Memory cyclicality - the supercycle that still has a cycle

$MU just crossed $1T in market cap. The weekly RSI is 86.2. The stock is +1,873% off its 252-week low. The forward PE is 8.76×.

That last number is the one that anchors the bull case - "Micron is still cheap" - and it is also the most dangerous number in the sector. Every prior cycle peak in memory looked exactly like this: a vertical chart, peak operating margins, and a single-digit forward PE that proved to be mid-cycle in retrospect and late-cycle in reality. The multiple is cheap because the earnings are about to roll over, not because the stock is undiscovered.

This article is the framework for telling those two states apart. The conclusion up front: memory remains cyclical. AI has dampened the cycle, not eliminated it. The drawdown when this cycle turns is more likely 30-50% than 50-75%, but it is still a drawdown that defines whether the trade is a winner or a flat-to-loss after round-trip. Position sizing has to assume the cycle is real until the structural data says otherwise - and most of the structural data has not changed.

Why memory is the most cyclical semiconductor sub-sector

Three structural features stack to make memory the most cyclical part of the semi tape. No other sub-sector has all three.

One - pure commodity output. A DRAM bit from Samsung is fungible with a DRAM bit from SK Hynix or $MU. There is no per-chip IP moat. Spot prices clear like a soft commodity. This is the opposite of logic semis ($NVDA's architecture, $ARM's IP, $AVGO's custom ASIC business) where customers cannot substitute one supplier's silicon for another without re-writing software stacks.

Two - multi-year capex lag against quarter-by-quarter demand. A greenfield DRAM fab is 3-5 years from groundbreak to high-volume manufacturing. Demand can shift in a single quarter - smartphones to PCs to servers to AI accelerators. The supply curve is rigid in the short term; the demand curve is elastic. This mismatch IS the cycle.

Three - three-player oligopoly with prisoner's-dilemma incentives. Samsung, SK Hynix, and Micron control ~95% of DRAM globally. When prices rise, all three rationally over-invest. When supply lands, all three rationally suffer. Discipline holds until one player breaks it - historically the Korean fabs have broken first, with Samsung in 2018 and Hynix in 2022 each adding capacity at the wrong point in the cycle.

Logic semis ($NVDA, $AVGO, $ARM) are less cyclical because IP is sticky, switching costs are real, and the upstream bottleneck (TSMC) is one supplier with capacity discipline. Memory has none of those buffers. It is the part of the semi tape that most resembles oil refining: high fixed cost, commodity output, brutal cycles.

The historical record - every up-leg ends in a drawdown

The empirical record on $MU's stock-cycle history makes the pattern unambiguous.

Three observations carry the rest of the framework.

First, every up-leg in the last decade ended in a 50-75% drawdown within 12-18 months of the peak. Not "if" - when. The bust phase is not an outlier; it is the back half of the structure. A trader who held $MU through any of the prior three cycles round-tripped most of the gains unless they exited within the late-cycle phase.

Second, the current up-cycle (+1,851%) is larger than the prior three combined. Either AI is a true structural step-change that bends the historical pattern, or this is the biggest mean-reversion setup the sector has ever produced. There is no third interpretation that does not require explaining away the magnitude.

Third, the drawdowns happen while the forward PE looks cheap. In 2018, $MU peaked at forward PE ~5×. The stock fell 57% over the next year. The multiple stayed "cheap" the whole way down - earnings collapsed faster than price did. This is the most important and most counter-intuitive lesson in memory: single-digit forward PE at the top of a cycle is the signature of a cycle peak, not an undervaluation signal. P expands relative to E during busts because E falls first.

The 5-phase cycle template

Each of the cycles in the table above followed the same five-phase shape. The shape is set by the lag structure - fast demand, slow supply - and is independent of which end-market triggered the cycle.

