What Ends the Cycle First Is Demand, Not Supply — Peak Margin Is the Starting Line of Mean Reversion

Micron posted its highest quarterly gross margin ever, 84.9%, and in that very earnings call said that price increases would "moderate meaningfully." SK Hynix's operating margin (OPM) cleared 72%, Samsung's chip margin topped 70%, and supplier DRAM inventory had dried to 2–3 weeks — down to the level of the 2018 cycle trough. This is the peak of strength. Yet in that same quarter the rate of price increase had already begun to roll over. Both are true. So is this the peak, or the mid-to-late stage?

My read is mid-to-late. The slowdown has begun but the shortage persists, and the inflection that ends the cycle comes first from the demand side — not the supply side that conventional wisdom points to. That said, the fundamentals being mid-to-late is one thing; the inflection that matters to investors (the deceleration in price increases) having already been priced through the fast channel is another. Mid-to-late does not mean the runway is long. Let me split out why I read it this way, with the data.

The Slowdown Has Begun, but Not the Decline

"The rate of increase has rolled over" and "prices are now falling" are different events. Whether prices rise or fall (the first derivative) is distinct from whether the size of that rise grows or shrinks (the slope of the rate of increase, the second derivative), and what has happened now is the latter.

Conventional DRAM contract prices rose +93–98% quarter over quarter in Q1 2026 (realized). That is a rate of price increase, not a rate of industry revenue growth. Constrained by shipments and mix, revenue grew only +81%. The Q2 forecast is +58–63% — the size of the rise has shrunk. That said, a deceleration in the quarterly rate off the +95% range falls out of the base effect alone. So the weight of the signal lies not in the number but in the supplier's own admission. Micron nailed "a meaningful moderation in price increases" into its guidance, and if it were merely a base effect there would be no reason to spell it out that way.

But +58–63% is still a large plus. Prices themselves climb much further. NAND is +70–75% in Q2, mobile LPDDR5X +78–83%. Far from easing, the shortage has spread even to older parts: legacy DDR2 contract prices are forecast to jump +35–40% in Q3. The reason conventional DRAM is drying up too is HBM (high-bandwidth memory, several DRAM dies stacked vertically and mounted on an interposer beside the accelerator). For the same capacity HBM uses roughly 3x the wafer area of DDR5 (HBM4 ~4x), so even though it is a minority in bits it eats wafers in bulk. As a result the global DRAM wafer cannibalization that was ~19% in 2025 climbs to about 20–25% (we read it as a range, given definitional variance across institutions), squeezing conventional DRAM capacity along with it.

Table: Prices still rise, but the size of the rise has shrunk. Slowdown (2nd derivative) and decline (1st derivative) are different events, and inventory and lead time keep prices from going negative (−) for a while. Source — TrendForce price forecasts (2026-03–06) · inventory · lead time (2025 Q3–first half 2026).

The onset of slowdown and the persistence of shortage are not contradictory. That is the definition of mid-to-late.

The Strength Held Up by Margins, Inventory, and Locked-In Demand

Margins at an all-time high and inventory at an all-time low mean prices cannot fall for a while. SK Hynix posted Q1 revenue of 52.58 trillion KRW (+198% YoY) and operating profit of 37.61 trillion KRW, with OPM leaping from 58% to 72% in a single quarter (net margin 77%, an unusual quarter in which large non-operating gains pushed net income above operating income). Samsung's semiconductor (DS) operating profit was 53.7 trillion KRW, 48x the prior year — a single quarter's operating profit exceeding the entire company's operating profit for all of last year. Micron's gross margin of 84.9% was an all-time high, with next-quarter guidance at ~86% as well. Logic (foundry) came along too. TSMC's gross margin was 66.2%, it raised its annual growth outlook to '30%+', and its supply constraint runs through 2027.

On top of this, demand has been locked in by contract. Micron secured 16 take-or-pay agreements (contracts that pay for the committed volume whether or not you take it), of which 14 carry minimum contracted revenue of about $100B, running 2026–2030 and covering about 20% of DRAM volume (with HBM as its strategic core) and about a third of NAND. Cash deposits (advance payments) alone run about $18B, and including related financing commitments the figure exceeds $22B. Demand visibility for this volume, HBM included, is in effect tied up through 2030. Combined with 2–3 weeks of inventory and 30–40 weeks of lead time, the downside cushion is thick. But this cushion is thick against a supply-driven price decline. As we'll see, this cycle's trigger is pulled first on the demand side, and when orders tighten the 2–3 weeks of inventory rebuild quickly (as they did in 2022), thinning the cushion. In other words, what the cushion blocks is 2026; the 2027 demand bill routes around that cushion.

