1. Begin with the actual question
Bill Igoe“Can we help investor Bill Ackman answer his question at the end of this image?”
Ackman asked whether demand for intelligence and energy could remain strong as rates rise, while higher interest costs feed into prices and potentially make rate hikes inflationary. We separated that into an investment-sensitivity hypothesis and a price-pass-through hypothesis. The Fed's September 16, 2026 statement records a quarter-point increase to a 3¾–4% target range. The available construction series stops before that decision.
2. Look for a change in the rate–inflation relationship
Bill Igoe“Can we show a regime shift using a binary_split?”
An initial BINARY_SPLIT diagnostic for PCE inflation and lagged federal funds changes identified breaks associated with the 2020–22 inflation cycle, rather than a clean break at the start of the AI buildout. In the latest segment the simple relationship had little fit. The diagnostic chose breakpoints from the data; it could not establish that policy caused an inflation regime.
3. Compare potential AI investment with other equipment
Bill Igoe“What about a DID analysis?”
We tried DID, using real computer investment as a rough treated-sector proxy and industrial and transportation equipment as comparisons, with 2023 as the post period. The estimated quarterly growth differential was about +2.69 percentage points, with a conditional 95% interval from −3.15 to +8.53 points. With 20 pre-period and 14 post-period quarters, the result did not distinguish the differential from zero. Computers also include substantial non-AI demand.
4. Broaden the sector check, then narrow the measure
Bill Igoe“Can we use a loop_object function for gathering different sectors? Then run a regression on Rates changes?”
Bill Igoe“Maybe a search would give you a better list.”
LOOP_OBJECTS compared six real investment categories against earlier rate changes. Their estimated slopes varied substantially: computers −5.43, software −0.88, industrial equipment −0.26, transportation equipment +3.49, other equipment −0.60, and broad nonresidential structures +3.88 percentage points of growth per one-point rate move. The mixed signs, broad categories and endogenous rates made this a poor direct test of AI-specific construction. We searched official sources and chose the Census Bureau's narrower data-center component of private office construction. The Federal Reserve's AI data note explains why buildings and computing equipment are complementary and incomplete proxies.
5. Make the public data import reproducible
Bill Igoe“READ_JSON_SERIES”
Bill Igoe“I put the data here.”
We prepared a three-series JSON file with monthly data-center spending, total office spending and the derived other-office series. Bill hosted it at rainbowstats.com/data/census-data-center-office.json. The RainbowStats plugin successfully loaded each named series with READ_JSON_SERIES. There are 151 monthly observations per series from January 2014 to July 2026. Other office equals Census Office minus Census Data center; both remain nominal, seasonally adjusted annual rates.
6. See the sector divergence and test the break
Bill Igoe“I found this chart showing substitution very helpful. ... Binary_split does help in spotting regime shifts.”
The chart shows data-center construction climbing while other office construction declines. Quarterly averages move from $17.2 billion versus $71.3 billion in 2023 Q1 to $66.1 billion versus $49.1 billion in 2026 Q2. July 2026 data centers reached 61% of private office construction. BINARY_SPLIT(dc,other) finds a final monthly regime beginning near August 2022 with a negative slope; the apparent substitution is an accounting shift in construction mix, not proof that any one conventional building was displaced. A second split on monthly relative growth and three-month-lagged rates has weak within-regime fit.
7. Separate investment resilience from inflation
Both construction series are converted to quarterly means, approximately deflated with a common construction PPI and transformed to log growth. The policy variable is the prior quarter's change in the quarterly average effective funds rate.
| Result | Coefficient | Conventional t | Observations |
|---|---|---|---|
| Full-sample data-center construction growth | +2.45 | 0.72 | 49 quarters |
| Full-sample other-office growth | −0.40 | −0.25 | 49 quarters |
| Recent data-center minus other-office growth, 2023 onward | +5.60 | 1.83 | 14 quarters |
| Data-center growth in the PCE inflation model | −0.0011 | −0.15 | 49 quarters |
| Data-center growth in the consumer electricity inflation model | −0.015 | −0.82 | 49 quarters |
Price models also condition on prior rate changes and commercial-electricity producer-price growth. The price coefficients do not detect national pass-through in this specification. These are descriptive regressions, without an identified exogenous monetary shock, local grid prices, or a test of delayed price responses.
8. Turn the analysis into a public script
Bill Igoe“Showcase the use of the rainbowstats plugin.”
Bill Igoe“I did make two corrections.”
The presentation combines TEXT_SLIDE, BULLET_SLIDE, Census charts, BINARY_SPLIT, regressions and source slides. Bill's two corrections remove SET_NAME wrappers from the two initial READ_JSON_SERIES assignments. The published RainbowStats version is the one used for the article.
Where we landed
Some investment can remain strong through tightening, and the data-center boom makes that fact visible. This investigation does not demonstrate that higher rates cause higher overall inflation or that the September 2026 increase was a mistake. Data end before the September decision. Identifying the proposed mechanism requires regional electricity prices and loads, financing exposure, other-sector demand controls, and an exogenous monetary-policy shock.