The 2017-2018 cycle is the closest historical analog to today's setup: a data-center demand thesis that the tape treated as structural, that turned out to be a real but compressible cycle. The 2018 peak came with margin commentary identical in tone to today's commentary, and with a forward PE near 5× that re-rated downward (and the stock down) for the next twelve months even as the multiple stayed numerically low.

Where Micron sits right now in the template

Map current state to the phase markers and the picture is unambiguous.

Eight of ten signals read late or peak. The two that have not triggered are (a) spot DRAM pricing rolling over, and (b) $MU issuing equity at the top - both of which have historically been the final confirmations of a cycle turn. $MU issued secondary at $60 in 2018 and at $90 in 2021, in both cases within weeks of the cycle peak.

The bull narrative leans on HBM lead times remaining long and HBM allocations being "sold out." Both are true today. Both were also true at the 2018 peak - server lead times were extended right up until they were not. Sold-out is a current-state observation, not a forward demand signal. It tells you the supply-demand cross is tight now. It tells you nothing about whether NVDA's inventory or hyperscaler capex digestion will leave that cross tight in two quarters.

What HBM and AI actually change - and what they do not

The honest assessment of the AI structural-break thesis is that it changes part of the cycle dynamics, not all.

What is genuinely different this cycle:

• HBM contracts are LTAs, not spot. Long-term agreements with NVDA, Google, Meta, Microsoft, and the other large customers give pricing visibility through 2026. The quarter-by-quarter margin volatility that defined prior cycles is dampened for the HBM portion of the book.
• Customer concentration is healthier. The current buy-side is a small set of credit-strong hyperscalers and AI accelerator OEMs. Prior cycles depended on 200+ smartphone OEMs with thin margins and weak balance sheets - easier to soften, easier to default.
• HBM is a different product. 3D-stacked, dramatically more value-add per wafer, with real IP differentiation across the three suppliers (SK Hynix has the lead, Samsung is catching, $MU's 12-Hi is shipping). It behaves less like vanilla DRAM and more like a specialty memory product with switching costs.
• Capex per wafer is much higher for HBM, so the supply response is slower and more expensive. This partially discourages the "all three over-invest simultaneously" pattern that broke prior cycles.

What has not changed:

• Vanilla DRAM and NAND are still commodities. Roughly 70% of $MU's bit output is not HBM. That portion continues to cycle exactly as it did in the prior decade.
• Supply will respond. Samsung committed to $200B+ in fab investment over 20 years. SK Hynix's M16 fab is ramping. $MU's Idaho facility - the first 1γ DRAM in the US - starts HVM in 2026. All three players are racing to expand HBM capacity simultaneously. By 2027-2028, that supply lands.
• Demand can still shift. If GPU shipments slow - through NVDA inventory build, hyperscaler capex pause, ASIC substitution, or end-customer software efficiency gains - HBM allocations free up. The 2018 cycle peaked when hyperscalers paused capex for two quarters. That was sufficient to turn the cycle.
• HBM4 transition risk. The HBM3 → HBM3E → HBM4 transition is the next major specification step. Prior product transitions have coincided with cycle turns: HBM2 → HBM3 in 2019 followed the 2018 peak. The HBM3E → HBM4 transition is expected H2 2026 to 2027 - exactly the window where the current cycle is likeliest to peak.

The net is that AI partially decommoditizes the ~30% of $MU's bit mix that is HBM. The other 70% is still cyclical commodity DRAM and NAND. The cycle is dampened, not eliminated. Drawdowns this time may be 30-50% rather than 50-75% - still violent enough to define whether the trade is held correctly or destructively.

For the structural background on HBM specifically - yields, packaging constraints, and which companies are positioned where - see HBM is the tightest bottleneck in the AI cycle.

The eight leading indicators that turn before the stock

The cycle does not announce itself in stock price. It announces itself in the underlying supply-demand data, with predictable lead times before earnings, and longer lead times before the multiple compresses. These eight indicators, in rough order of how early they turn, are the watchlist for anyone holding the trade.

