On this page
RAM prices have roughly doubled, server memory quotes change week to week, and the shortage is expected to last years. The cause isn't a factory fire or a pandemic — it's the AI buildout, and one piece of wafer math explains almost everything. Here's what happened to the memory market, in plain terms.
Last updated: July 2026 — Note: this covers a fast-moving market. Specific figures below are attributed to their sources and dated; treat them as snapshots, not quotes.
If you've priced a memory upgrade lately — for a server, a gaming PC, or anything in between — you've seen it: prices that would have seemed like typos two years ago. This isn't a normal market wobble. Industry analysts have called it the most severe memory dislocation in decades, and unlike past shortages, this one is structural. Understanding why takes about five minutes, and it changes how you should plan hardware purchases for the next couple of years.
The short answer
Three companies — Samsung, SK Hynix, and Micron — manufacture roughly 90–95% of the world's DRAM. Starting in 2025, all three redirected as much production capacity as possible toward two products the AI buildout consumes voraciously: high-bandwidth memory (HBM), the stacked DRAM that feeds GPUs in AI data centers, and high-capacity server DDR5. Both carry far richer margins than ordinary memory. Every wafer that becomes an HBM stack is a wafer that doesn't become a server DIMM or a stick of desktop RAM — and demand for everything else didn't go down. Prices did what prices do when supply shifts away from steady demand: they exploded.
The wafer math that broke the market
The detail that makes this shortage different is how expensive HBM is to manufacture in capacity terms. HBM stacks up to a dozen thinned DRAM dies into a single package, and by TrendForce's estimate it consumes roughly four times the wafer area per gigabyte of conventional DRAM. So even a modest-sounding shift of production toward HBM removes an outsized share of the world's total memory output. Layer on the demand side — research firm IDC has forecast that data centers will absorb on the order of 70% of all memory produced in 2026, a share that sat around 20–30% as recently as 2022 — and the squeeze on everyone else stops being mysterious.
Timing made it worse. The memory industry runs in boom-and-bust cycles, and 2022–2023 was a bust: makers cut output and postponed new fabs. When AI demand arrived at unprecedented scale, the industry was at a low point in capacity investment. New fabrication plants are being built — but a leading-edge fab takes years, and the major new capacity isn't expected to reach volume production before 2027.
How big the price moves have been
A few attributed snapshots convey the scale. Counterpoint Research measured conventional DRAM prices rising 80–90% in a single quarter around the turn of 2026. TrendForce-affiliated analysts tracked buyers paying 50% more quarter-over-quarter for standard DRAM in late 2025, with expedited orders commanding two to three times list. Projections for Q2 2026 pointed to a further 58–63% rise in conventional DRAM contract prices. S&P Global's consensus estimates have Samsung's revenue per bit of conventional DRAM more than doubling year-over-year in 2026. Even DDR4 — older memory in wind-down — spiked, because buyers grabbed whatever they could get. And in a sign of how thoroughly priorities shifted, Micron announced it would exit its Crucial consumer memory brand entirely in early 2026 to focus supply on data-center customers.
It's hitting SSDs too
NAND flash — the memory inside SSDs — is caught in the same current. AI data centers need enormous, fast storage tiers, enterprise SSD demand has surged, and NAND contract prices jumped sharply through late 2025 and 2026, with knock-on effects reaching consumer SSDs after years of falling prices. The storage side of this story has its own dynamics (including pressure from the nearline hard-drive market) and deserves its own article — we'll cover the SSD market in a dedicated follow-up post.
When will prices come back down?
The honest answer: not soon, and possibly not to where they were. Supply growth in 2026 is running below historical norms (IDC estimates around 16% for DRAM), HBM capacity for 2026 sold out entirely, and analysts broadly place the earliest meaningful relief in late 2027 to 2028, when new fabs ramp. Several add a caveat familiar from past cycles: if AI infrastructure spending cools just as that new capacity arrives, the industry could swing to oversupply — but planning around that possibility is speculation, not strategy. The two wildcards that could ease things sooner — an AI spending pullback or major memory-efficiency gains in AI software — are exactly that: wildcards.
