The 2026 Memory Supercycle, Explained: Why RAM Prices Exploded and What the Earnings Tape Shows

If you tried to buy computer memory this year, you already know something unusual is happening. A 32GB DDR5 kit that sold for well under $100 in mid-2025 now starts around $375 — and that's the cheap end. Behind that sticker shock sits the most violent move in memory-chip prices in at least fifteen years, a shortage some analysts have called a once-in-four-decades event.

This article walks through what actually happened, using the companies' own earnings reports rather than headlines: the supply-side mechanism that created the shortage, the record quarters it produced at Micron, SK Hynix, and Samsung, the hyperscalers writing the checks, the long history of memory booms that ended badly — and the specific things that would tell you whether this one is different.

The price tape: three quarters of vertical

Start with the raw numbers. Memory chips come in two basic kinds: DRAM, the fast working memory every PC, phone, and server thinks with, and NAND flash, the storage inside SSDs and phones. Both are sold two ways — on a retail “spot” market, and through contract prices that big buyers negotiate directly with manufacturers, which is where the serious volume moves. TrendForce, the research house whose contract-price surveys are the industry benchmark, recorded a 90–95% quarter-over-quarter increase in conventional DRAM contract prices in the first quarter of 2026 — enough to push the DRAM industry's quarterly revenue up 81% in a single quarter. The second quarter added another 58–63%, with NAND contracts up 70–75%. For the third quarter, TrendForce expects server DRAM to rise a further 13–18% — a slower pace, but notably slower because large buyers locked in multi-year supply agreements, not because demand cooled.

The squeeze reaches all the way down the stack. Shortages have spread to products as old as DDR2; Samsung roughly doubled mobile DRAM prices from early-2025 levels; and TrendForce cut its 2026 smartphone and notebook shipment outlooks because memory got too expensive to put in cheap devices. Even Apple — famous for squeezing its supply chain, not being squeezed by it — capitulated: CEO Tim Cook confirmed in a June 17 Wall Street Journal interview that price increases are “unavoidable,” saying Apple had “been trying to shield our customers from the increases, but the situation has become unsustainable.” Mac prices have already risen — by as much as 18% on some configurations, per Forbes — and research firm TechInsights estimates the memory and storage inside a Pro-tier iPhone now costs Apple roughly $200, up from about $50 a year earlier. Supplier inventories sit at roughly 2–4 weeks against a normal 13–17. TrendForce projects the total memory market reaching $551.6 billion in 2026, with DRAM revenue up 144% year over year — for context, that would make memory alone roughly as large as the entire semiconductor industry was in 2021.

The mechanism: HBM and the 3-to-1 rule

The easy explanation — “AI needs lots of memory” — is only half the story. The sharper insight is on the supply side: the shortage is largely self-inflicted by a change in what memory makers choose to build.

AI accelerators like NVIDIA's GPUs don't use ordinary memory sticks. They use high-bandwidth memory (HBM) — stacks of DRAM chips packaged directly next to the processor. HBM is extraordinarily profitable, but it is hungry for wafers — the silicon discs every chip starts life on, the industry's true unit of capacity. Producing one gigabyte of HBM consumes roughly 3–4× the wafer capacity of one gigabyte of standard DDR5. Every wafer a factory allocates to an AI accelerator is roughly three wafers denied to laptops, phones, and servers. HBM already consumes about 23% of industry DRAM wafer capacity, and its share grows with every generation — each new HBM generation stacks more chips per package.

So when the AI buildout pulled memory makers toward HBM, conventional DRAM supply didn't just grow slowly — it was actively starved while demand rose. That is the signature of this cycle: prices at record highs while bit shipments — the actual volume of memory sold — barely grew. Micron's DRAM shipments rose only single digits in its record quarters; the revenue explosion was almost entirely price.

The economics explain the behavior. HBM3 sold for roughly $200 per stack; HBM3E about $300; HBM4 — now ramping for NVIDIA's next platform — around $500. Micron reports its HBM4 ramp is tracking about twice as fast as HBM3E's did, with over $1 billion of HBM4 revenue already shipped. The HBM market was roughly $35 billion in 2025, is headed toward $60 billion in 2026, and Micron's management projects ~$100 billion by 2028.

