The semiconductor industry has a rhythm, and it has punished everyone who ignored it for sixty years. The pattern: demand surges, capacity lags, prices spike, everyone orders equipment, capacity arrives all at once, prices collapse, capex gets cut, and the survivors buy the distressed. The 2018–2019 downturn. The 2020–2022 shortage, when lead times stretched past a year and automakers idled plants for want of $2 microcontrollers. The 2023 inventory correction that halved memory prices. Ask a veteran where we are in 2026 and you will get a long pause — because this cycle looks familiar and alien at the same time.
Start with the familiar part. Wafer fab equipment spending is tracking toward roughly $153 billion in 2026, up about 28% from 2025 — the kind of capex number that has historically marked the top of the cycle, not the middle. Memory is in a full supercycle: DRAM prices up an estimated 125%, NAND up 234%, memory revenue nearly tripling to $633 billion. Gartner's analysts warn that this "memflation" will destroy or at least delay non-AI chip demand into 2028. Translation: we are in the phase where one segment's boom is eating the rest of the industry's lunch — a classic late-cycle signature. TSMC is guiding 2026–2028 capex "significantly higher" than the prior three years. ASML just raised prices on its most critical tools for the first time in its history. Every prior cycle teaches the same lesson: when the equipment makers have pricing power, the peak is closer than it looks.
The Alien Part: Structural Demand
Now the unfamiliar part, and the reason this cycle may run longer than the history books suggest. Three structural shifts are stretching the cycle's duration. First, AI infrastructure spending is behaving like a utility build-out, not a product cycle — hyperscalers have committed more than $1.1 trillion since 2023, with $745 billion more expected in 2026, and the spending is debt-financed at a scale ($400 billion raised this year) that signals multi-year commitment. Second, each new technology node requires more equipment per wafer: gate-all-around transistors, backside power delivery, and advanced packaging mean tool reuse is falling — demand for equipment has partially decoupled from wafer volume. Third, HBM consumes three to four times the wafer capacity of standard DRAM per bit, so AI servers structurally tighten memory supply in a way smartphones never did.
The honest synthesis: this is still a cycle, but it is a longer and higher-amplitude one than the industry's history would predict. The equipment analysts who called for $125–129 billion of 2026 WFE spending in January revised to ~$153 billion by mid-year — and admitted they had been too cautious, because memory makers and foundries committed capital faster than expected once AI demand proved durable.
How Cyclical Investors Actually Position
So how do you invest in a cycle that is both extended and, eventually, still a cycle? The practitioners' playbook has four rules.
Rule one: buy the trough in equipment, sell the peak in memory. Equipment stocks bottom when orders bottom — typically 6–12 months before the chip recovery shows in revenue. Memory stocks top when pricing power looks permanent; it never is. Additional memory capacity is expected to enter the market in 2027, and while AI demand should keep conditions tight, the history of DRAM pricing says the turn, when it comes, will be violent.
Rule two: watch the dashboard, not the narrative. The leading indicators are book-to-bill ratios at the equipment makers, foundry utilization rates (leading-edge logic and DRAM fabs are running at roughly full utilization today — a coincident indicator of tightness, and a contrarian one at extremes), memory contract prices, and inventory days at distributors. When lead times normalize and distributors start reporting rising inventory, the cycle has turned whether or not anyone has announced it.
Rule three: separate the AI signal from the cycle noise. Non-memory semiconductor revenue is forecast to grow from $589 billion in 2025 to $718 billion in 2026 — a healthy 21.9%, driven by CPUs, networking silicon, power management, and optical interconnect for AI clusters. That is real, durable demand growth underneath the memory fireworks. The investors who get hurt are the ones who extrapolate memory's 200%+ revenue growth to the whole industry.
Rule four: size for the drawdown you know is coming. The quotable version: in chips, you do not get paid for predicting the cycle — you get paid for surviving it with capital intact and dry powder ready. Position sizes should reflect that a 40–50% peak-to-trough move in memory names is a normal event, not a tail risk.
The 2026 cycle will reward the disciplined and invoice the euphoric, exactly as every cycle before it has. The only thing that has changed is the amplitude. Respect it.
Cycle discipline is an edge. semiconductorfund.com is available for acquisition. Inquiries: info@primedomaincapital.com.