SUPERSEDED 2026-05-18 by [[2026-05-18-memory-cycle-v1.1]]. Backtest evidence ([[01-projects/investing/backtests/2026-05-17-memory-cycle-v1-walk-forward]] + [[01-projects/investing/backtests/2026-05-17-memory-cycle-v1-real-anchor-rerun]]) showed mechanical price-based exits cost 19-192pp vs buy-and-hold on the 2019-2025 window. The thesis itself remained valid; the strategy spec was structurally wrong for multi-year cyclicals. v1.1 keeps the thesis + tranche-entry logic, KILLS all mechanical exits, ADDS anchor-based position-add (Druckenmiller doctrine), and EXPANDS the basket to MU + SMH + SNDK + INTC per 2026 Q1 smart-money entries by Druckenmiller / Tepper / Tiger Global. Read v1.1 for deploy-ready spec; v1 below preserved for audit trail.
Memory Cycle v1 — executable thesis
Upgrades [[2026-05-17-memory-cycle-thesis-v1-framing]] from framing to deploy-ready. Adds source-aggregated bull case with cited primary-source data, fresh-eyes bear case, cross-domain translation (Snowflake credit-glut analog), and quarterly review template. Parent: [[2026-05-12-innermost-loop-ai-infrastructure]] — memory is the chips layer's bottleneck sub-layer.
The thesis (one paragraph)
HBM (high-bandwidth memory) supply is structurally short through 2027 because hyperscaler AI training and inference demand has outrun the 18-24 month fab + packaging cycle that gates new memory capacity. Every HBM wafer also crowds out ~3 commodity DRAM wafers (Micron's stated conversion ratio), so the tightness cascades from HBM into DDR5, DDR4, and eventually NAND. The memory-maker oligopoly (Micron, SK Hynix, Samsung) has pricing power for the first time since the 2017-2018 supercycle — except this time the demand driver is multi-year capex from hyperscalers with $750B+ of committed 2026 spend, not a cyclical consumer-electronics pulse. Buy MU (US-listed pure-play) tranche-accumulated, use SMH as the semis-basket proxy, exit on phase-marker flip not calendar.
Bull case (structural argument)
The binding constraint
HBM packaging capacity is the binding constraint that gates AI compute scale-out. NVIDIA can ship Blackwell/Rubin, TSMC can fab the logic, but if SK Hynix / Micron / Samsung can't stack 12-Hi or 16-Hi HBM4 fast enough, the GPUs ship with less memory or wait. As of January 2026, all three memory makers' entire 2026 HBM production is pre-sold to Microsoft / Google / Meta / Amazon — anyone placing a fresh 2026 order is shut out [Source 4, Source 5]. SK Hynix executives state HBM shortage extends into 2028 [Source 7].
Demand-side anchors (with current-state data)
| Anchor | 2026 datapoint | Source |
|---|---|---|
| Hyperscaler AI capex (GOOG/AMZN/META/MSFT combined) | $750B 2026 (CreditSights revised UP from $650B) | Source 8 (parent thesis) |
| Micron HBM TAM forecast | $35B (2025) → $100B (2028); ~40% CAGR | Source 1 |
| AI share of global DRAM wafer capacity | 20% in 2026 (lead products: HBM, GDDR7) | Source 6 |
| HBM4 hyperscaler lockup | 100% of 2026 capacity sold across all three makers | Sources 4, 5 |
| Q1 2026 DRAM contract price QoQ | +90-95% (record) | Source 3 |
| Q2 2026 DRAM contract price QoQ | +58-63% (forecast) | Source 3 |
| Q2 2026 NAND contract price QoQ | +70-75% (forecast) | Source 3 |
Earnings inflection (the rerating moment)
Micron's Q2 fiscal 2026 (Feb 2026):
- Revenue $23.86B, +196% YoY, +75% QoQ
- Non-GAAP EPS $12.20 vs $8.79 consensus (+38.6% beat)
- Q3 guide: $33.5B revenue (±$750M), 81.0% gross margin, $19.15 EPS
- 2026 capex raised to $20B from $18B prior — entire raise targets HBM + 1-gamma node
- 2026 HBM supply fully pre-sold including HBM4
Source 1, Source 2, Source 9.
