Luan Gjokaj/Unsplash
For 70 years, memory got cheaper almost every year — a background law of computing so reliable that analysts stopped even tracking it. That law broke in 2025. A new analysis published Monday by Tom’s Hardware, drawing on PCPartPicker’s historical average-price database, confirmed what PC builders have been discovering in real time: DDR5 memory now costs what it cost in 2007, erasing two decades of exponential price decline in a matter of months.
The culprit is not a supply disruption, a natural disaster, or a geopolitical embargo. It is a deliberate capacity reallocation by the three companies that manufacture roughly 90% of the world’s DRAM — Samsung, SK Hynix, and Micron — who have redirected their fabrication lines toward High Bandwidth Memory, the stacked, ultra-fast DRAM used inside AI accelerators, where margins run three to five times higher per wafer than conventional DDR5. Every wafer retooled for HBM removes approximately three gigabytes of consumer DDR5 from the market for every one gigabyte of HBM it produces — a hard physical tradeoff built into how HBM is made. The result is a consumer memory market running on the scraps of a supply chain that has reoriented itself almost entirely around artificial intelligence.
Every Tier Has Been Hit: The Full DDR5 Price Table
For the first time since RAMageddon began, here is the full year-over-year damage across every common DDR5 configuration, drawn from PCPartPicker average pricing data for August 2026 versus August 2025.
Memory ConfigurationAugust 2025 Avg.August 2026 Avg.Change
DDR5-4800, 2×16GB
$90
$425
+372%
DDR5-5200, 2×16GB
$100
$480
+380%
DDR5-5600, 2×16GB
$116
$528
+355%
DDR5-6000, 2×16GB
$108
$572
+429%
DDR5-5600, 2×32GB
$191
$1,118
+485%
DDR5-6000, 2×32GB
$222
$1,272
+473%
DDR5-6400, 128GB kit
$329 (all-time low)
$3,399
+933%
Tom’s Hardware confirmed the 128GB figure independently: a DDR5-6400 128GB kit that reached an all-time low of $329 now lists at $3,399 — ten times its record low. That is a configuration popular with content creators and power users for whom memory capacity affects export times, virtual machine loads, and the ability to run large AI models locally. At $3,399, it has moved from a premium workstation line item to a price that rivals an entire entry-level system. The full Tom’s Hardware price tracker documents the damage in real time.
For mainstream gaming builds, the damage is severe but somewhat more legible. A 64GB DDR5-5600 kit that cost $191 last summer now clears $1,118. A 32GB DDR5-6000 kit that ran $108 — at the time a reasonable spec for a mid-range gaming PC — now costs $572.
Computer science researcher Daniel Lemire put the historical scale in precise terms in a post on X on August 5, 2026. Citing memory price data compiled by David Shim at Stanford University’s DAM Project, Lemire wrote that current DDR5 prices — now running between $11.41 and $13.28 per gigabyte — match 2007 levels for DDR2 in nominal terms, and roughly 2011 levels when adjusted for inflation. He called this a “historical anomaly,” adding: “I know of no other case where the price of technology hardware has reverted to levels from decades earlier.” Tom’s Hardware confirmed the findings against Stanford DAM Project data independently.
DDR4 Offers No Escape
The natural response for anyone on an older platform is to stay put and wait out the DDR5 crisis. That strategy no longer works. DDR4 has absorbed its own surge:
Memory ConfigurationAugust 2025 Avg.August 2026 Avg.Change
DDR4-3200, 2×8GB
$63
$163
+159%
DDR4-3600, 2×8GB
$75
$165
+120%
DDR4-3200, 2×16GB
$105
$281
+168%
DDR4-3600, 2×16GB
$120
$307
+156%
DDR4-3200, 2×32GB
$222
$614
+177%
DDR4-3600, 2×32GB
$300
$789
+163%
The cause of DDR4’s rise is the same force driving DDR5: manufacturers are winding down DDR4 production rather than maintaining it as a cheap parallel line, because the economics of concentrating output on DDR5 and HBM are better. A buyer who held onto their DDR4 platform specifically to avoid upgrading now faces memory prices that are 120% to 177% above where they were twelve months ago. The “wait on the old platform” exit is closed.
How HBM Manufacturing Removes Consumer DRAM From the Market
Understanding why this is happening requires a brief look at how High Bandwidth Memory is actually made.
