The Slowest Rung Sets the Clock

I was on a call yesterday where several sharp people agreed on two things. The industry’s classic boom-and-bust cycle will come back, and it won’t be in the next year or so.

I think they’re right on both counts. Where I’d push is on the mechanism. If you expect the old cycle to return the old way, you’ll be watching the wrong signals. That’s how you get blindsided by the turn, or talk yourself out of an upturn that has more road left than it looks.

How the old cycle kept time

For most of the industry’s history, a cycle had one binding constraint: wafer capacity. The industry controlled the fix. Demand ran ahead, prices rose, and everyone placed tool orders at the same time. Roughly 18 to 24 months later the new capacity arrived all at once, and supply overshot.

Memory ran this play with almost metronomic regularity. One analysis of past cycles notes that each one topped roughly 12 to 18 months after the capex arms race began, because that’s the lead time from groundbreaking to output. Spend money, wait, flood the market. The clock was set inside the fab, and everyone in the industry could read it.

That clock is why seasoned people expect the cycle back. It worked for forty years.

Why that clock stopped keeping time

Today the binding constraint isn’t one thing. It’s a ladder, and the shortage keeps moving down it.

Look at advanced packaging. TSMC’s CoWoS supply gap is projected to narrow from about 20% early in 2026 to about 10% by year-end. By the old logic, AI chip lead times should be easing. They aren’t, because the binding constraint simply relocated to the next rung.

The next rung is memory. All three major suppliers have described their 2026 HBM output as sold out or fully allocated, and one has warned of the most severe shortage on record in 2027. HBM also eats wafers. HBM3E can require about three times the wafer supply per gigabyte of conventional DRAM, so every shift toward HBM removes bits from the rest of the market.

Below memory are rungs the semiconductor industry doesn’t control at all: grid interconnects, transformers with lead times now running up to five years, construction labor, permits, local consent. No etch tool shortens a utility queue.

Here’s the part that changes the cycle math. When the slowest rung sits outside the industry and takes five-plus years to move, overbuilding a faster rung doesn’t create a glut. It just moves the shortage somewhere else. You can’t overshoot a system whose throughput is set by its slowest stage.

The number that looks like a glut in the making

The headline the cycle camp points to is real: combined capex at Micron, Samsung and SK Hynix is projected to rise nearly 340% between 2024 and 2027. On its face, that’s exactly the arms race that preceded every past bust.

Read it as bits, not dollars, and it looks different. A growing share of that spend goes into HBM, which delivers fewer bits per wafer. Greenfield fabs take three-plus years to build. One estimate has SK Hynix still years away from meaningful new volume from its key projects, landing in mid-to-late 2029. The capex line flatters the supply outlook. What the market eventually feels is bits shipped, and those arrive on a much longer fuse than the spending headline implies.

The parallel that fits: oil, 2003 to 2014

Inside our industry, I haven’t seen this combination before. Outside it, there’s a close match.

The oil supercycle of 2003 to 2014 had almost the same structure:

• Producers came in disciplined. After the busts of the 1980s and 1990s, the majors held back on capex. Memory makers today carry the same scar tissue from 2022 and 2023.
• Demand came from one structural source. China’s industrialization then, hyperscaler AI build-outs now. Neither was a consumer replacement cycle that rolls over on its own.
• The constraints were stacked. Rigs, crews, refining capacity and tankers each had their own lead times, and money alone couldn’t compress them.
• High prices lasted most of a decade. Apart from the 2008 financial-crisis plunge, prices stayed elevated long enough that “the cycle is dead” became a common view.

It wasn’t dead. Prices broke hard starting in 2014. The cause wasn’t conventional overbuilding. Shale opened a new route around the old constraint, and demand growth slowed at the same time.

That’s the lesson I’d take from it. The cycle survived. What changed was the period, which stretched from a couple of years to roughly a decade, and the trigger, which came from outside the old playbook.

How a re-clocked cycle eventually turns

If the slowest rung sets the clock, the turn won’t come from fab capacity 18 months out. It comes from one of three places.

1. Demand has to convert into return. AI infrastructure is a bet that usage turns into revenue. Today, relatively few companies are using AI well, and building that capability takes training, process change and time. If the return arrives slower than the build, spending eventually adjusts. I’ve written more on that in “The Return Doesn’t Ship With the Chip.”
2. An efficiency shock. Telecom learned this in 2001, when wavelength multiplexing multiplied the capacity of fiber already in the ground. The equivalent here would be an architectural or algorithmic leap that cuts the memory or compute needed per unit of useful work.
3. The slowest rung clears, and the other rungs land together. This is the least discussed and the most important. Each layer is building on its own timeline. If power and permitting finally catch up, capacity built separately across packaging, memory and logic can all land in the same window.

The third one is why I’d argue the eventual correction could be sharper than a normal memory cycle, not milder. But eventual is the operative word.

Why I don’t see it in the next couple of years

Match the horizon to the evidence. The lead times on the binding rungs run from 2027 into 2029: HBM allocated through 2026 with suppliers calling for tighter 2027 supply, packaging easing but not closed until around 2027, major new memory volume in 2029, and transformers and grid connections on multi-year queues. Demand is still being guided up, not down. Alphabet raised its 2026 capex guidance to $175 to $185 billion, and Amazon raised its outlook to $200 billion.

One caution on the memory side: a supplier forecasting a shortage it profits from is talking its book. I discount those quotes. Even so, the physical lead times say the same thing without anyone’s help.

That doesn’t mean ignoring the risks. It means watching the right signals, the ones that show up before the turn rather than with it:

• Order quality. The 2020 to 2022 shortage was multi-layer too, and it ended in the 2023 inventory glut partly because double-ordering had inflated demand. Watch for inventory building at the customer.
• Contract behavior. Spot prices slipping below contract, or long-term agreements being renegotiated on volume rather than price.
• How capex is funded. Spending funded from operating cash flow is a growth story with control. Spending funded increasingly by debt and outside capital is the same growth with less control.
• The slowest rung itself. When interconnect queues and transformer lead times start shortening, the clock starts running faster for everyone.

None of these are flashing today. Most are moving the other way.

The question worth asking

So yes, I expect the cycle to return. I just don’t expect it on the old schedule, and I don’t expect it to start in the fab.

The useful debate isn’t “cycle or no cycle.” It’s which rung is binding right now, and what would have to happen for it to stop binding.

Where do you see the clock being set today?