This is not a "demand might show up" thesis. The orders are signed, prices are locked, capex is committed. The question is whether the price-volume curve continues into 2027 (the thesis) or whether 2026 is the peak (the bear case, see below).
Locked tickers
| Ticker | Rationale | Max position |
|---|---|---|
| MU | US-listed DRAM pure-play; HBM3E shipping to NVIDIA; HBM4 ramps Q2 2026; 2026 supply pre-sold; capex raised to $20B | 2R ($10k) |
| SMH | VanEck Semiconductor ETF; basket proxy that captures HBM tailwind without single-name risk; ~10% MU + Hynix indirect + NVDA exposure | 2R ($10k) |
| Korean primaries (000660.KS Hynix, 005930.KS Samsung) | NOT in v1 — Alpaca tradeability limits | n/a |
| MRVL / WDC / ON | Defer to v2 — HBM-adjacent custom silicon, NAND/SSD exposure, analog respectively | n/a |
Memory bucket cap = 4R total ($20k). Initial deploy = 1R ($5k) split across MU + SMH.
Bear case
The strongest counter (steelmanned, no hedging):
This thesis dies the moment 2026 stops being the peak. Memory cycles have ALWAYS topped on signed orders and locked capacity — that's the textbook late-cycle signature, not a structural setup. In 2017 the supercycle ended with order books full and Micron printing record EPS; the stock lost 60% over the next 14 months as those orders shipped into a market with one quarter too much capacity. The 2026 setup looks identical: revenue growing 196% YoY with 81% gross margins is mathematically unsustainable, and the market knows it. The forward multiple compression on cyclical peaks is brutal — MU trades at trough P/E on peak earnings (5-7x EPS) precisely because the market is pricing the mean reversion.
The HBM-specific bull case is also weaker than it looks once you scratch it. SK Hynix lost ~200 engineers to a competing AI memory venture in early 2026 [Source 5], and a 50,000-worker strike at Hynix is threatening five-day supply disruption [Source 5] — these are signals of operational fragility, not pricing-power durability. Samsung is shipping HBM4 first while Hynix lags [Source 5], which means the oligopoly's pricing discipline is breaking down right when the bulls are calling it a moat. Three suppliers + price discipline = oligopoly profit; three suppliers racing for share = commodity reversion. Memory has ALWAYS oscillated between the two and is structurally biased toward the second over multi-year horizons.
The hyperscaler capex story has a tail risk too. $750B of 2026 capex was revised UP from $650B [Source 8], which sounds great, but capex always overshoots at cycle peaks. Meta's 2022 capex pullback came 9 months after they revised UP. If GOOG / AMZN / META / MSFT signal even a 10% downward 2027 capex revision in Q3 or Q4 2026 earnings, the entire memory bid disappears overnight because the memory order book IS the capex commitment. The 3-to-1 HBM/DDR5 wafer conversion ratio that bulls cite as constraint also runs in reverse: the moment HBM demand slows, that capacity floods back into commodity DRAM and crashes spot prices. The DRAM market has the highest beta to capex-cycle sentiment of any chip sub-sector — that's why it's tradeable, and that's why holding through the cycle is so painful.
China is the other wildcard. Washington's January 2026 clearance of ~10 Chinese firms to buy H200 [Source 10] reopens a demand channel that could either extend the cycle (more buyers) or accelerate it (Chinese firms typically order at lower margins, forcing price competition). YMTC + CXMT are scaling DDR5 production aggressively, and while they're a generation behind on HBM, they're 18-24 months from being a meaningful DDR5 supplier. The non-HBM DRAM market — which is still ~75% of memory-maker revenue — is one capacity announcement away from a glut.
Finally: the cross-domain analog bulls love (Snowflake credit-glut, see below) cuts both ways. Snowflake also had a moment where every CDO was committing to multi-year credit packages, gross margins were 75%+, and growth looked structural. Then in 2023 customers optimized usage and growth halved. The "fully sold out" signal is exactly when sophisticated buyers start optimizing. Hyperscalers don't sit still on a $100B HBM bill — they're already designing for less memory per FLOP (MoE routing, KV-cache compression, attention-window optimization). Software absorbs hardware. Always has. The bull case assumes hyperscalers keep paying memory tax forever; the bear case is they engineer around it inside 18 months, and we exit 2027 with Micron earning $4 EPS on $15B revenue instead of $19 EPS on $33B.