Standard DDR5 DRAM is a flat, single-die component that sits on a circuit board. It connects to the processor through a 64-bit channel, delivering roughly 60 to 80 gigabytes per second of memory bandwidth. Manufacturing one DRAM die is a single-layer process: deposit, etch, repeat. Fast, high-volume, commodity economics.
HBM is built entirely differently. To produce one HBM stack, engineers take between eight and twelve individual DRAM dies, thin each one to a fraction of a millimeter, then drill hundreds of thousands of microscopic copper columns — called through-silicon vias, or TSVs — vertically through each die. Those copper pillars connect the layers electrically. The completed stack mounts on a silicon interposer alongside an AI processor, with a 2,048-bit interface instead of DDR5’s 64-bit channel. That interface is what gives HBM4 its bandwidth: up to two terabytes per second per stack, compared to DDR5’s roughly 60 gigabytes per second — a 33-fold difference.
The TSV drilling, die thinning, stacking, and interposer bonding all require dedicated cleanroom equipment that cannot simultaneously produce DDR5. When Samsung began mass-producing HBM4 in early 2026, it did so by converting wafer starts from commodity DDR5 lines. Each gigabyte of HBM that came out of those lines removed approximately three gigabytes of potential consumer DDR5 from the market. That ratio is the physical reason consumer prices are where they are. It is not a business decision that can be reversed with a policy change; it is a consequence of the manufacturing physics of high-bandwidth memory.
The economics reinforce the physics. HBM modules sell for $60 to $100 each, compared to $5 to $10 for a comparable amount of conventional DDR5 — a margin differential of ten to twenty times. When a manufacturer with constrained wafer capacity must choose between a product that generates ten times the margin and one that does not, the allocation decision follows. Samsung’s semiconductor division posted record quarterly results in Q2 2026, with revenue of ₩89.2 trillion (approximately $61.7 billion) on the strength of AI memory — all-time highs for the division. The same dynamics that are making consumer DDR5 unaffordable are making Samsung’s profits historic.
What Are Prices Going to Do? Two Forecasts, One Optimistic
The consensus view from the executives who run the memory companies, and from the research firms that track them, is consistent: prices will not recover to 2025 levels, and 2027 may be harder than 2026.
SK Hynix CEO Kwak Noh-jung delivered the starkest public forecast on July 10, 2026, the day the South Korean chipmaker began trading on the Nasdaq in a record $26.5 billion ADR offering that became the largest US listing by a foreign company in history. In an interview with Reuters on that day, Kwak said: “We forecast that next year will be the worst year in the industry’s history from the supply perspective. Our customer demand continues to go up, while our capacity has limitations. We still forecast that customer demand will remain higher than our supply capacity even beyond 2030.” The Reuters interview with Kwak was published on the day of SK Hynix’s historic Nasdaq debut.
Micron CEO Sanjay Mehrotra told Bloomberg in May 2026 that Micron can only meet about 50% to two-thirds of customer requirements for HBM and DRAM from its key customers — a remarkable admission from the CEO of one of only three companies capable of making the product. He has characterized the gap between supply and demand as “the highest that we have ever seen.” Multiple outlets confirmed Mehrotra’s supply gap statement from his May 22, 2026 Bloomberg interview. Micron discontinued its consumer-facing Crucial brand in February 2026 to concentrate entirely on AI and enterprise customers, eliminating the most recognizable retail memory brand in the US market.
ADATA Chairman Simon Chen offered perhaps the longest timeline of any executive on record. In comments made in late 2025 and repeated in 2026, Chen dismissed the notion of an AI bubble that would pop and restore normal memory market dynamics, suggesting the crisis could extend for another decade.
The supply-side math for any recovery is punishing. New fabrication plants take two to three years to build and cost $10 to $20 billion each. Micron’s Idaho fab is not expected to ship meaningful consumer-grade DRAM volume until mid-2027 at the earliest. SK Hynix’s new Yongin Y2 DRAM fab targets its first cleanroom in June 2029. Samsung’s Pyeongtaek P5 megafab targets mass production in the latter part of 2028. Even when those facilities come online, hyperscalers — Microsoft, Google, Meta, and Amazon, which are collectively approaching $725 billion in AI infrastructure spending in 2026 — have reportedly already made advance deposits to lock in most of the 2027 output from all three manufacturers. The TechTimes South Korea chip report from August 15 covers the full supply allocation picture.
