	Provenance
A1 requests per developer per working day	40	An agent working through a task issues dozens of tool calls; an ordinary day. Control range 5–200.
A2 bytes read per request, the old way	150 KB	Table 2 measured 103,023, 243,640 and 241,859 bytes of whole files behind three answers on one production codebase. Tokens ≈ bytes / 4. Range 20–600 KB.
A3 price per million input tokens	$3.00	2026 list prices run from under a dollar for small models to fifteen for frontier models. Range $0.25–15.
A4 working days a year	250	Unattended agents push this to 365. Range 100–365.
B3 avoidable share (shared with item B)	63 %	1 − 1/2.7, from the 2.7× fewer output tokens reported on real coding tasks in the reference-coding case study. Table 2's own ratios (16×, 38×, 47×) would give 94–98 %; the release uses the conservative figure.

3Item B: the world scenario (Reference figures)

avoidable per year = base spend × inference share × agent share × avoidable share

Parameter	Reference value	Provenance
base spend	$2.59T	Gartner, 2026 forecast.
B1 inference share	25 %	Gartner puts AI infrastructure above 45 % of spend and models and platforms as the fastest-growing slice; inference is part of the remainder. A middle estimate; range 5–60 %.
B2 agent share	40 %	Share of inference done by agents that read files and documents: coding agents, document assistants, retrieval pipelines. No public census exists; an estimate; range 10–90 %.
B3 avoidable share	63 %	As above.
B4, B5 growth	47 % in year one, tapering to 20 % by year five	Gartner's 2026 rate; the taper is an assumption.

At the reference values the world scenario gives about 6.3 % of AI spend, on the order of $163 billion a year, as avoidable reading. The figure is a scenario: three of its four factors are estimates by construction, which is why each is a control on the page.

4Measurements (Table 2)

Three questions were put to an agent on one production Next.js codebase in September 2026, once with the agent reading whole files and once with a local index returning pointers and passages. The whole files behind the answers measured 103,023, 243,640 and 241,859 bytes; the index path read 6,405, 6,389 and 5,101 bytes. The index used was XERJ; any index that returns passages instead of files produces the same effect. The measurement script and its output are published with the pipeline.

5What the index does not claim

It does not claim that AI causes public debt. The Treasury figure is scale, not cause, and no cost on the page is attributed to it.
It does not claim the federal figure is total federal AI spending. It is the labelled floor of contract obligations.
It does not claim the world scenario is a measurement. It is a forecast multiplied by three stated shares.
It names organisations only from their own public disclosures or from published reporting, each linked to its source in Table 6. The avoidable column beside each row is the index's scenario at the reference factors, not the organisation's figure. Measured score cards appear only with written permission.

6Revision and correction policy

Source series are refreshed daily at 00:00 UTC; the release date at the top of every page is the refresh date. Method changes are versioned (currently 2026.1) and listed here with the date. Corrections: [email protected] or an issue on the public pipeline. Corrections are published on this page with the date.

Date	Change
2026-09-09	Method version 2026.1. First release.
2026-09-09	Table 6 added: reported AI spending and usage by sector and organisation, from public disclosures, with the scenario column. State releases datelined at the state capital; the national release carries no city.