The first three are the early-warning system. As of today, spot DRAM remains firm, HBM allocations remain sold out, and lead times remain long. None of the cycle-rolling signals have triggered. That is why the directional long trade is still on. It is also why the cycle-turn signals matter - they are the exit triggers, and the historical record says they print 1-2 quarters before the stock peak in earnings, and 4-6 months before the peak in price.

The QA memory bubble residualized correlation runs 0.71 against the SPY-stripped basket - see the memory bubble page for the full taxonomy and the live correlation refresh. When the cycle turns, the entire bloc - $MU, SK Hynix, Samsung, $SNDK, $WDC, and the second-order semi-equipment names $LRCX, $AMAT, $KLAC - moves together. Position-stacking inside the bloc multiplies the turn-risk, which is why the cycle-watch matters more than any single-name story.

What Aschenbrenner's straddle is telling you

The most informative positioning signal on Micron right now does not come from a sell-side note. It comes from Situational Awareness LP's Q1 2026 13F-HR. Leopold Aschenbrenner - author of the AGI-imminence essay and the most public AI-supercycle bull on the tape - disclosed:

• $584M MU puts (rank 8 in the book)
• $422M MU calls (rank 12 in the book)

Net delta on this pair is near zero. The position is long volatility on $MU, not directional. He is paying premium to be hedged in both directions because he believes - with conviction sized to $1B+ notional - that $MU prints big either direction in the next two quarters.

Translation: the smartest publicly-disclosed memory bull on the tape is not net long the stock. He is long the tails. The up tail (HBM allocation tightens further, prices rip higher), and the down tail (cycle turns, peak earnings re-rate, 40%+ drawdown). The body of the distribution - the "stock chops sideways while everyone is right about AI" outcome - he is implicitly underweight.

For a directional long position, that signal carries one operational implication: size for the down tail. Aschenbrenner has the put leg as his hedge. A retail equity long does not. The sizing of the position has to do the work the puts would do in a hedged book.

Full breakdown of the Aschenbrenner positioning: Aschenbrenner Q1 2026 - the chip-short pivot.

Trading template - what late-cycle memory actually looks like

Translating the framework into a position-management plan, the historical pattern supports the following playbook for anyone holding $MU or the broader memory bloc today.

While the cycle holds (now):

• Position size at roughly half what a mid-cycle entry would risk. Weekly RSI 86 is not a normal-size entry zone, regardless of how clean the fundamentals look.
• Hard stop-loss, ATR-based. The Fib structural support at $339 is a −63% drawdown. That is not a stop - it is a portfolio destruction event. Use 1.5× daily ATR ($55) as the stop distance, around $836 from current $919.
• Trail the stop up. Every 5-7% of additional upside, ratchet the stop. Late-cycle parabolas have steep tops; the trailing logic captures the upside without giving back the move.
• Trim into vertical days. If the position is up 10% in 3 sessions, sell 25%. The volatility works for you on the way up; it will work against you when the cycle turns, and the trim is what funds the survival of the remaining position through the turn.

Cycle-turn exit triggers (these are the action signals, not the news):

1. Spot DRAM rolls over −10% from peak across 4+ weekly prints
2. $MU underperforms $SOXX for 4+ consecutive weeks
3. NVDA or hyperscaler capex guidance softens on an earnings call
4. $MU issues secondary equity or convertible debt at ATH
5. HBM commentary on the $MU earnings call shifts from "sold out" to "managing allocations" to "balanced"

Any single trigger from this list should reduce position size by 33-50%. Two triggers within 30 days should close the position. The historical record says triggers come in clusters within 4-8 weeks before the cycle peak in price.

