Case study · Public evidence reference
A graded, sourced reference on who pays for AI
A public reference site that grades who actually bears the cost of the AI infrastructure buildout: electricity, water, grid equipment, and memory. Every number carries a primary source and two dates, published and last checked, and one person plus AI agents keeps it current.
What it is
whopaysforai.org is a public reference work that answers one contested question: who actually pays for the AI infrastructure buildout. The answer is spread across electricity, water, grid equipment, skilled trades, and memory and storage, and most public numbers about it are either marketing or outrage. The site treats each disputed number as a claim to grade.
Like the cheatsheets pipeline, this one is public, so every claim below is checkable on the live site while you read this page.
The evidence standard
The rule is written down and enforced: every public quantitative claim carries a primary source, a publication date, a separately recorded checked date, and an attribution grade. A number with a strong direct source and a number inferred two steps removed do not get to look the same. A reporter or a ratepayer advocate who needs a defensible figure in a hurry can take one off the page and see exactly how many steps it sits from the filing.
You can see the standard doing work on the claim report cards. Take a viral claim, for example that every AI prompt drinks a bottle of water, and the page grades what the evidence actually supports, which is rarely what either side of the argument came for. The resource ledgers do the same for whole cost categories, tracing water use or grid interconnection from primary filings to a graded figure.
The release gate
The site publishes an index only when it has earned it. A derived cost index does not ship until a documented release gate passes: the comparison set matches, the weighting is fixed, the data version is pinned, and an independent check clears. Holding the headline number back until the gate passes is the entire product. A reference that inflates its own confidence is worse than no reference at all, because somebody will cite it and then defend it.
This is the same invariant as the regulated-lender and cheatsheets work: a written standard decides what ships, and looking finished has never counted as passing.
How it is built and kept current
One person plus AI agents produce and maintain the full methodology, the ledgers, and the statistics catalog. The build is deliberately boring: a static HTML snapshot served from my own server behind Cloudflare, with no runtime to break. Machine consumers are first-class, a schema.org Dataset with a downloadable baseline CSV and a concise llms.txt route guide, so the data is easy to reuse and cite. A rendered SEO gate fails the build on a missing title, a duplicate description, or a broken page, and scheduled agents watch the news for rate cases and viral cost claims that a graded page can answer.
There is a public corrections path. Readers dispute a number through an intake form, and a disputed figure arrives as a correction to weigh, never as an instruction to obey.
Limits
This is a personal-scale reference. There is no funded research institution behind it, and no independent panel: the grades come from me and the pipeline, against a written standard. The method is consistent, sourced, and dated, and any individual grade is arguable. Coverage stays narrow on purpose, sitting on the inputs where the evidence is strongest, and it widens as sources firm up.
Where it fits
The Antech and regulated-lender case studies show governed AI carrying load inside companies, with the strongest numbers internal. The cheatsheets pipeline shows a governed build process at personal scale. This one is the evaluation half: grading contested claims when the underlying evidence is thin, in public, where you can check the method against the sources without asking me for anything.
How costs are attributed and graded, the methodology ↗ Email me about this Where I fit best ← All case studies