Two Markets, One Industry: What Happens to Your Backlog When the Data Center Music Stops
ABC’s July 2026 backlog data splits the industry into two economies. The 12% of contractors with data center work under contract are sitting on 11.4 months of backlog. The 88% without are on 7.5. On the Bricks & Bytes round table, Dustin DeVan warned that some specialty trades now draw more than 90% of revenue from data centers, and that the smart ones are already tracking that ratio at board level. The number to know is not your backlog. It is what share of it has a hyperscaler behind it.
Owen brought a stat to the round table that stopped the conversation for a second. Last month, construction was one of the only goods-producing industries in the US where jobs actually grew. Not because the industry is healthy. Because of data centers.
Then he put the backlog split on the table, and it landed harder. Contractors doing data center work have a pipeline that runs nearly four months longer than everyone else’s. Same country, same labor pool, same material prices, wildly different business.
Dustin DeVan, who sits on the board of a large subcontractor, has been watching this from inside a boardroom. His read was blunt: construction companies are now growing like technology companies, which nobody thought was possible, and the well-run ones are quietly asking themselves what happens when it stops. Owen asked the question nobody answered all episode. What happens after the boom?
The headline number is hiding a hole
Backlog fell to its lowest point since January, and the boom is what stopped it looking worse
Associated Builders and Contractors reported on 11 August that its Construction Backlog Indicator dropped to 8.0 months in July 2026, down 0.8 months on both the month before and the year before. ABC chief economist Anirban Basu did not soften it. He said backlog fell sharply, that it is the weakest reading since January, and that the data center boom masks how thin things are underneath, because no other segment has momentum. The 88% of ABC contractors with no data center work averaged 7.5 months. The 12% with it averaged 11.4.
Strip the hyperscalers out and you get a picture the headline number does not show. This is the same divergence Basu flagged back in April, when he described Silicon Valley and Wall Street as fine and Main Street as rocky. We covered that split in detail in Data Centers Are Keeping US Construction Afloat, and four months on the gap has widened rather than closed.
Martin pushed back on the doom framing with a useful corrective on rates. His point was that everyone is anchored to the wrong baseline, comparing today’s cost of borrowing to the near-zero years around 2020 rather than to the multi-decade norm. That is worth holding onto. Rates are not the anomaly here. The concentration is.
Concentration risk has a number, and most firms have not calculated theirs
Some specialty trades are now single-client businesses that have not admitted it
This was the sharpest practical takeaway of the episode. Dustin said the reasonable contractors, including the one he sits on the board of, are monitoring their total book and the percentage attributed to data centers. Then he named the tail risk: there are companies, particularly on the specialty trade side, where data centers account for more than 90% of revenue. There are software companies in the same position, with subscription growth driven almost entirely by the boom.
That is not a diversified contractor with a good year. That is a business whose survival depends on one client category’s capital allocation committee. Dustin was careful to say a pullback is unlikely right now. He was equally careful to say it is possible, and that the trigger would be financing rather than demand.
You have seen construction companies grow like technology companies, which was never thought possible before.Dustin DeVan, founder of Ediphi, on the Bricks & Bytes round table
The dislocation runs deeper than one sector. Dustin described markets he has never seen behave this way, where the Bay Area is healthy across the board and other metros have almost nothing moving. Martin read the same pattern as a warning sign. When a rising tide stops lifting everyone evenly, that unevenness is the signal.
The financing math nobody wants on the record
If chips die in two years instead of four, the capital requirement doubles
Dustin walked through the piece of this that keeps analysts up at night. The financial models underwriting these projects depreciate GPUs over a set life, and he understood the working assumption to be around four years. Patric Hellermann’s instinct was that the physical life is shorter than five. If the real economic life is closer to two, the whole capital structure changes.
Dustin’s version of the arithmetic was characteristically direct. If you planned to spend one hundred billion dollars on chips every four years and you end up spending it every two, you have just doubled what you need to finance. He called that a big boo boo, which is one way of putting it.
This is not a fringe view. Michael Burry made the same accounting argument publicly in November 2025, claiming hyperscalers use five to six year schedules when real economic life sits closer to two or three, given Nvidia’s product cycle. We broke the numbers down in Big Tech Promised $650 Billion in Data Centers, Most of It Isn’t Being Built. Microsoft runs six-year schedules, Meta 5.5, Amazon shortened from six to five. Whether Burry is right on timing is arguable. Whether the question matters to your covenant strength is not.
The construction briefing that runs the numbers before the narrative
Deals, data, and takes you will not get from a vendor blog. Read weekly by construction execs, ConTech founders, and AEC investors.
Join 3000+ ReadersThe bigger worry is macro, and it is uncomfortable
Patric’s Japan comparison, and why the Fed has no clean lever
Patric offered the contrarian read. His concern is that the US is closer than people think to a 1989 Japan situation, with asset overvaluation running hot in private and infrastructure markets rather than just equities. His argument is about what sits underneath: a lot of notional value that somebody owes, without the cash in the system to service it if the assets turn out to be worth less than the debt raised against them.