Is There a Contrarian Case?
There is one voice in the market making a materially different argument: Bloomberg Intelligence analyst Shuli Ren, who published research in July 2026 arguing the global memory shortage may have already peaked in the second quarter of 2026.
Ren’s model uses what she calls a “sufficiency ratio” — a supply-demand balance measure — and projects that conditions could stabilize by the fourth quarter of 2027, with the industry potentially swinging into oversupply by 2028 as the massive capital investment all three major manufacturers have committed begins to yield actual chips. The logic is not that AI demand will collapse, but that the semiconductor industry’s historical tendency to overbuild in response to high prices is already in motion, and that by 2028 the new capacity coming from Micron, Samsung, SK Hynix, and the maturing Chinese producer CXMT could tip the market from scarcity to glut. Multiple analysts confirmed Ren’s Bloomberg Intelligence research and its Q2 2026 peak thesis.
Even Ren’s optimistic scenario does not predict a return to 2025 prices. Her forecast for “stabilization” is a condition where prices stop rising rather than one where they fall meaningfully. And her timeline — easing in late 2027, potential oversupply in 2028 — is still 12 to 24 months away.
The practical implication for a builder today: the analysts most likely to be right are those with the most direct visibility into supply allocation commitments. SK Hynix and Micron both report that 2027 supply is already fully spoken for by hyperscalers. That visibility makes a “peak is in” thesis plausible for investor modeling purposes but does not translate into near-term retail price relief for a consumer who cannot access long-term allocation agreements.
Why Did This Reach Every Country?
The crisis is not a US phenomenon. German technology publication ComputerBase reported this week that average RAM prices in Europe have risen 345% compared to September 2025 — a trajectory broadly parallel to the US market. Their data also showed hard drives and SSDs up more than 125% in the same period, confirming that the squeeze has spread beyond DRAM into adjacent memory components that share fabrication infrastructure with NAND flash.
Major OEMs including Apple, HP, Dell, and Lenovo have raised device prices, with several explicitly citing memory costs in earnings calls and customer communications. Google confirmed that the Pixel 11 will carry a higher price tag than its predecessor, with Google’s SVP of Devices pointing to mobile LPDDR5X prices surging from $2.80 per gigabyte in 2025 to approximately $12 per gigabyte in 2026 — a roughly fourfold increase in the cost of smartphone RAM. Samsung’s own Galaxy phone division recorded its first-ever quarterly operating loss in Q2 2026 despite the chip division’s record profits — the same company was simultaneously the biggest winner and biggest loser from the memory crisis.
Even the gaming console market is feeling the pressure. Speculation has circulated within the industry that Sony’s PlayStation 6 could face supply challenges in securing DRAM allocation at scale, though no confirmed delay has been officially announced.
Is There a Legal Challenge?
Seventeen consumers and small businesses filed a class-action antitrust lawsuit on June 25, 2026 in the U.S. District Court for the Northern District of California, naming Samsung, SK Hynix, and Micron as defendants. The case — Garciaguirre v. Samsung Electronics, No. 3:26-cv-06345 — alleges the three companies coordinated a deliberate restriction of conventional DRAM supply, using the shift toward HBM production as cover to curtail DDR3 and DDR4 output while prices rose approximately 700% over four years. The lawsuit invokes Section 1 of the Sherman Antitrust Act.
The complaint cites the companies’ prior criminal history as pattern evidence: Samsung pleaded guilty to DRAM price-fixing charges in 2005 and paid a $300 million fine; SK Hynix (then Hynix Semiconductor) pleaded guilty to the same charges and paid $185 million, with executives in both companies serving prison time. Micron avoided criminal penalties by cooperating with investigators in the first case.
The allegations remain unproven. None of the three defendants had filed a court response as of this writing. The central legal challenge the case faces is identical to the one that killed a second civil class action in 2020: proving that parallel conduct — three companies independently making the same rational business decision to reallocate capacity toward higher-margin products — constitutes an illegal agreement under Sherman Act doctrine, rather than independent competitive behavior that happens to produce similar outcomes.
What Should Builders Do Today?
For anyone planning a PC build right now, the calculus is painful and the options are limited.
