Cost of retrieval by reading in AI systems: national release, September 10, 2026

1Main points

• An AI agent that finds information by reading whole files loads 16 to 47 times more bytes than the answer needs, measured on three ordinary questions on a production codebase (Table 2). On published coding tasks, retrieval-first cuts tokens 2.7 times end to end.
• Applied to worldwide AI spending, the reference factors put avoidable reading at … in 2026, … every second: money paid for text nobody needed (Technical note, item B).
• Applied to reported budgets: … a year for U.S. federal agencies on $7.2 billion of AI obligations, … for JPMorgan Chase on $2 billion, … across the customers of one vendor with a $65 billion run rate (Table 1).
• Federal prime contracts naming AI total $427.8M for fiscal year 2026 to date, a labelled floor of the Brookings figure; every state has a dated release with its own dateline (Table 4).
• The habit ends with a local index that answers "where is it" before the agent reads: open source, four commands, nothing leaves the building (section 10).

How the cost is cut, in three steps

1. Index once. A local engine reads the folder one time and builds a card index on the agent's own machine. One command; nothing leaves the building.
2. Ask "where" first. The agent asks the index and gets the file, the line and the symbol back in milliseconds instead of opening files to look.
3. Read only that. It reads the passage the pointer names. Measured: 103 KB became 6 KB, 244 KB became 6 KB, 242 KB became 5 KB.

First release of this series; the change line begins with the next one.

2Your organisation

Pick a published budget or type your own. The result applies the same three factors as the world scenario; each is a control under the Technical note. Nothing you type leaves the page.

Table 1Reported AI spending and usage, by sector and organisation, with the avoidable share

Public disclosures, each linked to its source, and what the index's reference factors imply for each. B1·B2·B3 for a general AI budget (inference share × agent share × avoidable share = 6.3 %); B2·B3 for vendor revenue, which is already customer spending on tokens (25.2 %); B3 alone for a coding-agent product, where the whole figure is agents reading code (63 %); none for infrastructure and prices. Move a factor under the Technical note and the column recomputes.

Score cards measured with the plugin and published by organisations themselves will appear as a separate row group, dated, only with written permission ([email protected]). None yet. The list is public: data/disclosures.json.

Table 2Cost of one answer, by retrieval method, measured

Three ordinary questions put to an agent on one production Next.js codebase, September 2026. "Read" is the size of the whole files behind the answer, which a grep-and-read agent loads into its context; "Index" is what the same agent read when a local index answered first.

Tokens ≈ bytes / 4. A published case study on real coding tasks reports 2.7× fewer output tokens end to end; this release uses the conservative 1 − 1/2.7 = 63 % as the avoidable share, not the 94–98 % above.

Table 3Estimated annual cost of retrieval by reading, by team size

Select a row to set Chart 1. Reading bill = developers × requests per day × working days × tokens read per request × price per million tokens. Avoidable = reading bill × avoidable share.

Reference parameters: . Adjust under Technical note, item A. Last column: avoidable cost divided by $185,000, a reference fully loaded cost of one senior engineer in the United States.

Chart 1Ten-year projection of the reading bill, two methods

Growth starts at the Gartner 2026 rate for worldwide AI spending and tapers to the long-run parameter. The shaded area is the gap: what is read and paid for without being needed.

Projection of the selected team's reading bill. Assumptions in Technical note, items A and B.

Table 4Federal contract obligations naming AI, by state of performance, FY2026 to date

Prime contracts whose descriptions name artificial intelligence, machine learning or language models. USAspending.gov, refreshed daily. A floor, not a total: AI work inside larger contracts is usually unlabelled. State debt at end of fiscal year 2023 from the Census Bureau's Annual Survey of State Government Finances, shown for scale. Each state has its own release page and a press release datelined at the state capital.

All 51 jurisdictions with the scenario column: state releases.

Tables 5 and 6Awarding agencies and recipients, same filter

Reference figuresShown as published, for scale

Official series reproduced without adjustment. They size the environment in which the cost above is incurred; nothing on this page attributes them to it.

The remedyA local index that answers "where" before the agent reads

The cost above has one cause and one cure. An agent that needs one function finds it by reading whole files; a local index on the agent's own machine answers "where is it" first, and the agent reads only the passage. The habit is called retrieval-first, and it is what Table 2 measured.

curl -fsSL https://xerj.org/get | sh
xerj --insecure --data-dir ./.xerj-data &
xerj autoindex .
xerj search "the question you were about to grep"

Technical noteEstimation method and parameters

Item A sets the team estimate (Table 3, Chart 1). Item B sets the world scenario, Table 1 and the figures in Main points. Every default carries its source; every control is yours. The full note, with what the index does not claim, is on the technical note page.

Data, embed and pressFor newsrooms, offices and teams

<script src="https://aiburnclock.org/embed.js" data-scope="world"></script>
<script src="https://aiburnclock.org/embed.js" data-state="TX"></script>

Notes and contactSources, limits, corrections

• Sources. Treasury, Debt to the Penny · USAspending.gov · Census Bureau, Vintage 2024 population estimates and Annual Survey of State Government Finances · Gartner, May 2026 · the disclosures linked in Table 1 · reference-coding case study.
• What this release does not claim. It does not claim that AI causes public debt; it does not claim the federal contract figure is total federal AI spending; it does not claim the world scenario is a measurement. Organisations are named only from their own public disclosures. Full list on the technical note.
• Corrections. Published on the technical note with the date: [email protected] or an issue on the public pipeline.
• Offices. A one-page memo with your state's or agency's figures, their sources, and a pilot that runs on one laptop with no procurement is available on request. Put the state or agency in the subject line.