He added a second layer that is specific to this build cycle. The assets going up are designed to substitute for a taxpaying workforce, which erodes the income tax base at the same time it accelerates. His suggested fix, a tax on installed robots and possibly on AI tokens, is deliberately provocative. He was clear this is a risk scenario rather than a forecast, and Martin’s counter was that the dollar’s reserve status gives the US room Japan never had.
Dustin’s framing of the policy trap holds either way. Ordinarily the Fed would cut to reaccelerate a stagnant economy, but cutting lets hyperscalers borrow more cheaply and pours fuel on the one segment that is already inflationary. Raising hurts everything else, which is far more rate-sensitive than the data center build. Blunt instruments, no targeting.
What to actually do about it this quarter
Three questions that separate a good year from an exposed balance sheet
| Position | Backlog (ABC, July 2026) | Primary risk | What to watch |
|---|---|---|---|
| Heavy data center exposure (over 50% of book) | 11.4 months average for the segment | Single-category demand shock if financing tightens | Hyperscaler capex guidance, GPU depreciation schedules, lease commitments on unbuilt sites |
| Mixed book with some data center work | Between the two segment averages | Margin compression as boom-driven labor costs carry into non-boom work | Trade wage escalation, subcontractor availability outside the boom lane |
| No data center work | 7.5 months average | Weak demand across every other segment | Regional dislocation, residential affordability, non-residential spending |
| Specialty trade, over 90% data center revenue | Longest pipeline in the industry | Effectively a single-client business | Diversification runway, contract termination provisions, bonding capacity |
On 11 August 2026, Associated Builders and Contractors reported that its Construction Backlog Indicator fell to 8.0 months in July, based on a member survey run from 20 July to 4 August. That is down 0.8 months from both the previous month and the previous year, and the lowest reading since January. Chief economist Anirban Basu attributed the relative strength that remains to data center construction, noting that the 88% of ABC contractors without data center work averaged 7.5 months of backlog against 11.4 months for the 12% with it. (Source)
Yes, and the share of contractors in the fortunate group has been shrinking. In June 2026, ABC reported 13% of members under contract on data centers with 11.0 months of backlog against 8.5 months for the 87% without, a gap of 2.5 months. By July the gap had opened to 3.9 months while the share with data center work slipped to 12%. Basu has also noted the split runs along company size, with far more contractors above $100 million in revenue holding data center work than below it. (Source)
That figure comes from Dustin DeVan speaking from board-level experience on the Bricks & Bytes round table, and describes companies he has visibility into rather than a published industry statistic. Treat it as an informed practitioner observation. The published ABC data supports the broader pattern of concentration, and Basu’s commentary confirms the split falls heavily along company size and sector exposure. If you want a hard number for your own business, the concentration ratio is one you can calculate internally in an afternoon. (Source)
It affects the financial strength of the tenant behind your project. If hyperscalers are depreciating AI hardware over five or six years when the real economic life is closer to two or three, their reported earnings look better than the underlying cash position, and a wave of writedowns could land in 2027 and 2028. That is roughly when current buildouts complete. A tenant absorbing an impairment cycle is a tenant less willing to sign for additional capacity, which shows up in your pipeline rather than theirs. (Source)
Not outright. Moody’s flagged the spending as unprecedented in a July 2026 note covering six firms, projecting $785 billion this year and close to $1 trillion next, with $1.2 trillion of signed lease commitments and more than $820 billion of that on buildings that do not exist yet. The agency was explicit that a downgrade is not close and framed this as a yellow flag. Separately, Sightline Climate’s tracking suggests a meaningful share of announced 2026 capacity may slip or cancel. Demand is real. The timing and the ownership of the risk are the open questions. (Source)
Because interest rates apply to everything at once. As Dustin framed it on the show, the economy outside data centers is stagnant enough that the Fed would normally cut, but cutting lets hyperscalers borrow more cheaply and accelerates the one inflationary segment. Raising would hit the rest of the market, which is far more rate-sensitive than a build cycle funded by corporate balance sheets. Martin raised the idea of differentiated borrowing costs by sector. Patric’s response was that this belongs to fiscal policy and legislation, using instruments like targeted subsidies or tax relief, not to the Fed’s mandate. Dustin agreed with that distinction on the show.
Related Articles
Data Centers Are Keeping US Construction Afloat, But the Rest of the Market Is Struggling
Big Tech Promised $650 Billion in Data Centers. Most of It Isn’t Being Built
Your Data Centre Client Is Changing (and We Don’t Understand Who)
ABC Construction Backlog Indicator, July 2026
ABC Construction Backlog Indicator, June 2026
ENR: ABC Backlog Indicator Inches Up to Ten-Month High
Associated Builders and Contractors, economic releases
Analysis of the GPU depreciation argument
