Newsletter/AI Infrastructure

Micron's $41 Billion Quarter and the Memory Supercycle

The headline is the record. The tell is buried in the DRAM line: bits barely moved, price jumped roughly 60%. That gap is the whole thesis.

By Upcurious · · 2,728 words

Micron's $41 Billion Quarter and the Memory Supercycle

Micron reported fiscal Q3 after the close on June 24, and the numbers are hard to overstate. Revenue of $41.5 billion, up roughly 346% year over year against expectations near $35 billion. Adjusted gross margin of 84.9% — a company record, and a number you almost never see in what was a cyclical commodity a few years ago. Adjusted EPS of $25.11 versus $20.60 expected. And a Q4 guide of $50 billion, plus or minus a billion, against a street near $43 billion. The stock jumped double digits and dragged the semis with it.

The most important figure is not the headline. It is one line in the DRAM segment: bit shipments rose only low single digits quarter over quarter, while average selling prices rose in the low 60s percent. Flat volume, price up sixty percent. That gap — barely more memory shipped, vastly more money collected — is what a physical bottleneck looks like in an income statement, and it is the whole argument for why this cycle is different.

You did not need the filing to feel it. A day later, Apple raised prices across much of its lineup and blamed memory outright, with Tim Cook calling the shortage a "hundred-year flood" and something he had "never seen anything like in any area in over 40 years." Same crunch, opposite ends: the supplier books record margins while the most powerful buyer on earth passes the cost to consumers.

The Earnings, Quickly

MetricQ3 FY2026Context
Revenue$41.5B+346% YoY; vs ~$35B expected
Adjusted gross margin84.9%Company record
Adjusted EPS$25.11vs $20.60 expected
DRAM revenue$31.3B76% of total; bits +low-single-digit, ASP +low-60s% QoQ
Data center revenue$25B~60% of total revenue
Enterprise SSD$5B20% of data-center revenue
HBM>$1B HBM4 already shippedBooked through CY2027, demand into 2028
Q4 guidance$50B ± $1Bvs ~$43B expected

Three things matter beyond the beat. The roughly 85% adjusted margin is a software-like number — you do not earn it selling a commodity that clears at spot. The mix has flipped: data center was $25 billion, about 60% of the company, so memory is now an AI-infrastructure business that still happens to sell some chips into PCs and phones. And the visibility is new: with HBM sold out into 2028 and roughly $100 billion in customer agreements, the oldest knock on memory — that you cannot see past two quarters — no longer holds. Micron can now see past two years.

Two Markets, Opposite Physics

The question the print forces is whether this is durable or the top. The answer turns on a fact the headline hides: "the memory market" is now two markets with opposite physics, fused in one income statement. One is shrinking in units, one is exploding, and the exploding one sets the price.

DimensionAI / hyperscaler memoryConsumer memory
ProductsHBM, high-capacity server DRAM, enterprise SSDPC DRAM, mobile DRAM, consumer NAND
Price sensitivityVery inelasticElastic
Current behaviorVolume flat-to-growing, spend up multiplesVolume falling double digits
Pricing mechanismMulti-year LTAs, pre-sold, no spot marketSpot + short contracts, price-sensitive
Who absorbs the costHyperscalers, from operating cash flowConsumers and OEMs, who push back
OutlookSold out into 2027+; pricing firm overall, though HBM itself faces price competitionDemand destruction, accelerating

On the consumer side, demand destruction is real. Apple raised Macs 15–20% and iPads 15–25%, sparing the iPhone only "for now." When even Apple — first in line for allocation — has to pass memory costs through, you are watching the elastic end in real time. It is squeezed enough that it asked Washington for clearance to buy DRAM from CXMT, China's largest memory maker and a company on the Pentagon's blacklist. That is what desperation for supply looks like. Everyone below Apple has worse options: raise prices, cut specs, or eat margin. Many are cutting specs — some flagship phones shipped this cycle with no RAM upgrade over the prior generation, a first in the smartphone era.

Micron's record margin and Apple's price hike are the same event, billed to different people.

On the AI side, none of this is happening. Hyperscalers are not trimming orders; they are signing multi-year agreements and locking 2027 supply during 2026, treating allocation as more urgent than price. Their willingness to pay is anchored not to the cost of memory but to the return on the cluster it sits inside — and that return still dwarfs the memory bill, even after this quarter's price jump.

