The semiconductor capex cycle is turning on tool orders, and the order book says 2027 is softer

Semiconductor cycles are usually narrated through chip prices, because chip prices are visible and move quickly. They are more reliably read through tool orders, because a lithography system ordered today ships in fifteen to twenty months and the decision to order it was made against a capacity plan drawn up a year before that. By the time spot memory pricing moves, the capital decisions that will define supply two years out have already been taken.

Those capital decisions are turning. Not collapsing, and not uniformly, but turning. Three of the five largest equipment buyers have now guided 2027 capital budgets below 2026, and the order books at the two largest deposition and etch suppliers show book to bill slipping under one in the June quarter for the first time since early 2023.

The 2027 estimate of $121bn represents a decline of about 9.7%. That is a mild downturn by historical standards. The 2023 contraction was 17.8% from peak, and 2019 was worse than that. What makes this one awkward is the composition, because the spending that is still growing and the spending that is falling belong to different customers with different economics.

Capital intensity is concentrating

A mature node fab producing power management chips or microcontrollers costs roughly $1.8bn to $2.6bn for 40,000 wafer starts a month. A leading edge logic fab at the current node costs between $22bn and $28bn for the same nominal capacity, and a meaningful share of that goes into a small number of lithography tools. The gap has widened at every node transition since 2016 and shows no sign of narrowing.

Two lines in that table are worth separating from the rest. Advanced packaging is the only category growing meaningfully, because high bandwidth memory stacks and chiplet architectures have moved packaging from a commodity back end step into a capacity constraint. It is also the smallest category, so a 25% increase adds about $1.5bn against an $11bn decline elsewhere.

Mature logic is the category falling hardest. Capacity added in 2023 and 2024 for automotive and industrial demand arrived into a market that did not grow as expected, and utilisation across that segment has run in the high sixties for four quarters. Nobody adds capacity into that.

What Sanborn Micro's deferral signals

Sanborn Micro, the specialty foundry, disclosed in its August update that it had moved roughly $1.9bn of planned 2027 capacity spending into 2028. The company framed it as phasing rather than cancellation, and the equipment orders behind it were reportedly rescheduled rather than cancelled, which matters for the suppliers. It is still the first deferral Sanborn has disclosed since 2023 and it came from a management team that had spent two years telling investors demand was supply-constrained.

Rescheduling an order is not the same as cancelling it, and equipment companies will tell you that all day. The distinction holds for one quarter. If a customer reschedules twice, the second reschedule is a cancellation that has not been written down yet.

Hannah Ilves, semiconductor analyst at Quillon Data

Lead times are doing the company's work

One reason a downturn in this sector takes so long to become visible is that equipment backlogs absorb the first year of weakness. Vasa Lithography carries a backlog equivalent to roughly five quarters of shipments. Revenue can keep rising for a year after bookings fall, which is exactly what happened in 2022 and what appears to be happening now. Investors who track revenue rather than bookings will find out late.

• Bookings lead revenue by four to six quarters at the lithography suppliers and two to three quarters at the deposition and etch suppliers.
• Utilisation below about 80% at a mature node fab reliably precedes a capex cut within two quarters, because the cash cost of an idle tool is small but the depreciation is not.
• Advanced packaging capacity decisions now move independently of front end capacity decisions, which breaks the historical rule that all semiconductor capex moved together.

The third point deserves more attention than it gets. For most of the industry's history, front end and back end capacity moved in step because back end was cheap and followed automatically. High bandwidth memory changed that. Packaging capacity is now a genuine bottleneck for a specific class of accelerator products, and it is being expanded during a period when front end spending is falling. The two halves of the industry are on different cycles for the first time.

Equipment shares have not priced a downturn. The sector trades at about 26 times forward earnings against a ten year median near 19, and the premium is being justified by the packaging story and by a structural argument that capital intensity per transistor keeps rising. Both arguments are sound. Neither prevents a cyclical business from having a cyclical year.

The most useful disclosure this autumn will come from the foundries rather than the toolmakers. If specialty utilisation stays in the high sixties through December, the 2027 estimate is too high. If it recovers into the mid seventies, the deferrals announced this summer get quietly reversed in the spring and the cycle turns out to have been a pause.