Three flat years of housing starts is a result, not a pause

Housing starts have moved by six hundredths of a million units across three annual readings: 1.42 million in 2023, 1.37 million in 2024 and 1.36 million in 2025. I would argue that a line that flat should be treated as the current state of American housing production rather than as a dip on the way back to something. The industry and its supporters have spent that entire stretch describing it as a pause.

The rest of the window makes the point sharper. Starts were 1.39 million in 2020, peaked at 1.6 million in 2021, and have fallen in every year since. The 2025 figure is below the level at which the window opened. Whatever has been attempted in housing policy over these five years, the count of units on which construction began is lower at the end than at the beginning.

The dollar total is not a count

Set against that, the figure that actually circulates. Construction value put in place, which covers residential and non-residential work together, rose from $1.65 trillion in 2021 to $1.9 trillion in 2022, $2.07 trillion in 2023 and $2.24 trillion in 2024. In 2025 it was $2.24 trillion again: the most recent observation shows no change at all.

That series is in current dollars, and the window it covers is one in which prices rose hard. The producer price index for all commodities went from 194.4 in 2020 to a peak of 264.5 in 2022, and stood at 260.3 in 2025. I am not going to claim that the whole increase in construction value is price, because these files cannot separate the two. What they can establish is the direction of each: the dollar total went up, the unit count went down, and only one of the two is a measure of how much housing got built.

The objection that rates did this

The obvious reply is that starts are a rate story and nothing else. That has real force here. The peak of 1.6 million units came in 2021, the same year the federal funds effective rate averaged 0.08 percent. Housing is the most rate-sensitive thing the American economy builds, and anyone arguing that five years of policy failed has to reckon with the possibility that the policy was simply swamped.

I accept the mechanism and still think the conclusion is wrong, for two reasons. The first is that the relief being waited for has not arrived: the ten-year Treasury averaged 4.21 percent in 2024 and 4.29 percent in 2025, so the long end against which mortgages and construction loans are priced went up, not down, in the most recent year. Treating the flat line as temporary is therefore a forecast, and this brief does not make forecasts.

The second is what the objection concedes. If starts are principally a function of the rate environment, then the zoning reforms, subsidy programmes and permitting changes offered as housing supply policy are being asked to overcome a force they cannot reach. That may be true. It is not an argument for continuing to present them as the fix, and it is certainly not an argument for measuring them by a dollar total that rises with prices.

There is a narrower objection worth answering, which is that a flow is not a stock. Even a flat 1.36 million units a year is 1.36 million units a year, and the housing that results accumulates whether or not the line slopes upward. That is true, and it is the correct way to think about the physical outcome. It is not a defence of the policy argument, because the claim being made since 2021 has not been that the country is still building. It has been that the country is building more. The series says the rate of addition settled three years ago and has stayed there, and a settled rate is exactly the thing a supply programme exists to move.

What to count instead

Three changes would make the next five years of this argument more honest.

1. Quote units, not spending, whenever the claim is about supply. Value put in place answers a different question and is not a proxy for one.
2. Say which year is the benchmark. Against 2021 the current level is down about 15 percent; against 2020 it is down 0.03 million units; against 2024 it is down 0.01. The comparison chosen is usually doing the arguing.
3. State the rate assumption behind any recovery claim. If a programme's case depends on the long end falling, the case should say so in the same paragraph.

A count that has not moved in three years is information. It is less useful than a rebound and less dramatic than a collapse, which is probably why it gets described as neither. But a flat line held for three observations is the most reliable thing in this series, and any programme being designed now should start from it rather than from the year it would prefer.