Last Friday, the Census Bureau confirmed what most of the industry already felt in its gut: single-family housing starts fell for the third consecutive month, permits dropped to their lowest level since August 2025, and the NAHB/Wells Fargo confidence index sat at 34 for July, extending what is now the longest sustained stretch of builder pessimism since 2012 to 15 consecutive months below 40. None of this was supposed to happen in a market with a structural housing shortage.
And yet here we are, with 496,000 new single-family homes sitting unsold across the country, a pile not seen since late 2007. Thirty-seven percent of builders have cut prices by an average of 6%, while 63% offer incentives like rate buydowns and closing cost credits, marking the 16th straight month that more than 60% of the industry has been paying buyers to take what it built. I have been tracking construction cycles for two decades, and the combination of a genuine national shortage with late-2007 inventory levels is something new. It means firms built the right number of homes in the wrong places at the wrong prices, and nobody caught the mismatch until carrying costs started eating margins alive.
What makes this cycle different from 2007 is that builders had tools available to forecast demand at the zip-code level before they ever broke ground, machine learning models that ingest permit velocity, employment migration data, mortgage application volume, and local inventory to predict absorption rates by submarket, commercially available for years now, not in prototype, not in pilot, but on the shelf waiting to be licensed. Instead, the industry's AI budget went somewhere else entirely.
Where the Money Actually Went
In July 2025, the NAHB surveyed single-family homebuilders about artificial intelligence adoption. Forty-nine percent said they use AI in some capacity, which sounds like progress until you look at what "some capacity" actually means. Twenty percent generate advertising and marketing materials with it. Eleven percent use it for market analysis and project planning. Fewer than 5% apply AI to any of ten other business functions, including scheduling, estimating, design, safety monitoring, or equipment operation, and just 1% use it to run construction equipment. Safety monitoring lands below half a percent.
When builders were asked about the likelihood of adopting AI in the next two years on a 1-to-5 scale, marketing scored 3.6, market analysis scored 3.0, equipment operation scored 1.7, and interacting with building and planning departments scored 1.9. Put differently, builders are twice as likely to adopt AI for writing ad copy as they are for predicting whether homes they are building will actually sell.
A more recent BuildOps survey of 600-plus contractors across both the United States and Canada found a higher headline number: 78% using or testing AI. But 45% of contractor AI budgets go to outside consulting, not to tools or subscriptions, just to consultants who teach people how to use what they already bought. Training is the top barrier. Money never was.
$7.8 Billion in Missed Signals
I ran the numbers on what that misallocation has cost, and they were grim enough that I checked them twice.
Start with price cuts: 37% of roughly 496,000 unsold homes means about 183,500 units marked down by an average of 6% off a median new-home price near $417,000, which works out to roughly $25,000 per home in evaporated revenue, or $4.6 billion across the pool. Add incentives: rate buydowns, appliance upgrades, and closing cost credits on 63% of inventory at a conservative $10,000 per home produce another $3.1 billion.
Combined, that is $7.7 billion in losses absorbed by an industry of roughly 45,000 homebuilding firms. Not losses from construction defects, not from supply chain disruptions, not from labor shortages. Losses from a collective inability to match what was built to where people were actually buying.
Now consider the cost of the tools designed to prevent exactly this outcome. AI-powered demand intelligence platforms, the kind offered by companies like Zonda, John Burns Research, and a growing cohort of startups with names nobody outside construction would recognize, run between $50,000 and $200,000 a year depending on firm size and market coverage. At a $100,000 average across 45,000 firms, equipping every builder in the country costs about $4.5 billion annually. At the expensive end, $200,000 per firm, it reaches $9 billion.
So the industry's aggregate losses from inventory misallocation roughly equal the aggregate cost of the tools purpose-built to prevent inventory misallocation, except the losses recur every year and compound as carrying costs pile up, while the tools are a subscription that gets smarter with each cycle of data.
What the Largest Builders Are Doing Instead
Pro Builder's 2026 survey of the 200 largest homebuilders reveals a spending pattern that mirrors these findings at a bigger scale: 77% invested in digital sales and marketing tools, 75% in business software integration, but only 54% in digital design applications and 53% in construction documentation. These firms closed 509,100 units in 2025, grew their collective market share to 34.2%, and command the budgets to invest in whatever they choose. What they chose, overwhelmingly, was the selling side of the business rather than the building-decision side.
A better virtual tour does not fix a $25,000 price cut on a home that should never have been started in that submarket eight months ago, and no amount of AI-generated staging renders will change the absorption rate in a zip code where three competing developments launched simultaneously because all three builders were reading the same six-month-old permit data by hand instead of letting a model flag the collision in advance.
