Why Is RAM So Expensive in 2026?

The answer, in one paragraph

RAM is expensive because the three companies that make almost all of the world's DRAM decided to build something else with their best capacity. High-bandwidth memory - the stacked DRAM that sits beside an AI accelerator - takes roughly 23% of DRAM wafer supply in 2026, against about 8% in 2024, and it earns three to five times the revenue per wafer that conventional DRAM does. The DRAM market is more than 95% concentrated across those three manufacturers, so the demand displaced by that decision has no fourth supplier to go to. This is not scarcity in the ordinary sense of a factory burning down or a mine closing. It is an allocation decision, taken rationally by each maker independently, whose combined effect is that the memory in your computer is competing for wafers with the memory in a datacentre and losing.

What actually happened to the prices

The consumer figures are the ones most people have seen, and they are worth stating precisely rather than as an impression. DDR5-5600 2x32GB kits averaged a little over $200 for much of the eighteen months to late 2025. Between October and December 2025 they rose to around $800, with DDR5-6000 kits at roughly $900 over the same window, on PCPartPicker's average-price trend data as reported by TechRadar Pro.

Those are reported figures rather than our measurements, and we label them that way throughout. We track ECC server memory, not consumer kits - the distinction matters and we come back to it below.

On the contract side, TrendForce revised its own first-quarter 2026 outlook upward rather than down. Conventional DRAM contract prices went from a previous estimate of +55-60% quarter-on-quarter to +90-95%, published on 2 February 2026, with the release noting that further upward adjustments might still occur. That revision is worth dwelling on: the widely-quoted 55-60% figure is the number its own author withdrew, and it still circulates as though it were the forecast.

The mechanism, and why it does not unwind quickly

Three facts describe why this is structural rather than a spike.

First, the concentration. More than 95% of DRAM comes from three makers. In a market with a dozen suppliers, one reallocating capacity creates an opening for the others; in a market with three, all facing the same incentive, it does not.

Second, the revenue gap. HBM earning three to five times per wafer what conventional DRAM earns is not a margin any manufacturer walks away from, and 2026 HBM capacity at SK hynix was reported committed well in advance. Committed capacity is not available capacity.

Third, the lead time. SK Group chairman Chey Tae-won said in March 2026 that the shortage is likely to persist through 2030, with wafer supply trailing demand by more than 20%, and that building the capacity to close that gap takes four to five years. That is the single most important number in this story for anyone deciding whether to wait: the fix, if it is coming, is measured in years rather than quarters.

The downstream effects are already visible in company guidance. MSI told investors that 2026 would be the most challenging year since the company was founded, on memory and GPU component costs, and signalled hardware price increases in the 15-30% range.

What our own data shows, from the inside

This is where we can add something rather than summarise. We track ECC registered server memory - the modules that go into rack servers - priced per gigabyte from live in-stock listings, with daily snapshots. The chart above is that series.

It matters because server memory is the market that HBM displaced INTO. Conventional server DRAM is the direct substitute for the wafer capacity that moved, so the squeeze shows up here first and most clearly. A consumer kit price is the same pressure after it has travelled through several layers of channel and inventory.

Two notes on how to read it. The chart is a levels series: it shows what the cheap end costs, not what the market did. Every percentage change we publish comes from a matched-model index that compares each listing against its own prior price, because comparing one month's median to another's measures which listings we added rather than what sellers did - a trap that once had our own hard drive figure at +10.3% when the real matched movement was +6.18%, and flipped the sign on two of three categories in the same month.

The other half of our data is the one almost nobody publishes: a continuously tracked DDR4 series. DDR4 became a legacy product when capacity moved to DDR5 and HBM, but the installed base that needs it cannot switch - a server with DDR4 slots takes DDR4. Shrinking supply against inelastic demand is the textbook setup for a price rise, and it is why DDR4 has been climbing faster in relative terms than DDR5 in much of the reporting, to the point that MSI is reported to be shifting its motherboard mix back toward DDR4 boards. Our live DDR4-versus-DDR5 split is on the server RAM forecast with its tracking window stated.

Server versus consumer: what transfers and what does not

Our figures are ECC server memory. If you are pricing a gaming build, do not read our floor as your price - the products are different, the channels are different, and the absolute numbers will not match.

What does transfer is the mechanism and the direction. Both markets are downstream of one wafer-allocation decision taken by the same three companies. That makes the server series we can measure genuine evidence about the consumer market we cannot: a different point on the same curve, not a different curve.

The practical difference is timing. Server contracts reprice on a quarterly cadence and move first. Consumer retail moves later, cushioned by inventory bought at older prices, then moves suddenly when that inventory runs out - which is exactly the shape the October-to-December 2025 consumer jump has.

What would change this

An explanation with no falsifiers is a story rather than an analysis, so here is what we are watching.

HBM's share of DRAM wafers falling back toward its earlier level would be the clearest signal, because that share is the mechanism itself. Committed 2026 HBM capacity being reallocated or expanded would be the next clearest.

On our own side: the matched-model index turning negative across consecutive monthly editions - a run rather than one month - and in-stock breadth recovering for reasons other than our own catalogue growing.

One thing that would NOT count: consumer kit prices stabilising at their elevated level. Prices that stop climbing after quadrupling have not fallen, and the weekly-delta framing of most coverage makes a flat month read as relief. The level is what matters, and the level is still roughly four times where it started.

We review these monthly and publish the result as a dated edition whether or not the read changed. The directional outlook, with the full falsifier list, is on the server RAM price forecast.