What this means if you run servers
For IT teams and homelabbers, four practical implications:
Extending existing servers got more valuable, not less. Memory and storage prices rose — but so did the cost of whole new servers, which contain the same expensive DRAM and NAND plus everything else. The economics that favored upgrading a healthy Gen10 or Gen11 server over replacing it have strengthened. Know your platform's headroom: our guides to DL380 Gen10 maximum memory, Gen11 memory configurations, and Gen12 memory configurations map what your machines can take.
If an upgrade is planned, sooner beats later. In a market where analysts forecast double-digit quarterly increases, deferring a known memory purchase has been expensive. This isn't a call to panic-buy or speculate — buying memory you don't need at elevated prices helps nobody — but a budgeted upgrade for a real workload is worth executing rather than sitting on.
Right-size with more care than usual. When DIMMs were cheap, over-provisioning was harmless insurance. At 2026 prices, matching capacity to the workload — and populating channels correctly so you get full performance from every module — matters. Our server memory upgrade checklist covers the process.
Installed memory is an asset now. Servers full of DDR4 or DDR5 hold more of their value than they did two years ago, which changes decommissioning math and makes matching existing modules (rather than ripping and replacing) the economical path when expanding.
Frequently asked questions
Why is RAM so expensive in 2026?
Because the AI data center buildout redirected the world's memory production. The three manufacturers who make roughly 90 to 95 percent of DRAM shifted capacity toward high-bandwidth memory and server DDR5 for AI systems, which carry higher margins, leaving less supply for everything else while demand stayed strong. Prices for conventional memory rose steeply through late 2025 and 2026 as a result.
What is HBM and why does it matter to memory prices?
High-bandwidth memory is stacked DRAM packaged tightly with AI accelerators to feed them data fast enough. It matters because it is extraordinarily capacity-hungry to manufacture, consuming roughly four times the wafer area per gigabyte of conventional DRAM by industry estimates. Shifting production toward HBM therefore removes an outsized share of total memory output from the rest of the market.
How much have memory prices actually risen?
Attributed snapshots include Counterpoint Research measuring conventional DRAM prices up 80 to 90 percent in a single quarter, analysts tracking 50 percent quarter-over-quarter contract increases in late 2025, and projections of a further 58 to 63 percent rise in the second quarter of 2026. Exact figures vary by product and week; the direction has been consistently and steeply upward.
When will memory prices come down?
Most industry forecasts place the earliest meaningful relief in late 2027 to 2028, when new fabrication plants reach volume production. Supply growth in 2026 is below historical norms and HBM capacity for the year sold out entirely. Several analysts expect prices to settle at a higher new normal rather than returning to earlier levels, though a pullback in AI spending could change the picture.
Does the shortage affect SSDs too?
Yes. NAND flash, the memory inside SSDs, is being pulled toward enterprise and AI storage demand, and NAND contract prices rose sharply through late 2025 and 2026, ending years of falling SSD prices. The storage market has its own dynamics, which we cover in a dedicated companion article.
What should I do if my server needs more memory?
Execute planned, budgeted upgrades sooner rather than later, since forecasts point to continued increases; right-size capacity to the workload instead of over-provisioning; match your existing modules when expanding rather than replacing; and verify your platform's limits and population rules before buying. Extending a healthy existing server remains more economical than replacing it, since new servers carry the same expensive memory plus everything else.
The bottom line
The 2026 memory crunch is what happens when an industry built on boom-and-bust cycles meets a demand shock without precedent: three suppliers, a four-to-one wafer penalty for the product everyone suddenly wants, and years of lead time on new capacity. For most of the tech world it means pricier everything; for people who run servers, it rewrites the upgrade calculus in favor of extending what you own, buying what you've planned, and sizing with discipline. If a memory upgrade is on your roadmap, browse genuine HPE server memory by platform, find modules for your exact machine on our HPE parts by server model pages, or contact our team — we'll confirm the right genuine HPE SmartMemory for your configuration, at today's price rather than next quarter's.
