Why is RAM so expensive? Because the three companies that make almost all of the world's DRAM shifted their factories toward memory for AI systems in 2025, that demand has absorbed everything they can build, and the consumer market is buying what's left. The result was the sharpest memory price rise in a decade: across the kits MemRadar tracks, DDR5 costs several times what it did a year ago and DDR4 is up more than 70%. DDR4 stopped climbing in the spring; DDR5 has kept grinding upward since. Nothing in the supply picture points to a quick reversal.

How much more expensive, exactly

Numbers first, because "expensive" gets thrown around loosely. Across the DDR5 kits MemRadar tracks, the median price is well over triple what it was a year ago. DDR4 is up more than 70%. The climb was not gradual. For DDR5, most of it happened between October and December 2025, and prices have kept grinding slowly upward since. DDR4 rose more evenly, kept climbing into spring 2026, then cooled off its peak and went flat.

The more telling measure is how the catalog sits against its own history. 207 of the 235 products MemRadar tracks are priced at least one and a half times their all-time low, and 152 are at more than triple it.

Tracked products grouped by current price divided by all-time low
Current price vs all-time lowRAMSSDAll
Under 1.5x20828
1.5x to 2x347
2x to 3x04848
Over 3x9656152
Tracked119116235

Each product compared against its own recorded low, not against a market index.

Those lows are not ancient history. For most kits, the all-time low was recorded in 2024 or early 2025, when memory was as cheap as it had ever been. Someone who paid $93 for a 32GB DDR5 kit in November 2024 and looks up the same kit today is not misremembering; the number on that kit's page is a multiple of what they paid.

Where the memory went

Three companies make almost all of the world's DRAM: Samsung, SK Hynix, and Micron. Between them they control the overwhelming majority of supply, they add capacity through multi-year fab projects that cost billions, and they have spent a decade learning to keep supply tight enough to stay profitable. That structure means a demand shock has nowhere to go but price.

The demand shock was AI. Training and running large models needs a specialised type of memory called HBM, high-bandwidth memory, stacked directly beside the processor. The accelerator buildout of 2024 and 2025 turned HBM into the most profitable product the memory industry has ever sold, and two facts about it explain most of what followed. HBM is made on the same fab lines and from the same wafer starts as ordinary DRAM, so every wafer that becomes HBM is a wafer that does not become DDR5. And HBM uses far more silicon per gigabyte than conventional memory, with lower yields, so a fab that shifts toward it produces fewer total gigabytes. The industry did exactly what its economics told it to do and moved capacity to the customers paying the most.

DDR4 was squeezed from a different direction. The major manufacturers announced the wind-down of mainstream DDR4 production during 2025, retiring lines that had run for a decade. A product being discontinued while the platforms that use it still sit in millions of homes is a recipe for a spike, and DDR4 delivered one before cooling.

NAND flash, the memory inside SSDs, followed a related path. The same datacenter buildout needs enormous amounts of fast storage, enterprise SSD demand rose, and NAND makers redirected supply toward it. SSDs did not get the HBM effect, but they got the same signal: the consumer market became the residual buyer.

Why it landed on consumers, and why so fast

Memory is sold in tiers. The largest buyers, server builders, phone makers, PC manufacturers, buy on quarterly contracts. A spot market handles what is left. Retail, the kits on Amazon and Newegg, sits at the end of the chain, priced off contract and spot with a lag of weeks. When contract prices jump, retail follows a month or two later, and it follows all the way; there is no cushion of margin to absorb it.

That lag is why MemRadar's charts show the explosion landing between October and December 2025 rather than earlier in the year, when the supply shift was already under way. It is also why retail prices can keep rising for a while after the underlying market has cooled, and why they will lag on the way down.

This has happened before

This is not the first DRAM spike, and the earlier ones are useful because they all ended the same way. Between 2016 and early 2018, memory prices roughly doubled on a smaller supply crunch, as manufacturers migrated to new process nodes while smartphone demand surged. MemRadar's longest-tracked kits show it clearly: a steep climb through 2017, a peak in early 2018, and then a long slide as new capacity arrived and the market tipped into oversupply. By 2019 prices were below where the spike had started, and the years from 2020 to 2024 were the cheapest memory has ever been.

G.SKILL RipjawsV Series DDR4 16GB 3200MHz 2x8GB, tracked continuously since 2015. The 2017 to 2018 spike, the long slide to 2024, and the 2025 surge on one axis.

That is the memory industry's cycle in one chart. Shortage lifts prices, high prices fund new fabs, new fabs arrive together, oversupply crashes prices, and the cheap years fund the next round of demand. Every DRAM spike in thirty years has ended in oversupply. The open question is never whether, only when, and the answer depends on how long the current demand lasts and how fast the new capacity arrives.

What would make prices fall

Four things, and none of them is buyers waiting harder.

New capacity. The fabs announced in response to the shortage will not ship meaningful volume until 2027 at the earliest; a memory fab takes years to build, equip, and qualify. Until then, total supply is roughly what it is.

AI demand levelling off. If accelerator demand plateaus, or the manufacturers' HBM capacity catches up with it, wafers flow back toward conventional memory. That is the single largest lever, and it is not one anyone outside the industry controls.

DDR4 working through. The discontinuation shock has largely played out in MemRadar's data; DDR4 prices cooled from their peak and have been flat for months. Further movement depends on remaining inventory against a shrinking installed base, which points to stable rather than cheap.

The lag running in reverse. When contract prices turn, retail follows weeks later. MemRadar's charts will show the turn a month or two after the industry feels it. That is the cost of tracking the prices people actually pay rather than the ones manufacturers quote.

What this means if you are buying

That is a different article, and it is here: Should I Buy RAM Now?. The short version is that waiting for 2024 prices is not a plan the data supports this year, but paying a fair price within the current market is, and fairness is checkable against each kit's own history.

Questions this page gets asked

Is this AI's fault? Mostly. The demand shock is AI-driven, but the price response is a product of how concentrated DRAM manufacturing is. In a market with a dozen suppliers, a demand surge pulls in capacity; in a market with three, it pulls up prices. Both halves are true and neither is going away soon.

Will DDR4 get cheap again? Not 2023 cheap. Its spike has already cooled and flattened in MemRadar's data, which suggests the discontinuation squeeze has mostly passed. But a product that is no longer being made does not usually get cheaper as its supply runs down, and DDR4's remaining buyers are people who cannot switch platforms. Expect stable, not falling.

Are SSDs expensive for the same reason? A related reason, not the same one. SSDs use NAND flash rather than DRAM, and NAND has its own set of manufacturers and its own cycle. The datacenter buildout pulled on both, but through different mechanisms, and SSD prices behaved differently through 2026. That story gets its own page.

Figures on this page update automatically every few hours from tracked prices. Page last regenerated: October 13, 2026. The industry narrative is reviewed when conditions change meaningfully.