The estimate landed on November 5 and it does not look like a memory market number at all. Counterpoint Research put global DRAM revenue for the third quarter at close to $200 billion, up 369% from the same quarter a year earlier. That is very nearly five times the year-earlier figure, in twelve months, for a commodity that is sold by the gigabit.
The mechanism behind the increase is the part worth reading twice. Counterpoint attributed the gain mainly to prices, not to a jump in the number of chips shipped. Unit shipments rose about 4% from the previous quarter. The average selling price reached $1.90 per gigabit.
Run the arithmetic and the shape of the quarter becomes clear. A 4% sequential increase in volume against a 369% annual increase in revenue means the revenue line is almost entirely a price line. The industry sold roughly the same number of bits it sold three months earlier, and collected several times what it collected a year ago for each of them. At that run rate the quarterly figure annualizes to roughly $800 billion, a sum that rests on price rather than on bits.
DRAM is the working memory in every computer, phone and server: the pool of cells a processor reads from and writes to while it computes. It is manufactured in a small number of plants owned by three companies, and it is sold on contracts that reset every quarter or every month. Because the product is standardized and the plants are enormously expensive, the industry has spent four decades alternating between shortages that make it wildly profitable and gluts that make it unprofitable. The price per gigabit is the single number that determines which of those regimes is in force.
The current regime has a specific cause, and it is the same cause behind most of the AI buildout’s supply problems. High-bandwidth memory, the stacked DRAM that sits next to an AI accelerator, consumes far more wafer capacity per bit than the commodity DRAM used in a laptop. When manufacturers shift wafers to HBM, the supply of ordinary DRAM falls even though nobody has stopped buying laptops. Add a year of customers building inventory defensively, and a market that was in surplus becomes a market where buyers accept whatever allocation they are offered.
The transfer of margin is the consequence that shows up outside the memory industry. A phone maker that plans a bill of materials a year ahead now finds the memory line in that plan moving against it, and a PC maker that competes on price has to choose between absorbing the increase and shipping less memory. Server builders face the same math at a larger scale, and for cloud operators the memory content of a machine is now a material part of what a rack costs to fill.
For the memory makers, the revenue figure is not the same as a profit figure, and the distinction matters more at this point in a cycle than at any other. Wafer capacity is added in multiyear increments: a fab is a building, a permit, a set of lithography tools and a customer qualification process that takes years from decision to output. Revenue that arrives because prices tripled does not create the capacity to serve the next customer, and the capital that gets committed at the top of a cycle is the capital that gets written down at the bottom.
The industry has run this experiment repeatedly. In the downturns of 2018 and 2022, prices per gigabit fell far enough and fast enough to push the smaller manufacturers into losses while the largest ones kept building. Each cycle ended with fewer suppliers and a more concentrated market, which is part of why a price move of this size is possible in the first place. Consolidation does not prevent cycles. It changes who survives them.
What makes this one different from a routine shortage is the demand behind it. Previous memory booms were driven by PCs, then by smartphones, then by cloud servers, each of which has a predictable replacement cycle. The current demand is driven by training and inference for large models, which is a use case with no installed base to replace and no clear ceiling, and by a customer base that has demonstrated it will pay almost anything to avoid being short of accelerators. That is a better demand story than a phone refresh. It is also a demand story that can be revised downward in a single earnings call.
The $1.90 per gigabit figure is the one to watch, because it is the number that will appear in the contracts negotiated for the December quarter. If the average price holds, the revenue run rate holds with it. If it slips, the annual comparison stops being a multiple of five and starts being a comparison of two prices, and the entire narrative of the market changes with it.
Counterpoint’s estimate describes a quarter in which an industry that sells a commodity at a price set by supply extracted the value of a shortage. Whether that is a durable new level or the peak of an old cycle is a question the next two quarters will answer, and no one who has watched this market for a full cycle is willing to answer it early.
