Storage-device PPI121.4
+30.4% exposed ÷ +7.5% control (Core PPI (final demand less foods and energy))
Storage prices rose 30.4% against a 7.5% core-PPI control, the widest gap in the basket.
Crowding Index
AI infrastructure cost index and methodology
The index compares price and availability changes in AI-exposed markets with matched markets that have less direct AI demand. January 2024 is the common baseline of 100.
Current score: v2 covers 3 scored components with locked baselines and a matched core-PPI control. Other basket markets remain indicators until each has a locked comparison.
The current calculation
121.4 storage × 106.9 commercial power × 103.4 industrial power = 110.3
Storage prices rose +30.4% while the core-PPI control rose +7.5%, a 121.4 reading. Commercial electric power rose +14.9% against the same control, a 106.9 reading, and industrial electric power rose +11.1% for a 103.4 reading. Their equal-weight geometric mean is the 110.3 composite. The control basket is the core producer price index, final demand less foods and energy, chosen because the v1 all-commodities control was itself moved by the 2025-26 energy surge.
- 01Rebase
Set every exposed and control series to 100 in January 2024.
- 02Compare
Divide each exposed change by its preregistered matched control.
- 03Combine
Take an equal-weight geometric mean of eligible component ratios.
<100Less relative pressure
100January 2024 baseline
110.3Current score, June 2026
200Twice the relative pressure
No capExtreme scarcity can exceed 200
Scored components
What each measure contributes to 110.3
Component readings range from 103.4 to 121.4. A reading above 100 means the exposed market outran its control; below 100 means it lagged. Weights are equal; the sensitivity range is simply the individual component readings.
Commercial electric power PPI106.9
+14.9% exposed ÷ +7.5% control (Core PPI (final demand less foods and energy))
Commercial electric power rose about twice as fast as core producer prices, with a sharp June 2026 jump.
Industrial electric power PPI103.4
+11.1% exposed ÷ +7.5% control (Core PPI (final demand less foods and energy))
Industrial electric power rose well above the core-PPI control too, with a similar June 2026 jump.
Formula definition
A ratio of ratios, then a geometric mean
For component i, divide its exposed-market change since baseline by the matched control change. Invert measures where a lower value means more scarcity. Combine the resulting positive ratios so that proportional moves are treated consistently.
d is +1 when higher means more pressure and −1 when lower means more pressure.
Family weights sum to 1. Missing or unmatched series do not silently become zero.
Indicators outside the scored basket
Series that stay indicators until they pass exposure validity
These official producer price indexes are candidate component inputs, published for reference even though they are not part of the composite. The transformer index remains an indicator because its current BLS series has no January 2024 observation to rebase against; it joins the score once a comparable baseline is locked. The semiconductor PPI was a scored component in v1 and is now excluded on exposure-validity grounds: Demoted from scored component to published indicator on exposure-validity grounds. The domestic semiconductor PPI does not price the imported accelerators and memory AI buyers actually purchase, and the aggregate series has a long-run quality-adjusted downward drift unrelated to AI demand.
Specification sensitivity
How the score moves under other reasonable specifications
The 110.3 headline is one specification choice among several defensible ones. Every variant below is precomputed from the same raw observations in the downloadable CSV, including the v1 spec at 96.8, so an unfavorable reading is as visible as a favorable one.
v1 published spec96.8
Storage and semiconductors scored against the all-commodities PPI control (the original v1 spec).
v1 basket, core control102.8
Same storage-and-semiconductors basket, re-run against the v2 core-PPI control.
v2 basket, semiconductors scored104
The v2 four-series basket with the semiconductor PPI left in as a scored component, against the core-PPI control.
v2 basket, all-commodities control103.8
The v2 three-component basket, re-run against the v1 all-commodities control instead of core PPI.
Family-weighted112.9
Storage weighted 50% and the electricity family (commercial and industrial power) weighted 50%, instead of equal weights across all three components.
AI-exposed basket
Markets with direct AI demand
- HBM / advanced-memory capacity
- Enterprise storage devices
- Power and distribution transformers
- Switchgear and power-control equipment
- Gas-turbine manufacturing slots
- Electrician and mission-critical trade labor
- Large-load electricity and grid capacity
- Data-center corridor construction inputs
Control rules
Use a fair comparison or leave it blank
- Same measurement family and release cadence
- Similar input intensity before January 2024
- Low direct AI exposure, with energy excluded from the core-PPI control
- The all-commodities PPI stays published as a sensitivity check, not the control
- Any comparison changes are logged in the changelog before recalculation
- Unmatched series remain component indicators outside the matched score
Calculation standard
Six rules define the score
- 1Series coverage
At least 80% of the exposed and comparison baskets have current observations.
- 2Expected direction
Each series defines how greater scarcity moves the value.
- 3Comparison match
Every scored series has a documented lower-AI-exposure comparison.
- 4Weighting
Published results show equal-weight and sensitivity calculations.
- 5Data version
Raw values, calculations and retrieval dates are available for download, including revisions.
- 6Independent check
A reviewer outside the project can reproduce the result from the published file.
Open data
Download the baseline evidence
Each row includes the observation, unit, date, grade, evidence boundary and source URL, plus the control series and composite so the 110.3 score can be reproduced.
The original Who Pays for AI dataset compilation is licensed under CC BY 4.0. Source data remains subject to each cited publisher’s terms.
Common questions
Crowding Index questions and answers
- What is the Crowding Index?
- The Crowding Index measures how prices and availability in AI-exposed input markets have changed relative to a matched control series, rebased so January 2024 = 100. It is an index level, not a probability and not the percentage of a price caused by AI.
- What is the current Crowding Index score?
- As of June 2026 the Crowding Index reads 110.3 (January 2024 = 100), 10.3% above the baseline. v2 scores 3 components against a core-PPI control (final demand less foods and energy): storage at 121.4, commercial power at 106.9, industrial power at 103.4.
- How is the Crowding Index calculated?
- For each component, divide the AI-exposed market's change since January 2024 by a matched control series' change over the same window, then combine the eligible component ratios as an equal-weight geometric mean. The control basket is the core producer price index, final demand less foods and energy, and every raw value, date and calculation is published in a downloadable CSV so the score can be reproduced independently.
- Why did the Crowding Index change from 96.8 to 110.3?
- The v1 score of 96.8 was an artifact of two instrument problems, not weak AI demand. First, the semiconductor component failed exposure validity: the domestic semiconductor PPI does not price the imported accelerators and memory AI buyers actually purchase, and the aggregate series carries a long-run quality-adjusted downward drift. Second, the v1 all-commodities control was contaminated by the 2025-26 energy and metals surge, a market AI demand itself plausibly contributes to. v2 demotes the semiconductor PPI to a published indicator, switches the control to the core PPI (final demand less foods and energy), and adds two electricity components with locked January 2024 baselines. The v1 score is preserved in the changelog and in the sensitivity table below, alongside variants that read less favorably than 110.3.