The single fact that would force me to capitulate to the bull case: any single hyperscaler signing a 3+ year HBM offtake contract at fixed price. That would prove the demand is structural rather than spot-cycle. Until that happens, this looks exactly like every memory peak I've ever seen.
Cross-domain translation (load-bearing)
This thesis is structurally identical to the 2020-2022 Snowflake credit-consumption cycle that you instrumented at phData.
Walk it through:
The binding constraint shape. In Snowflake's case, the constraint was warehouse-credit provisioning vs. customer query growth — usage scaled faster than enterprises could budget for, so credit consumption ASPs rose and Snowflake's gross margins expanded through 2022. In memory's case, the constraint is HBM packaging throughput vs. AI training compute demand — wafer starts scale on an 18-24 month cycle, but hyperscaler capex commits 6 months out, so HBM ASPs rise and Micron's gross margins go from 30% (commodity) to 81% (Q3 2026 guide). Same shape: a slow supply-side clock against a fast demand-side clock.
The reflexivity loop. Snowflake customers signed multi-year credit packages when they saw their queries growing 3x YoY — which validated more queries — which pulled forward more credit purchases. Hyperscalers are signing 2026 HBM packages while watching their inference load grow 5x YoY — which validates more capex — which pulls forward more 2027 HBM commitments. The order book IS the demand signal AND the demand pull. You saw this exact pattern when phData clients started prepaying Snowflake credits in 2021.
The "fully sold out" peak signal. When Snowflake's product gross margin hit 75% and every enterprise CDO was on a 3-year contract, you flagged it internally as a cycle marker — because that's when sophisticated buyers start optimizing. They did. Snowflake's NRR dropped from 178% to 127% over 6 quarters as customers compressed query patterns, materialized views, and pushed work back to Spark. The exact same compression is coming in HBM as MoE architectures, KV-cache quantization, and ternary weight schemes reduce memory per FLOP. The question for the memory thesis is the SAME question that decided Snowflake's 2022-2023 trajectory: how fast does software absorb the hardware tax?
The exit signal you'd actually have seen at Snowflake. It wasn't price. It wasn't growth. It was NRR compression in a single quarter combined with a downward forward-guide. The memory equivalent is: hyperscaler capex guide flat or down in any quarterly cycle, AND DRAM spot ASP -15% over 90 days. Two clocks, same shape.
If you've seen Snowflake 2020-2022, you've seen this thesis. The structural anchors are identical: oligopoly supply, customer concentration in hyperscalers, multi-year demand visibility through signed contracts, gross-margin compression as the leading exit signal. The thing that breaks it is also identical: customer optimization eating the hardware tax faster than supply catches up. Your read on Snowflake's cycle was right on timing — apply the same instrument here.
This is what gives the thesis a real handle: you don't need to be a semiconductor analyst to ratify it. You need to remember what Snowflake's GM curve looked like in 2021 and what it looked like in 2023, and ask whether HBM is in the 2021 part of the curve or the 2023 part. My read: 2026 looks like Snowflake mid-2021 (still pulling forward, optimization not visible yet). The bear case is that we're actually at Snowflake late-2022 already and just don't know it.
Demand-side anchors (observable signals confirming thesis)
- Hyperscaler capex direction (combined GOOG/AMZN/META/MSFT). Data source: quarterly earnings transcripts, CreditSights summaries. Bullish: flat or revised up. Bearish: revised down >10% any quarter.
- DRAM spot price 30d trend. Data source: TrendForce monthly summaries, DRAMeXchange. Bullish: flat or up. Bearish: -15% over 90 days.
- HBM capacity-online cadence. Data source: vendor earnings calls (MU, SK Hynix, Samsung), CES announcements, TrendForce capacity reports. Bullish: matching projected ramp (HBM4 Q2 2026, HBM4E samples H2 2026, HBM4E mass prod 2027). Bearish: oversupply announcement OR >2-quarter delay across 2+ vendors.
