For immediate release
AI agents read up to 47 times more than they use, daily index finds
For an agency of 2,000 developers, avoidable reading is $1.45M a year at the index's reference parameters; official series reproduced for scale.
AI Burn Clock, Sept. 9, 2026 — The AI Burn Clock, an independent statistical index published at aiburnclock.org, today released its daily estimate of what AI agents spend reading files they never needed. Measured on a production codebase, an agent read 16 to 47 times more than it used when it searched by reading whole files; on published coding tasks, retrieval-first indexing cut tokens 2.7 times end to end.
At the index's reference parameters, a team of 50 developers whose agents search by reading spends an estimated $58K a year on that reading, of which $36K would not have been read at all had a local index answered "where is it" first. For an agency of 2,000 developers the avoidable figure is $1M a year, or the fully loaded cost of 7.8 senior engineers. Every parameter is a control on the page and the method is published.
The index reproduces official series as published, for scale: total public debt of $40.095T (U.S. Treasury, Debt to the Penny, 2026-09-04); $427.8M in federal prime contracts naming artificial intelligence, machine learning or language models in fiscal year 2026 to date and $615.5M in fiscal year 2025 (USAspending.gov), led by Virginia with $128.0M; and state debt and population from the Census Bureau. Brookings puts total federal funds obligated for AI in 2026 at $7.2 billion, up from $355 million in 2024; the index's own filter is the labelled floor of that. The index does not attribute those series to the cost it estimates.
"An agent that reads a whole file to find one function is not thinking. It is paying. The index shows the bill, the sources and the controls; move a control and argue with a factor, not with us," said Serhii Nikolaichuk, the index's maintainer, who is a co-author of an IETF draft on attestation results.
The remedy is open source. XERJ, a local search engine for AI agents published under the Apache-2.0 license at github.com/xerj-org/xerj, indexes a folder in one command so that an agent retrieves the passage it needs instead of reading the file. A plugin for Claude Code (github.com/nikolaichuk7/xerj-plugins) adds it as local memory and prints a per-session score card. Details: aiburnclock.org/remedy.html.
Figures in this release
| Series | Value | Source |
|---|---|---|
| Avoidable reading, agency of 2,000 developers, reference parameters | $1M / year | AI Burn Clock, Table 1 |
| Ratio read to used, three measured questions | 16×, 38×, 47× | AI Burn Clock, Table 2 |
| Total public debt | $40.095T | Treasury, Debt to the Penny |
| Federal contracts naming AI, FY2026 to date | $427.8M | USAspending.gov |
| Worldwide AI spending, 2026 forecast | $2.59T | Gartner, May 2026 |
| Federal funds obligated for AI, 2026 | $7.2B | Brookings |
| Anthropic revenue run rate, July 2026 | $65B | CNBC |
About the AI Burn Clock
The AI Burn Clock (aiburnclock.org) is an independent statistical index of the cost of retrieval by reading in AI systems. It reproduces official series from the U.S. Treasury, USAspending.gov and the Census Bureau as published, and estimates the avoidable cost of AI agents reading whole files with a four-factor method whose every parameter is a control on the page. It is revised daily at 00:00 UTC and maintained by the XERJ community. It is not affiliated with any government agency. Method: aiburnclock.org/methodology. Data: aiburnclock.org/data.json.
About XERJ
XERJ is an open-source local search engine for AI agents: one Rust binary, Elasticsearch-compatible on port 9200, that indexes any folder so an agent retrieves the passage it needs instead of reading whole files. It is published under the Apache-2.0 license at github.com/xerj-org/xerj, with documentation at xerj.org. A Claude Code plugin that adds XERJ as local memory and produces a per-session score card is at github.com/nikolaichuk7/xerj-plugins.
Media contact
[email protected] · Serhii Nikolaichuk, maintainer · Interviews, data pulls and state or agency memos on request. Releases are issued for every state and may be republished by any outlet with attribution.