What Demand Intelligence Actually Looks Like
A demand forecasting model does not replace a land acquisition manager's instinct, and anyone selling one as a replacement deserves the suspicion they get. What these tools do is surface the signals that manager would otherwise need three months of drive-bys, broker conversations, and county records searches to gather, and they surface them before a builder commits capital to a site.
Inputs are not exotic: permit velocity at the census-tract level, employment-weighted migration data from IRS returns and postal change-of-address filings, mortgage pre-approval volume by zip code, school enrollment trends, competitive inventory within a 15-minute drive time, and days on market for recent closings in the price band. Cross-referenced, these signals produce an absorption-rate estimate telling you how many homes of a given type and price point a submarket can absorb per month without triggering the price-cut-and-incentive spiral currently devouring $7.7 billion industry-wide.
If a model says a submarket absorbs 12 homes per month and a builder is planning 30 starts, that is a conversation worth having before the concrete trucks arrive. At most firms today, that conversation never starts because the tool that would initiate it was never purchased. Marketing got the AI budget instead.
Why the Counterargument Is Half Right
Skeptics will point out, fairly, that no model predicted the 60-basis-point mortgage rate spike triggered by renewed US-Iran hostilities at the end of February. Geopolitical shocks are exogenous by definition, and demanding that a demand model foresee military conflict is an unfair standard.
But inventory was already climbing before February, and builder sentiment had been below 40 for more than a year when rates spiked. A model's job is not to predict wars. Its job is to look at the data in October 2025 and say: your submarket is absorbing 8 homes per month, you have 47 sitting finished, stop starting new ones until the pipeline clears. That signal was hiding in plain sight, inside publicly available Census and NAHB data, for anyone with the tool or the discipline to look. Most builders had neither, because they spent their technology budget on the customer-facing half of the business while the operational half kept running on instinct, spreadsheets, and the same "drive the market and see what's selling" methodology that produced the mid-2000s wreckage now being used as a historical comparison for the current inventory pile.
What Washington Just Did, and What It Cannot Fix
On July 12, the 21st Century ROAD to Housing Act became law without a presidential signature, after passing 358-32 in the House and 85-5 in the Senate. It brings pattern-book grants for pre-approved designs, NEPA environmental review waivers for infill construction, chassis rule removal for manufactured housing, and a cap on corporate landlords owning more than 350 single-family homes. Four days later, California signed AB 179, projecting $60,000 to $70,000 per unit in cost reductions through impact fee reform and streamlined financing for affordable developments.
Every one of those measures is supply-side. They make building cheaper and permitting faster. But none of them address the demand-side mismatch that produced 496,000 unsold homes, because the problem was never that homes were too expensive to build. It was that builders could not tell, before committing capital, whether a specific submarket would absorb what they were planning to build at the price they needed to charge. Supply-side reform and demand-side intelligence are different prescriptions for different diseases, and Washington just filled one of them.
For Builders and Buyers
If you are a general contractor or production builder doing $2 million to $50 million in annual revenue, here is the reallocation the data supports: take 30% of whatever you are currently spending on AI-powered marketing tools and redirect it to a demand intelligence subscription. Entry-level cost runs $50,000 a year for a platform like Zonda's builder analytics or John Burns' regional forecasting package. If that exceeds your overhead, a low-cost approximation exists. Census publishes monthly new residential construction data, including authorized-but-not-started permits by state and metro area, free. Freddie Mac publishes mortgage rates weekly. Combine those with local MLS days-on-market data and you can approximate, in a spreadsheet, what the expensive platforms model automatically for your specific market.
If you are a homebuyer looking at new construction right now, understand what 37% and 63% mean for your negotiating position: a builder who is discounting built something the market did not ask for at that price, and the discount is a floor, not a ceiling. Ask for both the incentive and the price cut. If a completed home has been sitting unsold for more than 90 days, carrying costs are eating margin, and your leverage increases with every month it stays empty.
Limitations
This analysis uses national averages and assumes roughly uniform distribution of price cuts and incentives across the unsold inventory pool, when in reality those losses concentrate heavily in Sun Belt markets like Texas, Florida, and the Mountain West where overbuilding has been most acute, while coastal markets with persistent undersupply experience a fundamentally different dynamic. NAHB adoption survey data dates from July 2025 and may not capture shifts in the past 12 months, though nothing in Pro Builder's 2026 Top 200 survey suggests a meaningful reallocation of technology spending toward demand-side tools. BuildOps surveyed contractors broadly, not homebuilders specifically, and adoption patterns may diverge between the two groups.
What this analysis cannot determine is whether any specific demand model would have prevented any specific loss, because absorption models are probabilistic, not deterministic. A builder who ignores a model's warning and succeeds has not disproved it, just as one who follows it and takes a loss has not proven it useless. Across 45,000 firms and 496,000 unsold homes, better demand intelligence at the go/no-go decision point would almost certainly have reduced the $7.8 billion figure. By how much remains unknowable without firm-level data the industry does not publish.