The Ren thesis offers a possible argument for waiting: if prices ease in late 2027, a builder who defers a 64GB DDR5 build by 18 months might save $600 or more on memory alone. But “easing” in Ren’s model means prices stop rising, not that they fall to 2025 levels. And the majority of analysts — including the CEOs of the companies making the memory — forecast continuing increases through 2027 before any stabilization.
For builders who cannot wait, the full-spectrum table above provides the most precise available signal: know exactly what your target configuration costs at today’s prices, build with the minimum viable RAM for your actual workload now, and budget for an upgrade when — or if — fabrication capacity catches up to demand.
The era of RAM as a cheap, predictable line item in a build budget is over for the foreseeable future. The Tom’s Hardware tracker that once helped enthusiasts chase $3-per-gigabyte deals is now helping them understand how to spend $11 to $13 per gigabyte instead. For content creators, developers, and power users whose work requires 64GB or more, the question is no longer where to find the best deal — it is whether to build at all.
Frequently Asked QuestionsWhy did RAM prices increase so much, and is this a one-time thing or permanent?
The increase is structural, not cyclical in the traditional sense. Samsung, SK Hynix, and Micron — the three companies that make roughly 90% of the world’s DRAM — have redirected their fabrication capacity toward High Bandwidth Memory, which earns three to five times the revenue per wafer of conventional DDR5. Each gigabyte of HBM manufactured requires approximately three times more cleanroom capacity than a gigabyte of DDR5, and the manufacturing process (stacking dies with through-silicon vias) physically cannot run on the same equipment as flat-die DDR5 production. Whether this resolves as new fab capacity comes online (2027–2029 for most new facilities) or persists further depends on whether AI infrastructure demand keeps absorbing that new capacity. The manufacturers themselves — whose CEO statements are quoted in this article — project that demand will outstrip supply through at least 2030.
How bad are current RAM prices historically?
They are historically unprecedented in technology hardware. Computer science researcher Daniel Lemire, analyzing data from Stanford University’s DAM Project memory price database compiled by David Shim, found that DDR5 now costs $11.41 to $13.28 per gigabyte — prices last seen in 2007 for DDR2 memory in nominal terms, or roughly 2011 levels adjusted for inflation. Lemire called this a “historical anomaly” and noted he could not identify a prior case where the price of technology hardware had reverted to levels from decades earlier. Technology hardware prices typically fall exponentially over time as manufacturing processes improve; the current crisis is the first documented instance of that trend reversing significantly for a mature consumer product category. Tom’s Hardware independently confirmed Lemire’s findings against Stanford DAM Project data.
Is DDR4 a viable lower-cost alternative to DDR5 right now?
No. DDR4 prices have risen 120% to 177% year-over-year across common configurations — less than DDR5’s 355% to 485% surge, but severe enough that it no longer represents a meaningful escape valve. A 2×32GB DDR4-3200 kit that cost $222 in August 2025 now averages $614. PCPartPicker’s DDR4 price data confirms the increase across all common configurations. The underlying cause is the same: manufacturers are winding down DDR4 production alongside DDR5, since dedicating lines to either consumer format generates lower margins than redirecting them toward HBM. A buyer holding onto a DDR4 platform to avoid upgrading will still face memory prices that are 120% to 177% above last year’s levels.
Is there a class-action lawsuit over RAM prices, and what could it mean for buyers?
Yes. Seventeen consumers and small businesses filed a class-action antitrust lawsuit on June 25, 2026 — Garciaguirre v. Samsung Electronics, No. 3:26-cv-06345 — in the U.S. District Court for the Northern District of California, alleging Samsung, SK Hynix, and Micron coordinated a supply restriction that drove DRAM prices up approximately 700% over four years. The lawsuit invokes Section 1 of the Sherman Antitrust Act and cites the three companies’ prior 2005 criminal guilty pleas for DRAM price-fixing as pattern evidence. If the case proceeds and results in a settlement or judgment, it could mean damages for a broader class of consumers who purchased products containing DRAM during the shortage period. However, a comparable civil case was dismissed in 2020 after courts found insufficient evidence that parallel conduct constituted an illegal agreement — the new case must clear the same bar, using the HBM allocation agreements as its primary evidence of coordination. The allegations remain unproven, and none of the defendants had responded in court as of this writing.