Q3 2026
The Q3 Fern Quarterly opens on a contradiction. Through the second quarter the broad market decided the recession scare was over: the yield curve un-inverted, recession-probability models fell back toward their long-run floor, and the bank stocks that investors treat as sentiment bellwethers rallied hard. Read the construction tape at the same moment and you see the opposite picture. Spending is falling year over year, housing starts just printed their weakest month since 2020, and the architecture billings that lead nonresidential work by roughly a year are still in contraction.
Both readings are true, because they are about different economies. This issue is about the gap between them, and about the single force that explains most of it: an AI-and-power construction boom running at record highs, statistically masking an otherwise broad private-building recession. Section one measures the divergence. Section two follows the boom into the one place it is most legible to an estimator, the order books and lead times for the electrical equipment that a data center or a generating plant cannot open without.
This is the opening cut of the Q3 issue, anchored to the May 2026 prints that closed Q2. It will be re-anchored as the Q3 data lands. Every figure traces to a cited public source or a Fern computation; where a claim rests on estimator practice, the assumption is shown so you can substitute your own.
Section 01
The Bifurcation: Optimism Meets a Contracting Tape
Two economies, one quarter
By the usual financial tells, recession fear drained out of the market in Q2 2026. The 2-to-10-year Treasury spread un-inverted, the New York Fed's twelve-month recession model eased to roughly 15%, and high-yield credit spreads sat historically tight near 2.6 to 2.8%. Large-bank equities, the stocks investors treat as a proxy for how they feel about the economy, rallied well ahead of the index.
Construction spent the same quarter going the other way. As of the May 2026 print, total construction spending was running at a $2,210.2 billion annual rate, down 1.5% year over year, and housing starts had fallen to a 1,177 thousand annual rate, down 8.7% year over year and the weakest month since 2020. Private nonresidential spending fell for an eighth straight month. Neither series is a market mood; both are Fern's own tape.
The forward-looking gauges agree with the tape rather than the market. Builder sentiment (NAHB HMI) sat at 35 in June, its 14th straight month below 40, the longest such streak since 2011. The AIA Architecture Billings Index, which leads nonresidential construction spending by roughly a year, read 44.5 in May — below the 50 line that separates growth from contraction, with every region and sector declining. What leads construction is still pointing down.
Why the two economies can both be right
The gap is not a data error; it is a difference in what each signal measures. A bank-stock rally reads financial conditions and market sentiment. Construction reads the cost of money and the willingness to commit capital to a multi-year project, and on that score nothing eased: mortgage rates held near 6.5%, and the Fed held its policy rate at 3.50 to 3.75% in June and removed its easing bias, with markets pricing roughly 80% odds of no cuts at all in 2026. Rate-sensitive, project-driven demand does not turn on improved sentiment; it turns on rate relief that is not coming.
The broad economy itself is more ambiguous than the rally implied. Q4 2025 GDP grew just 0.5%, the Atlanta Fed's GDPNow estimate for Q2 2026 fell to 1.2% by July 1 from 3.3% three weeks earlier, June payrolls added a soft 57,000 jobs with the unemployment rate falling only because participation dropped to a four-year low, and consumer sentiment stayed near record lows. The honest read is that the broad economy stopped deteriorating rather than reaccelerated — and construction sits on the wrong side even of that.
The contraction is concentrated, not universal
Here is the part an estimator has to get right: "construction is contracting" is a statistical average of two opposite stories.
| Segment | Direction | Reading |
|---|---|---|
| Single-family residential | Contracting | spending ~-4% y/y; starts -8.7% y/y |
| Office | Contracting | -11.9% y/y; office CMBS delinquency 12.34%, an all-time high |
| Warehouse | Contracting | -8.5% y/y |
| CHIPS-era manufacturing | Contracting | -22% y/y, 14 straight monthly declines |
| Data centers | Booming | ~3.6x y/y, ~$50B annual rate |
| Power / electric | Booming | record, ~$158B annual rate |
| Transportation infrastructure (IIJA) | Booming | record, ~$209B annual rate |
The bust and the boom are close enough in size that the headline nets to a mild decline. The Dodge Momentum Index rose 33.8% year over year, but strip out data centers and the commercial component is up only 6.6%; ABC's backlog sits near a three-year high, but at 11.6 months for data-center contractors versus 8.6 for everyone else. The segment you build in, not the market's mood, determines which economy you are bidding into.
Fern's read
For a bid today, the practical translation is to distrust top-down optimism and price the segment. Residential, office, warehouse, and traditional commercial face soft demand, subcontractor hunger, and negotiating room that did not exist two years ago — but no rate relief to pull work forward. Data-center, power, and infrastructure work faces the opposite: record backlogs, stretched trades, and the long-lead equipment constraints that section two takes up directly. The one macro item to watch is whether Q2's real-economy softening — the 1.2% GDPNow, the 57,000-job payroll — is a durable turn or noise; it resolves with the Q2 GDP and July jobs prints, and Fern's Pulse module will carry it forward.
