Why Memory Always Crashes and What's Different This Time - Part 2

If you have not yet read yesterday's blog post, please do so before continuing here.

Three Reasons This Cycle Looks Structurally Different

1. The oligopoly is dramatically tighter: During the early 90s supercycle, there were roughly twenty meaningful DRAM suppliers. Today, three companies - Samsung, SK Hynix, and Micron - account for roughly 95% of global DRAM. In HBM specifically, SK Hynix sits at 53%-57%, Samsung at around 35%, and Micron at 11%. The historical capital cycle assumes competitive pressure forces individual suppliers to ramp even when collective discipline would be more profitable. When the supply base is twenty companies, someone always defects. When it's three, defection is visible, costly, and reciprocally punishable. The game theory has fundamentally changed.
2. Capex is being managed with unusual discipline: In the 2017 to 2018 boom, Samsung raised capex by more than 50% year on year. That's the textbook driver of the 2019 collapse. In the current cycle, SK Hynix is reportedly raising capex by 17% for 2026. Samsung is raising capex by roughly 11%. Both are emphasizing process upgrades and advanced packaging over greenfield capacity. Micron has explicitly deferred its New York megafab project. There are two ways to read that. The optimistic read is that the operators learned from 2019 and 2023 and are deliberately managing the cycle. The skeptical read is that capex always lags pricing, and the real expansion announcements are still ahead. Both are defensible. What's not in dispute is that the capex print, eighteen months into a price boom of this magnitude, is materially below the historical baseline.
3. HBM is not a commodity in the way DRAM is: Standard DRAM is interchangeable. HBM is none of those things. Customer qualification takes months. The customer base is concentrated, with NVIDIA alone accounting for roughly 90% of SK Hynix's HBM supply. Each generation is co-designed with the accelerator it pairs into. Capacity is increasingly booked through customer prepayments and multi-year supply agreements. That structure looks more like contracted aerospace than spot-market DRAM. The disciplining mechanism of commodity oversupply - anonymous buyers walking away when prices drop - doesn't operate the same way in this market.

Where the Cyclical Risk Still Lives

The conditions that produce memory busts haven't disappeared. Instead, they've moved. Capacity is still coming. Samsung is reportedly looking to expand HBM production capacity by roughly 50% in 2026. SK Hynix has announced infrastructure investment more than four times its prior figure. Even with disciplined capex relative to history, the absolute capacity additions in 2027 and 2028 will be significant.

Demand concentration is a feature and a risk. The same concentration that gives HBM its specialty-product economics also creates fragility. If hyperscaler capex normalizes, if NVIDIA's order book softens, the demand side compresses fast. The cycle doesn't need oversupply to bust. It can bust from demand normalization.

The HBM4 ramp creates its own cyclical pocket. If yields jump faster than demand can absorb the volume, the suppliers themselves create the oversupply. This is roughly what happened in earlier cycles when process shrinks dramatically increased bits per wafer. And the 2010 to 2011 cycle also looked structurally different at the top. Prices fell 50%-70% from their peak in less than two years. Every cycle that crashes looks structurally different on the way up. That's part of what makes them crash.

How to Watch It

• Capex as a percentage of production. Currently well below the 30% threshold that marked the 1995-1997 top. A meaningful inflection upward, particularly in greenfield announcements rather than packaging upgrades, would be the clearest historical signal.
• Inventory days at the supplier level. Sold-out conditions can flip to inventory builds inside one to two quarters once demand pull slows.
• HBM yield prints. Yield improvements show up as effective capacity additions without a single new fab being announced.
• Hyperscaler capex guidance. The top eight cloud providers are the demand engine. Revisions on accelerator deployment and capex envelopes are the leading indicator.
• Executive language. The shift from "sold out" to "demand adjusting" is the qualitative tell. Memory executives have done this dance for thirty years and the vocabulary is consistent.

The Cycle Is Re-Shaped, Not Repealed

The memory cycle hasn't been abolished. It's been re-engineered. The oligopoly is tighter than it has ever been. The capex response is more disciplined than in any prior cycle. The product mix has shifted toward HBM, which behaves more like contracted aerospace than commodity DRAM. These are structural shifts in how the market clears. None of this means a bust never comes. The 2027 and 2028 capacity additions are real, the demand concentration creates new fragility, and the HBM4 ramp is the kind of supply-side event that has historically produced oversupply without anyone needing to add a fab. The conditions for a cyclical correction haven't disappeared. They've moved.

Every cycle ends. This one will too. The interesting question is whether it ends the way the last six did, or whether the structural changes underneath produce something the historical pattern doesn't anticipate. The honest answer is that the conditions look meaningfully different - and that's exactly what every cycle has looked like at the top.