The earnings tape: three record printers

Shortages make headlines; earnings prove them. All three memory majors printed the best quarters in their history in the first half of 2026 — and each print carried the same signature of price-driven, supply-constrained growth.

Micron is the cleanest read because it reports on an offset fiscal calendar, giving the freshest data. Its FQ3 2026 (March–May, reported June 24) delivered $41.5 billion of revenue, up 346% year over year, at an 84.9% gross margin — meaning that of every revenue dollar, roughly 85 cents survived the cost of actually making the chips. That is commodity-memory economics replaced, within a single year, by margins that look like enterprise software. Micron beat its own quarter-old guidance by ~24% on revenue and ~31% on EPS, and guided the next quarter to $50 billion. Management extended its supply-tightness horizon to “beyond calendar 2027,” with no line of sight to supply catching demand.

SK Hynix's first quarter — normally a seasonally weak one — produced record revenue, operating profit, and net profit simultaneously, at margins (72% operating) that memory companies have simply never reported. Samsung earned more operating profit in one quarter than in its entire prior fiscal year, and its memory chief publicly warned that, based on orders already received, the supply gap in 2027 is expected to be even wider than 2026's.

There's a fourth confirmation, one step upstream. ASML, the sole supplier of the EUV lithography machines that print nanometer-scale circuits onto wafers — equipment every leading-edge chip factory requires — reported that memory makers' share of its tool sales flipped from ~30% to 51% in Q1 2026. Memory capex is landing in the toolmaker's books — the expansion is real and being paid for.

Who's paying: the hyperscaler capex wave

Record seller revenue is someone else's record spending. The buyers are the hyperscalers — the handful of companies that operate the world's giant cloud data centers — and their 2026 budgets are extraordinary: Microsoft guided to roughly $190 billion of calendar 2026 capital expenditure, Alphabet to $180–190 billion, and Meta to $125–145 billion. Most of that buys NVIDIA GPUs (data-center revenue $75.2 billion in a single quarter, +92%) and custom accelerators co-designed with Broadcom — and every one of those chips ships with stacks of HBM attached.

Two of the three now cite memory prices by name in their guidance — the cleanest public admission that AI spending is feeding memory inflation. Microsoft attributed roughly $25 billion of its capex total to higher memory pricing; Meta raised its range citing “higher component pricing.” When the customers start itemizing your price increases in their own earnings calls, the shortage is no longer a rumor.

The spending is heavy enough to bend even the strongest cash machines. Alphabet's free cash flow — the cash left over after running the business and paying for all that construction — fell 47% year over year in Q1 2026 (it paused buybacks and issued $31 billion of debt); Microsoft's fell 22%. Meta is the counter-example so far — its Q1 free cash flow still grew 20% despite capex up 45%, showing an ad business that can absorb the buildout. Watching hyperscaler capex guidance is watching the memory cycle's demand side: it is the leading indicator for the entire chain.

The bear case: memory has never not been cyclical

Now the uncomfortable history. Memory is the most cyclical major industry in technology, and every prior boom ended the same way: fat margins attracted capacity, capacity arrived together, prices collapsed.

• 1995–97: DRAM prices fell roughly 51%, then another 65%, as mid-90s expansion flooded the market.
• 2018–19: the cloud-datacenter boom reversed; Micron's revenue fell ~30% peak-to-trough and its gross margin compressed from 59% to 27%.
• 2022–23: pandemic over-ordering unwound; DDR4 spot prices fell from over $5 to $1.63, and SK Hynix posted a full-year negative ~28% net margin as recently as 2023.

That last point deserves emphasis: the company now printing 72% operating margins was losing money on every chip three years ago. Skeptics also note that every cycle top in history was accompanied by the claim that “this time is structural.” The strongest-ever bull argument is precisely what the peak of a cycle sounds like — which is why the honest way to hold this thesis is to name, in advance, what evidence would break it.