Why the Blowout Happens: Mix

If half your market is in demand destruction, how do you guide to $50 billion? Because the demand destruction is a volume story, not a dollar one. TrendForce puts the 2026 memory market near $552 billion — DRAM ~$404 billion (+144%) and NAND ~$147 billion (+112%) — more than double 2025. Data center is now the majority of it and the source of most of the growth. Consumer revenue is up too, but only because price is overwhelming falling units, not because anyone is buying more. HBM, for its part, led on growth rate — ~$17 billion in 2024 to ~$35 billion in 2025 to ~$60 billion this year — yet at roughly 15% of DRAM revenue it is still a minority of the AI memory bill: even tripling, it did not outgrow the conventional DRAM around it.

SegmentDirection in 2026Price sensitivity
Data center / AI memory (HBM + server DRAM + enterprise SSD)The majority — consumes ~70% of high-end DRAM; most of the growthInelastic
Consumer (PC + mobile)Large but shrinking share; dollars up on price, units down double digitsElastic
Auto / industrial / otherNiche; modest growthModerate

Micron's own quarter is this table in miniature: data center was $25 billion of $41.5 billion, about 60%. Consumer is shrinking in units while AI grows on both price and volume — which is why "PCs are weak, so memory is rolling over" reads the wrong instrument. Consumer weakness is a unit story and a side effect of this cycle, not a leading indicator of it.

Why the Whole Stack Rises, Not Just HBM

If HBM is only a sliver of the bill, what makes up the rest? We had large datacenters long before AI — but never ones this memory-hungry. Look inside one AI rack.

MemoryGB200 NVL72 (AI rack)Conventional 2-socket server
HBM13.5 TB—
System DRAM (LPDDR5X / DDR5)17 TB~0.5–1.5 TB
Total fast memory~30.5 TB~1 TB

Two things jump out. Even in NVIDIA's flagship rack, there is more conventional DRAM (17TB of LPDDR5X) than HBM (13.5TB) — the HBM gets the headlines, the system memory around it is the bigger pile. And against a traditional server, the AI rack carries roughly 30x the memory and adds an entirely new HBM tier on top. Multiply by every rack in every AI datacenter and the conventional-DRAM demand alone dwarfs the old cloud.

Two more forces finish the job. Every wafer redirected to HBM — at three DRAM dies to one — starves conventional DRAM supply, so the existing base reprices higher without a single extra bit shipped. And AI has started bidding for the low-power LPDDR once reserved for phones, dragging even mobile-class memory into the crunch.

The clearest proof is in the margins. HBM's price per gigabyte actually fell over the past year — HBM3E drifted from roughly $17–20/GB in early 2025 toward $13–17 entering 2026 as Samsung discounted to win sockets — even while commodity DRAM prices exploded. The result: TrendForce found that in Q1 2026, plain server DDR5 — a 64GB RDIMM — became more profitable per wafer than HBM3E, enough that the makers began steering capacity back toward it. The single most AI-specific product on the planet was, for a quarter, out-earned by a commodity DIMM, while its own per-GB price was falling. That HBM softness was mostly a 2025 dynamic — as makers shifted wafers to DDR5, HBM supply tightened and TrendForce sees its pricing firming again into 2026 — but it makes the point: HBM is the spark, not the fire, and right now the commodity layers are quietly carrying a surprising share of the profit.

The Triopoly Holds the Pricing

This is not the fragmented commodity market of past gluts. Three companies control roughly 93% of DRAM, and the order just changed: SK Hynix overtook Samsung as the largest DRAM maker for the first time since 1992, on the strength of memory — the product driving the cycle.

Company~DRAM shareHBM position
SK Hynix~34% (now #1)~62% — clear HBM leader
Samsung~33%~17% — lost ground, chasing
Micron~26%~21% — passed Samsung for #2

They have also moved customers from fixed-price contracts onto variable-price long-term agreements, where the buyer now bears the pricing risk — the mirror image of the buyer-friendly 2022–23 cycle. Differentiated product, concentrated supply, self-funding buyers: none of those were true in the cycles that ended badly. That is the durability the record margin is quietly reporting. The one crack is CXMT — the ~5% Chinese maker customers like Apple are now lobbying to qualify. For commodity DDR it could one day be a pressure valve; for HBM it is years away and, for now, blacklisted.

