Power Construction Put in Place: The Monthly Number Comes From an Annual Forecast

On September 1, 2026, the Census Bureau put private construction spending at a seasonally adjusted annual rate of $1,614.2 billion for July, 0.5 percent below the revised June estimate of $1,622.9 billion. Total construction came in at $2,157.6 billion, private nonresidential at $755.2 billion. Inside that table sits the line that has become the most-cited monthly proxy for grid investment in the United States: private power construction, at $161,541 million at an annual rate in July, 0.5 percent above June and 6.5 percent above July 2025.

That number gets read as a measurement of July. For a large part of the line, the agency's own documentation says it is something else. The Census Bureau methodology page for Construction Spending states that for the electric, gas, railroad, oil and TV cable components, "construction put-in-place estimates are based on annual capital expenditure reports compiled by federal regulatory agencies and private organizations." The same page adds that "preliminary monthly estimates for TV cable, electric, and gas construction are based on annual forecasts from S&P Global, the Edison Electric Institute, and the American Gas Association; and oil estimates are projections from the latest final year of Federal Energy Regulatory Commission data."

This is a definitional property of the series, not a defect in it. The Census Bureau publishes the method in plain language and has done so for years. What follows is an attempt to size the consequence: how far the preliminary path has historically sat from the eventual one, and what that implies for anyone using monthly construction spending to track the electricity buildout.

One Published Line, Two Collection Routes

The power category is not a single measurement process. The definitions page for the survey lists three subcategories under Power: Electric, Gas and Oil. The Census Bureau definitions for Construction Spending describe Electric as including "power plants (nuclear, oil, gas, coal, wood), nuclear reactors, hydroelectric plants, dry-waste generation, thermal, wind and solar energy facilities."

The category name is already broader than the use it is put to. Power is not a grid line. Alongside Electric, the same page carries Gas, defined as "buildings and structures for the distribution, transmission, gathering, and storage of natural gas," and Oil, defined as "buildings and structures for the distribution, transmission, gathering, and storage of crude oil." A gathering system and a substation land inside the same Power total. The methodology page notes that among the utility components "monthly estimates are published only for communication and electric," so an electric figure does exist at monthly frequency, but the Power number carried in the headline table is the combined one.

Cutting across those subcategories is a second split that matters more. The methodology page states that "expenditures made by nonregulated utilities are gathered in the same method as private nonresidential construction." That method is a project sample: "privately-owned nonresidential, state and local, and federal projects are selected from lists compiled by Dodge Data & Analytics (DDA), (and supplemented with a small sample of projects in non-permit issuing areas)."

So one published figure carries two different collection regimes. Regulated utility spending enters through annual reports and trade-association forecasts. Nonregulated spending enters through the same monthly project machinery as an office tower or a warehouse. The release does not decompose the line, and the methodology does not publish the weights. A month-over-month move can originate in either regime, and the reader cannot tell which from the table.

When the Forecast Gets Replaced

A forecast-based month is not permanent. It is replaced on a schedule, and the schedule is published. The Census Bureau survey description for Construction Spending sets out three distinct mechanisms.

The first is the rolling window: "two months of data are revised along with the release of each month's preliminary estimates." The second is seasonal: "with the release of preliminary May data on the Construction Spending release, seasonally adjusted annual rates for the previous 88 months are also revised to reflect updated seasonal factors." The third is the benchmark: "annual estimates of value-in-place estimates for the most recent two years are finalized with the release of data for May of the following year."

The third mechanism is where the utility components change character. The methodology page states that "the Federal, Utilities, and Railroad estimates are benchmarked or derived from outside sources," that "some of these sources are unable to provide estimates until a year or two after the reference period," and that "once the source data is available the series are revised during the following annual revision period." A reference month therefore has to survive two annual revision cycles before its annual level stops moving on this account.

The most recent cycle is dated. The Census Bureau release for May 2026 (CB26-112), published July 1, 2026, states: "with this release, unadjusted data have been revised back to January 2024 and seasonally adjusted data back to January 2019." Everything printed since then, including the July figure discussed above, sits inside the pre-benchmark stage.

One further line from the methodology page belongs beside the calendar, because it bounds what the survey's published uncertainty actually covers. The page notes that "the standard error and CV only measure sampling variability" and that "they do not measure any systematic biases in the estimates." A revision that arrives because a regulatory filing finally exists is not a sampling event. Nothing in the survey's error statistics is designed to describe it, which is why the size of a benchmark move has to be measured from the vintages rather than inferred from the precision printed beside the monthly change.

What Three Vintages of December 2024 Show

The schedule tells you when a number can change. It does not tell you by how much. For that, the honest method is to take one reference month and read it in three published vintages.

December 2024 is the useful case, because as of September 2026 it has passed two annual revision cycles and sits in box D. The Census Bureau release for December 2024, published February 3, 2025, printed private power at $129,454 million at an annual rate. The December 2025 release of February 27, 2026, which is after the 2025 annual revision and before the 2026 one, carried December 2024 at $135,275 million. The current figure, published September 1, 2026 and available as FRED series PRPWRCONS from the St. Louis Fed, is $148,995 million.

The first step added 4.5 percent. The second added 10.1 percent. Together they moved the month 15.1 percent above its original print. The revision was back-loaded, which is what the methodology's language about source data arriving "a year or two after the reference period" would lead a reader to expect.