What's genuinely different: contracts, not spot

The structural case doesn't rest on demand enthusiasm — it rests on a change in how memory is sold. Historically, memory was a spot commodity: identical chips, price set daily, no loyalty in either direction. That is visibly changing, from both sides of the table.

On the seller side, Micron disclosed 16 take-or-pay supply agreements — contracts that oblige the customer to pay for the volume whether or not they end up taking delivery — covering roughly 20% of its DRAM and a third of its NAND volume, carrying about $100 billion of contractual floor revenue backed by $22 billion in customer cash deposits — with floor prices management describes as sitting above peak quarterly margins of any past cycle. Roughly half of Micron's revenue is heading under such contracts. A price floor doesn't repeal the cycle, but it structurally blunts the classic glut-driven collapse for the contracted book.

On the buyer side, TrendForce reports that major cloud providers signed multi-year long-term agreements that restrict suppliers from raising prices on contracted volume — which is why Q3 2026 price growth is moderating even as the shortage persists. Ceilings for buyers, floors for sellers: both sides are paying to convert a spot commodity into contracted infrastructure. That is the de-commoditization argument in signed paper, not assertion.

The 2028 question: what would end it

Here is the tension every investor studying this market should sit with: the same Micron that signed those price floors is also spending harder than it ever has — ~$27 billion of capex this fiscal year, stepping up further next year with more than half the increase going to construction: two new Idaho fabs — chip plants — targeted for mid-2027 and late-2028, a New York site, a second Taiwan cleanroom, and a Singapore packaging plant, plus a multi-year EUV tool agreement with ASML. The rest of the industry faces the same decision, and history says capacity tends to arrive together. TrendForce doesn't expect meaningful new capacity before late 2027 — but it is coming, and clustered capacity arrivals are exactly how every previous cycle died.

So the debate has a date on it. Through mid-2027, supply is essentially fixed and the shortage is contractual fact. The open question is whether the 2028 supply wave lands into demand strong enough to absorb it. Signals worth tracking, in rough order of importance:

1. Hyperscaler capex guidance — the demand side's leading indicator. Any broad guide-down flows through the whole chain within quarters.
2. Contract-price trajectory into late 2026 — moderation under long-term agreements is expected; an outright roll-over while capex is still climbing would be the classic glut setup forming.
3. The contracted book — does the take-or-pay count keep climbing, and do floor prices hold if spot softens? Renegotiations would be the tell that de-commoditization was a boom-time artifact.
4. Fab timelines — whether the 2027–28 greenfield capacity arrives on schedule, and whether makers throttle it if demand wobbles (they have explicitly reserved the right to).

How to research this chain yourself

The memory supercycle is best understood as one connected chain, and each link is a public company you can study: the hyperscalers funding it (Alphabet, Microsoft, Meta), the accelerator designers the money buys (NVIDIA, Broadcom), the memory maker at the center (Micron), and the toolmaker underneath all of it (ASML). Reading any one of them in isolation misses the point; the same dollar flows through all of them — which also means they share the same downside if the spending wave rolls over.

Sources

• TrendForce: Q1 2026 DRAM contract prices +90–95%, industry revenue +81% QoQ · Q2 2026 outlook (DRAM +58–63%, NAND +70–75%) · Q3 2026 server DRAM +13–18%, long-term agreements
• Micron FQ2/FQ3 2026 earnings materials (reported Mar 18 and Jun 24, 2026)
• SK Hynix Q1 2026 results · CNBC coverage
• Samsung Q1 2026 results · Counterpoint Research on the $50.4B memory quarter
• ASML Q1 2026 results (memory 51% of system sales by end-use)
• Alphabet, Microsoft, and Meta Q1 2026 / FY26 Q3 earnings releases and calls (capex guidance, free-cash-flow figures, memory-pricing commentary)
• Tom's Hardware RAM price index (July 2026)
• Apple confirms price increases due to the RAM shortage (Tim Cook, WSJ interview, Jun 17, 2026) · Forbes on Mac price increases