What It Means for the Trade

The strange thing about memory stocks is that they look cheapest exactly when earnings peak. Micron carries a rich trailing multiple but only single digits on forward FY27 estimates — and that low forward multiple is not a discount. It is the market pricing in the down-cycle it assumes is coming. So the trade is not "memory is cheap." It is a bet on duration: if the supercycle runs structurally longer than a normal cycle, today's "peak" earnings are not the peak and the compressed multiple is the opportunity. If it is a normal cycle, the multiple is right and you are buying the top.

That framing sorts the ways to play it.

VehicleThe readThe catch
Micron (MU)US pure-play, most liquid, HBM #3 and gainingPriciest of the three; most crowded US trade
SK HynixPurest HBM bet — ~62% HBM share, sold out, cheapest of the threeKorea-listed for now; concentration risk if share slips
SamsungThe laggard/turnaround; cheap relative to its breadthExecution bet, diluted by other businesses
NAND (SanDisk, Kioxia)The under-owned second derivative; enterprise SSD demandSmaller, more volatile, less HBM exposure
Roundhill Memory ETF (DRAM)The "I believe the cycle, not a name" basketYou own the laggards too

All of them have already run hard — MU up roughly 270% year-to-date, SK Hynix ~295% (it just passed Samsung to become Korea's most valuable company), Samsung ~200%, and SanDisk a staggering 600%-plus as the Nasdaq-100's top performer — which is its own warning: a cyclical at record earnings with nearly everyone already long and analysts at near-uniform Buy. The honest risk is that these are still cyclical stocks at record pricing. The down-cycles have been fast and brutal, and the multiple usually compresses while earnings are still rising — the market sells the peak before it arrives. None of the triggers has printed yet. But when one does, the cheap forward multiple stops looking cheap fast. (This is analysis, not advice — position sizing is yours.)

So What Actually Ends It

Durable is not permanent. The cycle does not depend on PCs recovering — that is the point. It depends on the inelastic AI buyer staying inelastic. So the things that end it are not the loud, visible risks. They are three quieter ones, on three clocks.

Now — demand could prove softer than it looks. The bull case rests on hyperscaler willingness to pay being anchored to AI returns. That holds only as long as the returns keep showing up. The first quarter a hyperscaler guides capex flat-to-down year over year is the first genuine crack.

Next — the leverage is hidden one layer down. Hyperscalers fund capex from cash, but the system around them does not. A slice of GPU demand sits with neoclouds financed by debt secured against the chips, and a slice of AI revenue is AI companies selling to other AI companies. If that layer cracks, it cascades backward into HBM demand without any hyperscaler changing its mind.

Later — 2027 is when supply and demand could collide. All three makers are building fabs for 2027 volume, the classic glut setup. The counter, which Micron leaned on this quarter, is that the orders already exist before the wafers do — HBM is booked years out. That pushes the risk out; it does not erase it. The cycle breaks only if demand decelerates and the new capacity lands into that deceleration. Two things have to go wrong together.

And the risk that is not a risk: falling PC volumes. The next weak PC quarter is not the warning; the next hyperscaler capex guide-down is.

What I'm Watching

Hyperscaler capex guidance. Watch the big four's year-over-year direction above all else — it leads the cycle, and the suppliers roll over only after the buyers do.

AI revenue disclosure quality. As long as the hyperscalers keep quantifying the revenue AI generates, the inelastic-buyer story holds. The quarter they stop putting a number on it is the quarter to assume it got worse.

The booking horizon. Bookings that stop extending past 2027, or fab timelines pulled forward, are the two sides of the capacity-collision warning.

Neocloud solvency and the China wildcard. A neocloud default would expose the hidden leverage; Washington clearing Apple to buy CXMT would mark the first real, if small, dent in triopoly pricing power.

This is part of an ongoing series on AI infrastructure economics. Earlier pieces covered the physical bottleneck sequence, the memory demands created by agentic work, the model race as an inference-demand engine, and why cheaper intelligence keeps pulling more computation into reach.

For more on the AI stack and where value flows, visit theupcurious.com.

Sources

Get the next analysis

AI infrastructure, model economics, and agentic software, delivered through Substack.

Subscribe

Related research

Model Economics Fable 5, GLM-5.2, and the Shrinking Frontier Premium

The capability gap is holding. The market willing to pay for it is narrowing.

 AI Infrastructure AI Is Still a Compute Story. For Now.

Demand is real. Spending is extreme. And the missing pieces aren't what most people think.