No large industry cycles like semiconductors. Every three to four years: shortage, expansion, glut, repeat. And the most counterintuitive scene plays out again and again — prices fall below cost, the whole industry is losing money, and yet the fabs keep running flat out. That is not madness. It’s arithmetic.
Shutting Down Is More Expensive Than Losing Money
A fab is heavy industry in its purest form: tens of billions invested per line, one to two years to ramp up. Once the money is spent, it’s sunk — so the loss from shutting a line down exceeds the loss from selling chips below cost. Stopped machines need requalification, trained crews scatter and never come back whole, customers defect for good. When the next upturn arrives, restarting an idled fab costs far more than keeping a loss-making one warm. So the rational move at the trough is exactly the one that looks irrational: produce at a loss, spread the fixed costs, keep the team and the capacity alive, and outlast the marginal competitor.
This arithmetic is why the cycle never dies. When demand rises, capacity can’t keep up — fabs take a year or two to build, and by the time they’re done, demand may have moved on. When demand falls, capacity has just come online at full tilt. Every glut is the late-arriving bill for the previous shortage. The 2022–2024 semiconductor winter was the textbook case: post-pandemic demand faded, inventories ballooned, memory prices fell below cash cost, the entire industry bled — and not a single major producer dared to stop.
Historic, but Structural
“This time is different” resurfaces every few years, usually justified by “this boom is historic.” Line up the three most recent peaks and that justification loses its shine.
The 2000 dot-com bubble was the industry-wide valuation peak — Cisco and Intel traded at triple-digit P/E ratios, and it was called historic then too. The 2017–2018 memory supercycle tripled or quintupled SSD and DRAM prices and sent Samsung and Micron into frantic expansion — but it was a structural peak: only memory was boiling, while logic and analog chips were merely fine. The 2025–2026 AI-driven cycle is the same animal in extreme form: GPUs, AI memory, and optical modules enjoy genuinely historic demand — Nvidia’s revenue doubled in a year — while mature nodes, consumer electronics, and automotive chips never overheated across the board. “Historic” and “universally hot” are different things, and 2000 already proved it with a crash.
Watch Utilization, Not Headlines
The current call: the AI-compute-driven upcycle is entering top territory. With the industry-wide expansion wave now underway, 2027–2028 is likely to bring a fresh downcycle of overcapacity.
The basis for that call is not bullish headlines but the industry’s own gauges: utilization rates and the expansion wave. Full utilization plus concentrated capex spending means capacity is being stacked onto the future; when that capacity lands together while demand growth cools, a glut becomes a matter of time. The cycle’s position is written in factory blueprints, not in news headlines.
A business that must produce at a loss is a business that must endure the trough at a loss. The cycle never dies — it just changes the name on the chip. The boundary as usual: 2027-2028 is a deduced position in the cycle, not a trading date; no stock picks, no price targets here. Utilization and capex readings follow official disclosure — misread the data and the cycle reads wrong in turn; that risk belongs to every reader personally.