Table: Margins at a peak across the three memory makers and foundry too. Source — each company's IR (SK · Samsung 1Q26 2026-04, Micron FQ3 2026-06, TSMC 1Q26 2026-04).

But a peak is no grounds for comfort. The opposite, if anything. The operating leverage that produced a 72% OPM runs in reverse at the same speed once prices turn. An all-time-high margin is the starting line for mean reversion (the tendency of unusually high values to return to a long-run average), and because part of that peak margin is non-recurring — like the non-operating gains seen above — the reversion could be larger. That the strength is real and that the strength is a peak are true at the same time.

What Ends the Cycle

The textbook closers of a memory cycle are two. Either the three makers expand capacity simultaneously and bring on a glut, or China catches up and crushes prices. Both are supply-side. This time, though, the order is reversed — because both supply-side variables are late variables.

Supply amplification comes late to begin with. But this applies only to new greenfield fabs. A greenfield takes more than five years to build, so Samsung's Pyeongtaek P5 reaches mass production in the second half of 2028, SK's Yongin Y1 takes equipment delivery in Q2 2027, and high-volume production is impossible before late 2027 to 2028. That said, raising bits per wafer through node transitions and shrinks at existing fabs is a faster lever measured in quarters, so it does not rule out bit-supply growth before 2027.

China is a different story again. CXMT's DRAM share jumped from 3.97% in Q2 2025 to 7.67% in Q4, and its wafer output rose from 100,000 to 290,000 sheets a month. Its target is the low end — the very legacy (DDR2) segment we read as a strength signal. So China has to be split in two. In legacy, state-subsidized capacity expansion is an early-supply risk that could press older-part prices down faster than we assume; in HBM, it is the latest variable of all. HBM is blocked because the core-die node is constrained by the EUV lithography cutoff, and on top of that — the bigger wall — it is blocked again by TSV stacking and packaging yield. Even HBM3 mass production is hard to achieve within 2026. So long as export controls block EUV, China's HBM pursuit stays trapped at the low end, and the variable that would crush HBM prices is pushed out past 2027. The variable the market fears most is in fact the latest to arrive.

What comes first is the demand side. The four big US hyperscalers (Amazon, Microsoft, Google, Meta — the operators of hyperscale data centers) are forecast to spend about $725B on AI capex (capital expenditure) combined in 2026, +77% from ~$410B in 2025, with 2027 set to top $1 trillion. That absolute scale itself cannot continue indefinitely.

The real trigger is pulled by prices themselves. The memory price surge by itself pushes memory's share of a hyperscaler's AI hardware budget from about 30% in 2026 to about 36% in 2027 (a single estimate). A big buyer whose budget is being eaten away has no choice but to tighten its next order. That is, this supercycle carries its own thermostat — a self-extinguishing loop in which the stronger prices are, the faster they pull the demand bill forward.

An auxiliary signal that accelerates this loop is doubt about returns. One bear-case analysis holds that about 95% of enterprise AI projects fail to produce measurable ROI (return on investment); in the same source are an estimate that extending GPU depreciation from three years to five years books about $176 billion less in expense over 2026–28, and the circular-financing suspicion that CoreWeave's $99.4B backlog dwarfs its $2.1B in quarterly revenue. That said, the 95% is an enterprise adoption statistic, not a return rate on hyperscaler capex. So this should be read not as direct grounds for a big buyer to cut orders, but as a demand-side sentiment signal that their customers' demand is cooling.

Table: The textbook puts the supply side (2·3) first. The data puts the demand side (1) ahead. Source — hyperscaler guidance (2026 Q1) · new-fab timelines · CXMT share.

The inflection comes not because supply collapses but because demand receives the bill.

