How we attribute AI data center costs

The ruling

Every claim gets a verdict, not a shrug

A claim report card opens with a one-word ruling. When the cited evidence settles the question, the ruling is true or false and the site says so plainly. “Unproven” is used only when no published measurement supports the claim in either direction, and it is itself a definite finding: the claim is not established, and anyone asserting it is going past the evidence.

The rubric

Four attribution grades, plus Background

The ruling answers the claim. The grade records how much of the measured effect the evidence actually attributes to AI. Background is also deliberate: it marks a figure that describes market or policy conditions without assigning AI attribution.

Unit of analysis

Use the narrowest claim the evidence can answer

“Data centers increase electricity demand” and “AI raised my bill” are different claims. The first has national evidence. The second requires utility-level evidence. Each page keeps the original wording in view so a narrow finding does not turn into a broader claim when it is repeated.

Evidence hierarchy

Start with primary sources

1. Measured public seriesBLS, EIA, FRED, queue and tariff data.
2. Agency analysisMethods and models with definitions, ranges and dates.
3. Filings and producer statementsCapacity and order-book evidence from commercially interested sources.
4. Peer-reviewed researchEvidence for mechanisms and uncertainty; publication dates define its market period.
5. Trade press and anecdotesLead sources that require confirmation from primary evidence.

Causal checklist

Five questions on every report card

1. What exactly moved: price, wage, allocation, lead time or capacity?
2. What is the proposed AI-demand pathway?
3. Which non-AI buyers and supply shocks share the market?
4. What fair comparison would help isolate the effect?
5. What evidence would change the grade?

Index policy

How the index is calculated

The Crowding Index divides each exposed market’s change since the January 2024 baseline by a matched control, then combines the eligible component ratios as an equal-weight geometric mean. The control is the core PPI, final demand less foods and energy, because the all-commodities PPI used in v1 was itself moved by the 2025-26 energy and metals surge, a market AI demand plausibly contributes to. A scored component must price the market where AI actually buys: this exposure-validity test excludes the domestic semiconductor PPI, whose survey frame misses imported accelerators and memory, and it now publishes as an indicator instead. A series without a locked baseline or matched comparison stays an indicator instead of silently becoming zero. All specification changes, including this one, are logged in the changelog with the prior score preserved.

Review process

A person reviews every update

Software fetches releases, flags changed values and drafts summaries. A named editor approves every publication and grade change. Each update records what changed, why it changed, the source, the check date and the editor’s approval.

Conflicts and corrections

The capital-markets ledger gets extra scrutiny

The editor is a partner in a crypto fund whose returns benefit when capital moves away from competing AI infrastructure. The capital-markets ledger therefore requires a preregistered comparison method and external review. Corrections are judged on the evidence and logged whether they strengthen or weaken an AI attribution.