Post-peak (the bust phase):

• Do not try to short the first leg down. Memory squeezes are brutal on the way down; short interest covers fast on any pause.
• Wait for the daily SMA50 to break and fail to reclaim on retest. That is the technical confirmation that the cycle has turned, not just a normal pullback.
• Short via puts rather than equity. Volatility expands on the bust, and put premium becomes worth paying. The asymmetric payoff favors defined-risk structures.
• Cover when forward PE expands past 25×. Counter-intuitively, that is the trough signal: price is holding while earnings collapse, the multiple expansion is the bottoming process.

Next cycle re-entry:

• Wait for $MU to trade below book value. Book is currently $64, and the 2022 trough breached it at $48 per share. The cleanest historical re-entry points have been below-book.
• Wait for analyst targets to round-trip 50%+ below the prior peak price.
• Wait for "memory is dead" headlines. Historically, the cleanest long entries are when sell-side capitulates, not when they upgrade.

The two-scenario probability frame

Honestly assessing the next 12 months from current price, three scenarios cover most of the distribution:

Expected-value math at $919 entry:

• Scenario A: +30% to +63% × 35% ≈ +16%
• Scenario B: −10% to +10% then −30 to −40% × 45% ≈ −11%
• Scenario C: +9% then −55% × 20% ≈ −10%

Probability-weighted: roughly −5% over 12 months at current entry, before any active trading.

That number is what the cycle is asking you to respect. The trade is positive-EV only if (a) the position is actively managed against the eight indicators above and exited on the first 1-2 cycle-turn signals, or (b) the entry is on a pullback to $750-800 where the scenario math shifts back to net-positive. Buy-and-hold from $919 has negative expected value against the historical base rate.

What this means for the bubble allocation

The memory bloc remains the cleanest validated bubble in the QuantAbundance taxonomy. The 0.71 residualized correlation is the tightest among the 12 sub-themes, and the fundamental thesis - AI-driven HBM demand, sold-out allocations, FY26 op-income +684% YoY for $MU - remains intact. That is why the bloc is on the platform's primary leader list, why $MU ranks #1 in Below SPX Forward PE at 6.0× and #3 in Semis by Forward PE at PEG 0.08.

What changes at late-cycle is not the bubble's validity. It is the position-management profile. A bubble that is correctly identified and structurally supported can still produce −40% drawdowns when its underlying cycle turns. The bloc-level identification (residualized correlation, bubble membership, ETF flow) tells you which names move together. The cycle framework tells you when to be in the bloc and when to trim.

The two layers are complementary. The bubble taxonomy gives you the what. The cyclicality framework gives you the when. Memory needs both because it is the only validated bubble where the underlying cycle dominates the multi-year return.

Closing - the discipline the cycle requires

The historical record is clear and the math is unsentimental. Memory cycles, dampened or not, end in drawdowns that round-trip most of the late-cycle gains for anyone without an exit plan. The current cycle is +1,851% off the lows; the cycle template puts the current state at late or peak across 8 of 10 structural signals; the smartest memory bull on the tape is positioned long volatility rather than long stock.

None of this is a sell call on $MU. The bloc thesis is real, the HBM allocations are sold out, and the next 6 months are probably the last leg of an up-cycle that has been the largest in the company's history. What this is, is a framework for ensuring the trade is held correctly - sized for the down tail, stopped against the ATR, trimmed into vertical days, and exited on the first two of the eight indicators that say the cycle has turned.

The single most useful thing a memory trader can do in late-2026 is build the indicator-tracker before they need it. By the time the headlines say "memory cycle has rolled," the stock is already 25% off the peak and the multiple has stayed cheap the whole way down. The eight indicators above are leading by months, not days.

That is why memory remains the cleanest validated bubble on the platform - and why it is the one bubble where the cycle framework matters more than the bubble framework. The supercycle is real. So is the cycle inside it.

To trade $MU and the rest of the memory bloc from a US-retail or LLC account, see /stack/ibkr. Bubble shifts and rule-based alerts on the memory cohort, the kind that fire when the cycle indicators roll, are part of /pro.