Why This Time Is Different, and Why It Still Ends

Past cycles all ended the same way. Demand saturation and simultaneous expansion by the three makers blew up together. From the 2017–18 peak to the trough took 6–7 quarters, 2019 revenue fell -23%, and the 2021–22 boom gave way to a collapse from late 2022 into 2023 — Samsung cut production by as much as ~50% and SK posted a -28% net margin in 2023. This time the structure is different. A five-year take-or-pay ($100B-class, with HBM as its strategic core) has tied up demand visibility through 2030, and the oligopoly discipline of the three makers (SK overtaking Samsung in revenue for the first time in 26 years, and signs of allocating DDR5 profit ahead of the HBM4 ramp) restrains supply disorder.

The permanent-bull case starts here: "With the lock-ins, the discipline, and the lowest inventory ever, this cycle has no end." The data backs half of this claim. Volume locked through 2030 is clearly a force that delays the inflection, which keeps a later-than-2027 scenario (BofA 2030) alive. But what a lock-in guarantees is a floor margin, not the constancy of volume and price. Historically, long-term contracts in a downturn have flowed into renegotiation and deferral rather than being enforced against the largest customer. So the breaking point lies in a different volume pool. Volume already locked delays the descent, but the volume and price of new and maturing contracts get marked down once hyperscalers start scrutinizing returns. Lock-ins and discipline only flatten the slope of the descent; they cannot abolish the descent itself.

The opposite extreme — the imminent-collapse case — is dismissed by the same data. Prices are still positive, inventory is at rock bottom, and supply amplification is after late 2027. So I stand at mid-to-late. The slowdown has begun but the shortage holds it up, and the inflection comes first from the demand side, not supply.

So When? — As a Distribution, Not a Point

Trying to nail the peak to a single point is a fool's errand. The consensus itself is split. Morgan Stanley sees a 2027 peak (volume entirely sold out through 2027, a slight slowdown in 2028); Goldman sees supply shortage through 2027; BofA sees strength through 2030. Two of the three cited (Morgan Stanley, Goldman) cluster near 2027. The peak is 2027, and BofA's 2030 is an upside tail risk. My base case is also a 2027 peak followed by a 2028 turn — not a point estimate but a weighting read directly off this distribution.

So what does one watch? The signals that read position rather than direction are these.

Table: A checklist that reads position, not direction. The basis for each signal maps 1:1 to the [f#] ledger in the text.

Supply constraint is not only in price. Another bottleneck (binding constraint — the limit that hits its ceiling first and caps total output) that sets the upper bound on GPUs is packaging and CoWoS (TSMC's 2.5D advanced packaging · interposer) allocation, an axis covered in Packaging Sets the Ceiling on AI Accelerators: Why the Back End Became the Center of Gravity. And how much of this cycle is already priced into share prices (valuation) is outside this piece's scope. The structure by which the KOSPI has become a derivative leveraged to memory is covered in SK hynix Has Passed Samsung. The KOSPI Is Now Two Companies.

So return to the surface. Spectrum-analyze that surface — supercycle, all-time-high margins — and the depth reveals three things: the rate of price increase (the second derivative) has already rolled over; the strength carries a self-extinguishing loop in which the stronger prices are, the faster they pull the demand bill forward; and so what ends the cycle first is demand, not supply. Read that depth as debt, and the question splits into who gets the bill, and when. The cost heads to three places: the investor who extrapolated the peak as a trend and bought at the top, the three makers at a peak facing mean reversion, and the hyperscalers who scaled up capex of uncertain return. The contract-price slowdown of +93–98% in Q1 2026 (decelerating to +58–63% in Q2 2026) is a fast channel, so it is already priced in; the hyperscalers' return burden is a slow channel and has not yet moved — that lag is the mechanism that actually drives this cycle. No single party corrects the timing for you. That the consensus is split between 2027 and 2030 is the evidence, so the work of dating the peak is left to the reader. And there the question shifts — from who carries it, to what to prepare for and what is predictable. The peak is 2027, BofA's 2030 an upside tail, so bet the peak as a distribution, not a point, and read your position within that distribution off the checklist above: the second derivative of price, inventory and lead time, capex guidance, take-or-pay maturities, new-fab timelines, the consensus spread. The one thing to guard against most is extrapolating the fast channel as a trend. The first-in-line bill goes to the investor who falls for that — fundamentals being mid-to-late does not mean the trade has runway left.

When the headlines shout "supercycle, all-time-high margins," what divides them is one thing: is it rising, or is the size of the rise shrinking? Whoever reads that distinction first gets a first hand on the correction the market has deferred.