Methodology and limitations
Sources: Census C30 construction spending and New Residential Construction releases
(Fern's construction_spending_total and housing_starts series, matched to the
published figures); NAHB
and AIA ABI
for sentiment and the leading indicator; NY Fed,
Atlanta Fed GDPNow,
BLS, and the
Federal Reserve
for the broad-economy read; segment splits compiled from Census and
trade reporting.
Limitations: Census reclassified data centers out of the "office" category in April 2026, so a few segment year-over-year comparisons straddle a definitional change; the direction of each is robust, the exact level less so. The broad-economy verdict is genuinely two-sided — financial signals say the scare passed while hard activity data softened — and that tension, not a single number, is the finding.
Section 02
The Grid Buildout, Read Through the Order Book
Following the boom into the order book
Section one left the construction economy split in two, with data centers and power generation on the booming side. That boom is easy to assert and hard to price, because most of it is still announcements and interconnection requests. The place it becomes real — and legible to an estimator — is the order book for the electrical equipment a data center or a generating plant cannot open without. Demand you cannot verify becomes a backlog you can.
The anchor is the twenty-year power agreement structure now common between hyperscalers and generators: a large technology buyer contracting a decade or two of offtake against new generation built largely on a single turbine maker's equipment. These are not soft letters of intent. They underwrite multi-year equipment orders, and those orders show up in filed backlogs.
What the buyers demanded has become what the factories are booked to build
| Maker / item | Backlog or lead time | Source period |
|---|---|---|
| GE Vernova (turbines, total) | $163B backlog; gas sold out through 2029 | Q1 2026 |
| Siemens Energy | record EUR 154B backlog; book-to-bill 1.72 | Q2 FY2026 |
| Hitachi Energy (transformers) | ~$43B backlog; 30-40 month lead times | 2024-2026 |
| Eaton (electrical) | backlog +48% y/y; data-center orders ~+240% y/y | Q1 2026 |
| Large power transformers | 128-144 weeks; ~30% supply deficit | 2025-2026 |
| Large standby generators | 52-70 weeks; Caterpillar sold out through 2027 | 2026 |
Behind the backlogs, the demand signal is corroborated at the grid level: NERC's 2025 Long-Term Reliability Assessment raised its ten-year US peak-demand forecast by 224 gigawatts, up 69% over the prior year's forecast, attributing most of the jump to AI and data centers, and the EIA now describes 2026 to 2027 as the first multi-year run of rising US power demand since 2007. The maker filings put hard numbers under it: GE Vernova reported a $163 billion backlog with gas turbines sold out through 2029, Siemens Energy a record EUR 154 billion backlog, Hitachi Energy a roughly $43 billion backlog with transformer lead times of 30 to 40 months, and Eaton electrical backlog up 48% year over year with data-center orders up roughly 240%.
Why this holds even if the AI forecasts are wrong
The load forecasts driving the headlines are almost certainly inflated. When AEP Ohio was required to back interconnection requests with capacity payments, its data-center pipeline fell 57%, from about 30 to 13 gigawatts. ERCOT found that in-service data centers drew, on average, only about half of the capacity they requested. Even the largest buyer pulled back, with reports of Microsoft cancelling or deferring up to 2 gigawatts of leases in early 2025. The queue gigawatts are an upper bound, not a delivery schedule.
But the estimator's conclusion survives the skepticism, because the constraint Fern tracks is set by already-booked backlog, not by future orders. Turbines and transformers booked today convert to deliveries three-plus years out regardless of what next year's orders do. So the long-lead electrical equipment on the Pulse tracker — switchgear, transformers, large generators — stays elevated through 2027 and 2028 even if data-center demand softens. The near-term risk to lead times is a supply-driven plateau, not a collapse.
Fern's read
The signal to watch is the split between firm orders and slot reservations in the turbine backlog: a wave of reservation cancellations would precede any real loosening, and would show up before the lead-time trackers move. Until then, the practical guidance is unchanged from the tracker — carry an explicit lead-time allowance on long-lead electrical scope, and sequence procurement early, because on this equipment the schedule risk is now larger than the price risk.
Methodology and limitations
Sources: NERC 2025 LTRA and EIA for demand; company filings and releases from GE Vernova, Siemens Energy, Hitachi Energy, and Eaton for backlogs; the counter-evidence on speculative load from AEP Ohio, ERCOT, and reporting on Microsoft lease deferrals. Lead-time ranges reconcile with Fern's sourced Pulse tracker.
Limitations: several counter-evidence items are second-hand or paywalled channel checks; treat them as directional. Published queue and peak-demand figures are upper bounds inflated by speculative interconnection requests, so this section deliberately leans on filed backlogs and disclosed lead times rather than gigawatt forecasts. The demand-side durability of AI beyond roughly 2028 is a genuine tail risk, but it sits outside the horizon on which today's backlog is already committed.