Why RAM Got So Expensive, and Why Older Memory Is No Escape

Memory prices climbed through 2026 because AI data centers and ordinary computers are supplied from the same factories. That much you can read anywhere. What follows is the part that is harder to find: why the usual trick of buying a generation older stopped working, and why the simple version of the story turns out to be wrong.

The surge did not begin this year. TrendForce revised its outlook for conventional DRAM contract prices in the fourth quarter of 2025 upward from an earlier forecast of 8 to 13 percent growth to 18 to 23 percent, and the increases got steeper from there. By the time most people noticed at a checkout page, the market had already moved twice.

Why an AI accelerator competes with your laptop

The memory stacked next to a data center AI accelerator is not the same product as the memory in a PC. It is High Bandwidth Memory, or HBM, which Micron describes as stacking multiple DRAM dies vertically on top of a base die, connected by thousands of through-silicon vias so that far more data can move in and out at once. Different package, different customer, same starting point: it is DRAM, made on the same kind of wafer, in the same fabs.

Those fabs have a fixed number of wafer starts per month, and someone has to decide what goes on them. TrendForce described the decision suppliers were making at the start of the year, forecasting that DRAM makers would “continue to reallocate advanced process nodes and new capacity toward server and HBM products to support rising AI server demand.” Reallocation is the operative word. No factory burned down and no shipping lane closed; the same capacity was pointed somewhere else.

Here is where the popular explanation goes wrong. It is usually told as a simple price chase, with manufacturers abandoning consumers because HBM sells for so much more. That was true, and then it stopped being true: TrendForce reported that HBM3e commanded a price premium of more than four times DDR5 as of the second quarter of 2025, but projected in October 2025 that DDR5 profitability would overtake HBM3e from the first quarter of 2026 onward. For most of the period when consumer memory was getting expensive, the memory in your laptop was not the less profitable product.

What kept capacity flowing toward the data center anyway was volume and commitment. Cloud providers were securing supply through long-term agreements, which is a different kind of pressure than a spot price. A fab planner allocating wafers eighteen months out weighs a signed multi-year contract for enormous quantities against a retail channel that buys in small lots and walks away when prices rise, and the contract wins that argument even in a quarter when the per-gigabyte margin does not favor it.

The escape route that closed

When a component gets expensive, the standard response is to step back one generation. That has worked for decades in PC hardware and it is close to useless right now, because the crowd went there first.

TrendForce documented the pattern in June 2026, reporting that supply shortages in consumer DRAM were “cascading down through successive technology generations,” with “DDR4 designs being replaced with DDR3 solutions, while certain DDR3-based products are being redesigned to use DDR2.” DDR2 was standardized in 2003. Its contract prices were estimated to have risen 55 to 60 percent in the second quarter of 2026, with a further 35 to 40 percent expected in the third.

It is worth being precise about what that does and does not show, because the tempting version overstates it. Legacy memory did not rise faster than everything else across the whole period: conventional DRAM overall was forecast up 58 to 63 percent in that same second quarter, slightly ahead of DDR2. The gap opens in the third quarter, when conventional DRAM growth was forecast to moderate to 13 to 18 percent while DDR2 kept climbing at 35 to 40. So the claim that holds up is narrower than the headline version, and more useful: shortage reached generations everyone had written off, and it stayed there after the broader market began to cool.

Nor is the cascade caused by HBM alone. The same June report notes that Winbond was reducing DDR2 production to concentrate on higher-margin DDR3, DDR4 and LPDDR4, while ESMT expanded DDR2 capacity through PSMC to fill the gap. Individual suppliers exiting and entering old product lines for their own reasons is a second, quieter force acting on the same shelves.

This reaches further than gaming builds. DDR3 and DDR4 are common in equipment with long design lifetimes, where the memory type is fixed at the schematic and cannot be swapped the way a desktop kit can, so a manufacturer in that position ends up bidding for old parts against everyone else who was pushed down the same ladder.

A brand that stopped selling to consumers

The clearest single marker of where this went is that one of the best known consumer memory brands left the retail market entirely.

On 3 December 2025 Micron announced it was exiting the Crucial consumer business, continuing shipments through the consumer channel only until the end of its fiscal second quarter in February 2026, with warranty service and support for existing products continuing after that. The company framed the move as aligning to “secular, profitable growth vectors in memory and storage.” Sumit Sadana, Micron’s chief business officer, put the cause plainly in the announcement: “The AI-driven growth in the data center has led to a surge in demand for memory and storage.”

Read that as a capacity statement, not a sales one. Micron did not describe consumer demand as weak; it described data center demand as large enough to be worth clearing a retail brand out of the way for.

Where it stands in September 2026

Prices are still rising, and the rate of increase is expected to fall sharply. TrendForce forecast conventional DRAM contract prices up 55 to 60 percent quarter on quarter for the first quarter of 2026, 58 to 63 percent for the second, and 13 to 18 percent for the third, a figure it repeated on 7 September. The reasons it gave for the slowdown were weakening consumer demand, smartphone brands becoming more conservative in procurement, and high base effects, which is the arithmetic point that a large increase on top of an already large increase is harder to sustain.

Three caveats belong with those numbers. They are contract price forecasts published by a research firm that sells memory market analysis, not retail prices at a checkout. The third-quarter figure describes a quarter that has not closed yet. And a smaller increase is still an increase, layered on the ones before it, so nothing in the current forecasts describes prices returning to where they were.

What would genuinely change the picture is added capacity instead of reallocated capacity. While a supplier can shift some legacy production inside a quarter, as ESMT did through PSMC, new leading-edge wafer capacity is a multi-year commitment. Until more of it arrives, the same allocation question gets asked again every quarter, and the answer has not changed.

The other half of this story is the silicon itself, and how AI chips work explains why those factories are worth reallocating. For the layer above the hardware, start with what a language model actually is.