The size only means something against a base. Here is the same month across four private lines, first print against current.

December 2024, private, annual rate	First print, Feb 3 2025	Current, Sep 1 2026	Change
Power	$129,454M	$148,995M	+15.1%
Office	$86,547M	$99,316M	+14.8%
Manufacturing	$236,093M	$240,906M	+2.0%
Total private construction	$1,688,531M	$1,733,436M	+2.7%

The Office Line Moved Just as Far

The table refuses the tidy story. If the forecast-based construction of the power line were the whole explanation, power would stand alone. It does not. Office moved 14.8 percent over the same interval, within four tenths of a percentage point of power, and office is collected through the ordinary project sample rather than from utility forecasts. Manufacturing moved 2.0 percent and the private total moved 2.7 percent.

Two readings survive that. The first is that power and office each revised roughly five and a half times as far as the headline in this episode, while manufacturing revised less than the headline did. Whatever produced that spread, it is not one collection method, because the two lines that moved were collected differently. The second is that the documented reason differs by line even when the magnitude does not. For power, the methodology names the mechanism: late-arriving regulatory and trade-association source data. For office, the methodology names a sample frame drawn from a commercial project list, and neither the release nor the methodology attributes the December 2024 move to any particular cause. Stating a mechanism for one line does not license inventing one for the other.

The office comparison carries a second edge. Under the definitions page, Office contains three subcategories, and one of them is Data center, defined as "buildings that contain the hardware needed for storing, processing, and transmitting digital information." So the two lines an analyst would pair to track the artificial-intelligence buildout against the grid that has to serve it are precisely the two that moved about 15 percent in this episode. In the July 2026 data, office ran 21.3 percent above July 2025 while power ran 6.5 percent above. That gap is the headline comparison. It is also a comparison between two provisional numbers with different revision futures, as the earlier discussion of how benchmark revisions concentrate on a single month showed in a different series.

What Would Invalidate This

The frame here is narrow, and several findings would cut against it.

One episode is not a base rate. December 2024 is a single reference month observed in three vintages. If the same exercise on months from 2022 and 2023 showed category revisions of one or two percent rather than fifteen, the correct conclusion would be that the 2024 vintages captured an unusual benchmark, not a standing property of the series.

The mechanism may not be doing the work. Because office moved nearly as far as power without any forecast-based input, the late-utility-data explanation is at most partial even for power. If a future annual revision moved power sharply while office, manufacturing and the total held still, that would support the mechanism. The December 2024 evidence does not separate them.

The blend may have shifted. The methodology puts nonregulated utility spending in the ordinary project sample. If nonregulated developers now account for most new capacity, the forecast-derived share of the power line may be smaller than the methodology's prominence implies, and the line may be closer to a measurement than this piece treats it as being. The Census Bureau does not publish that split, so the point stays open rather than resolved.

Finally, revisions have direction only in hindsight. Every change described above was upward. Nothing in the documented method requires that, and a benchmark that arrives during a capital-spending downturn could as easily cut a preliminary path as lift it. Treating the revision as a reliable upward bias would be a straightforward misreading of four observations.

Concrete Framework

For a reader who has to use this series anyway, the following checks are cheap and repeatable.

Locate the month in the revision calendar first. Before quoting a category figure, establish whether it sits in the two-month rolling window, outside it but pre-benchmark, or past one or two annual revisions. The branching diagram above covers every case.

Separate the headline from the detail. The private total and the category lines have behaved as different objects. Applying a headline-sized confidence to a category line is the error the December 2024 table exposes.

Quote the vintage with the number. A power figure without a print date is not reproducible, because the same reference month has carried at least three different published values.

Do not compare across collection regimes without saying so. An office-versus-power growth comparison pairs a project-sample estimate with a partly forecast-derived one. The comparison is still usable; it just needs the caveat attached, in the same way that a ratio whose denominator is re-estimated every year needs its denominator named.

Prefer annual figures for level questions. Monthly detail answers direction questions poorly at this revision scale. For questions about how much capacity spending has actually occurred, the benchmarked annual estimates are the appropriate object.

What to Watch Next Week

The first thing to watch is that the relevant release is not next week. The July 2026 report states that "the August 2026 Monthly Construction Spending Report is scheduled for release on October 1, 2026." Any power or office figure quoted between now and then is the September 1 vintage, whatever the framing around it suggests.

That leaves three checks available in the interval. The first is the pre-benchmark stack: May, June and July 2026 are still inside the two-month rolling window, the earlier months of the year sit outside it with no finalized annual level behind them, and so the entire year-to-date path for power and office is provisional in the same way December 2024 was in early 2025. The second is the utility earnings and regulatory filings that will eventually feed the benchmark, since the methodology makes those filings, not the monthly release, the ultimate source for the regulated components. The third is the gap between the two proxy lines, currently 21.3 percent against 6.5 percent year over year, which is the number most likely to be cited as evidence that data center construction is outrunning grid investment.

The narrower claim is the defensible one. Monthly construction spending is a real and carefully documented series, and the Census Bureau states exactly how each component is built. It does not support a month-to-month reading of grid capital expenditure, and its own documentation is where that limit is written down. The earlier look at where the EIA electricity tables stop reached the same boundary from the demand side. Both times the constraint came from the release notes, not from the data.