Why Memory Always Crashes and What's Different This Time - Part 1

Samsung raised the price of a 32GB DDR5 module from $149 to $239 in a single month. Contract DDR5 prices roughly tripled across 2025. HBM capacity is sold out across SK Hynix, Samsung, and Micron through the end of 2026. Server memory pricing is reportedly up to 70% higher for Q1. Micron exited the consumer memory market entirely in December. This is what a memory boom looks like. Anyone who's been around the sector for more than one cycle knows what usually comes next.

For thirty years, memory has been the most reliably cyclical business in technology. Prices spike, manufacturers race to add capacity, supply floods the market eighteen to twenty-four months later, prices collapse, everyone bleeds, capex gets cut, and the whole thing resets. It has destroyed shareholder value, bankrupted operators, and made every memory analyst suspicious of the phrase "this time is different." And yet, the people who run these companies are telling a different story. Capex isn't ramping the way it did in prior cycles. The customer mix has changed. The product mix has changed. The oligopoly has tightened to three players from twenty. Something structural is happening underneath the price action.

The question isn't whether this is a boom; that part is obvious. It's whether the conditions that have always produced the bust are still in place.

The Capital Cycle, Memory Edition

DRAM and NAND are commodities in the textbook sense. A 16GB module from Samsung is functionally interchangeable with one from SK Hynix. Customers buy on price, qualify on spec, and switch suppliers without much friction. Demand is elastic and price-sensitive. Supply is the opposite. Fabs take years and tens of billions of dollars. Capacity decisions made today produce wafers 18-24 months from now. Once a fab is online, the marginal cost of running it is low and the incentive to keep it running is high.

That combination - elastic demand, inelastic capex-heavy supply, undifferentiated product - is the textbook setup for a capital cycle. Edward Chancellor laid this out in Capital Account, drawing on Marathon Asset Management's work: commodity industries follow boom-bust cycles where returns on capital compress as more capital floods in, eventually producing oversupply, losses, and forced supply discipline.

In memory, the cycle runs in four beats. Shortage develops and prices rise. Capex follows as suppliers race to capture margin. Supply floods 18-24 months later into a market that has normalized. Prices collapse, margins compress, the cycle resets. Every memory cycle of the past thirty years has followed this pattern.

Three Crashes, One Pattern

• 1995 to 1997: The early Windows PC era produced a DRAM boom that Korean industrial history still references. Gross margins cleared 50%. Then capex hit. Roughly 50 fab construction plans were announced between 1995 and 1996, with capex exceeding 30% of semiconductor production. DRAM prices peaked in late 1995 and collapsed: down 51% in 1996, then down another 65% in 1997. The resulting shock contributed to the Asian Financial Crisis.
• 2017 to 2019: The hyperscaler buildout produced one of the cleanest examples of the cycle. Samsung raised capex by more than 50% year on year during the boom. That capacity addition is what caused the 2019 bust. Supply growth re-accelerated, demand normalized as end markets digested inventory, and pricing collapsed. The industry crashed itself by overbuilding in a period of euphoria.
• 2022 to 2023: The post-pandemic correction produced one of the worst down cycles in recent memory. Inventory built up across the channel, hyperscaler purchasing slowed, and suppliers operated below cash cost on parts of their portfolios. The big three responded with deliberate supply discipline, slashing utilization and deferring fab projects. That discipline is what set up the conditions we're in now.

Three different end markets. Same mechanism every time: shortage, capex, supply, collapse.

What's Actually Happening Right Now

Set the historical pattern aside. Here's the current setup in plain numbers:

• DRAM contract prices are forecast up roughly 55%-60% in Q1 2026, with server DRAM climbing more than 60%. NAND flash up 33%-38%.
• Samsung and SK Hynix reportedly hiked server memory by up to 70% for Q1. Combined with 2025 increases, server pricing has roughly doubled inside twelve months.
• HBM3E pricing was hiked nearly 20% for 2026 deliveries, unusual given HBM4 is supposed to be the mainstream product.
• HBM capacity is sold out across SK Hynix, Samsung, and Micron through the end of 2026.
• Micron exited consumer memory in December 2025 to focus entirely on AI and enterprise customers.
• Hyperscaler capex among the top eight cloud providers is projected to exceed $600 billion in 2026, up roughly 40% year on year.

The mechanism is also what reshaped the cycle. HBM consumes roughly three times the wafer capacity of standard DRAM per gigabyte. Every wafer reallocated to HBM is three wafers' worth of capacity coming out of the commodity pool. AI isn't just buying more memory. It's eating the manufacturing capacity that used to make everything else.

Come back tomorrow for Part 2 of this topic!