How much will US data center construction spending be in October 2026?

Change carried forward from September

The base rate is the prior month adjusted by the rate of change. Five ways of computing it were tested month by month against first prints; the last month's change missed least and had the smallest worst miss, so it is the one used. Adding the change in the change made the misses larger.

Adjustment for the current state

−$1.5B. The Dodge index's October decline and Microsoft's pullback each add marginal downward pressure, together −$1B to −$2B. The early-November tech stock sell-off came after October's work was done, so it is left out. I rely most heavily on the base rate, and the marginal adjustment should be exactly that, marginal. The signals on the Projects page are better suited to flag a major regime change: a major reversal, or a big acceleration given large news or massive outlook changes.

Standard deviation

$3.0B. The base rate's first-print misses from March to September, leaving out May (the yearly revision), have a standard deviation of 2.2%, or $2.4B at October's level. Widened by a quarter, to $3.0B, because six months is a small sample. Census's revised figures miss by more, but mostly from revisions, which a first print can't contain.

Skewness

+0.3, slightly to the upside, given the aggressive, rapid build-out and the lack of negative news to the contrary. Too few months to measure it.

Kurtosis

4, slightly fatter tails than a normal distribution (3). Too few months to measure it.

• The outcome follows a sinh-arcsinh distribution with the input mean, standard deviation, skewness and kurtosis.⁷ Some pairs of high skewness and low kurtosis cannot be reached; the readout above the chart shows the values used.
• October follows the same rule as the months before it.
• The December 1 release carries no yearly revision. Census makes that each July.⁶

• Few first prints. The backtest has seven months, one of them distorted by the yearly revision.
• Most of the figure is estimated, not reported. In 2025, 78% of private nonresidential dollars were filled in by Census for projects that had not reported. No rate is published for data centers alone.³
• No published error for the data center line.

• First-hand reports of a large project starting, pausing or finishing.
• A change to Census's methods or to the release date.

• How much of the figure comes from a few very large projects?
• Can project-level tracking predict which way the base rate will miss?
• Why has Census raised every recent month after first publishing it?

2026-11-10

First draft.

Suggest a change on GitHub, or copy the repository and send your version (how to contribute). The same calculation as a short script: strike_probabilities.py.