Disqualifying conditions (thesis-archive triggers)
- Hyperscaler 2027 capex guide -10% or worse in any quarter
- HBM4 demand-side cancellation/renegotiation signal from any tier-1 hyperscaler
- DRAM spot -15% over 90 days AND HBM ASP -10% in same quarter
- Any major memory maker announces capacity expansion that closes the bottleneck in <18 months
- CXMT or YMTC achieves volume HBM3E equivalent (low-prob 2026, high-prob 2027)
Two consecutive quarters of any single signal flipping = close entire position. Single quarter flip = trim to 1R, await confirmation.
Phase markers
| Phase | Definition | Data sources | Position posture |
|---|---|---|---|
| Phase 1 — Tightness (current) | All 3 anchors bullish: HBM sold out, capex flat-or-up, DRAM spot flat-or-up | TrendForce monthly + vendor earnings + hyperscaler 10-Qs | Tranche-accumulate to full size |
| Phase 2 — Plateau | 1 anchor flips, other 2 hold (e.g. DRAM spot flat to mildly down but capex + HBM cadence intact) | Same | Hold, no new tranches, raise trim alert |
| Phase 3 — Rollover | 2+ anchors flip for 2 consecutive quarters | Same | Close entire position per disqualifying-condition logic |
Watch surface: Ray runs quarterly review skill (see template below) and reports anchor state to founder.
Position sizing (R-unit, tranche schedule)
R-unit framing per parent README: 1R = $5,000. Memory bucket cap = 4R ($20k). Max per name = 2R ($10k).
Initial deploy (this thesis, founder click-through gates)
| Ticker | Tranche 1 (initial) | Trigger |
|---|---|---|
| MU | 0.5R = $2,500 | Founder APPROVE on /decisions/ page |
| SMH | 0.5R = $2,500 | Founder APPROVE on /decisions/ page |
Total initial deploy: 1R = $5,000 (paper).
Tranche schedule (per name)
| Tranche | Trigger | Add size | Running total |
|---|---|---|---|
| 1 (initial) | Founder APPROVE | 0.5R | 0.5R |
| 2 (scale-in) | -5% from tranche-1 avg price OR thesis-confirming catalyst (new HBM capex announcement, hyperscaler capex guide-up, ASP-up TrendForce monthly) | +0.5R | 1.0R |
| 3 (full) | -10% from tranche-1 avg OR earnings catalyst with thesis-positive data (MU Q3 26 results, Hynix HBM4 ramp news) | +1.0R | 2.0R (full) |
Exit rules
| Trigger | Action |
|---|---|
| 2+ phase markers flip bearish for 2 consecutive quarters | Close ENTIRE position (thesis broken) |
| Single phase marker flips bearish, other 2 hold | Trim to 1R per name (de-risk, await confirmation) |
| Price +50% from tranche-1 avg | Trim 0.5R (book profit), let rest run |
| Price +100% from tranche-1 avg | Trim another 0.5R (let core ride) |
| Thesis fundamentals improve materially (hyperscaler signs 3+ year HBM offtake at fixed price) | Add up to +0.5R (override cap to 2.5R per name) |
| Founder kill-switch (channel message: "halt memory" or "pause investing") | Close immediately, log to vault |
No stop-loss on individual trades. The thesis itself has a stop (anchor-break condition), and when the thesis stops, the position stops.
Quarterly review template
Ray runs each quarter and surfaces to founder via channel + /decisions/ page if any anchor flips.
## Memory Cycle v1 — Q[N] [YYYY] Review
### Anchor state
1. Hyperscaler capex direction
- Source: GOOG/AMZN/META/MSFT Q[N] earnings (link)
- Current: $[X]B combined ([+/-Y%] vs prior quarter guide)
- State: bullish / plateau / bearish
2. DRAM spot price 30d trend
- Source: TrendForce [month] release (link)
- Current: $[X]/Gb ([+/-Y%] over 90 days)
- State: bullish / plateau / bearish
3. HBM capacity-online cadence
- Source: vendor earnings + CES/announcements
- Current: [HBM4/4E ramp status per vendor]
- State: bullish / plateau / bearish
### Position state
- MU: [X shares] @ [avg price]; current $[Y]; [+/-Z%]
- SMH: [X shares] @ [avg price]; current $[Y]; [+/-Z%]
- Bucket total: [X]R deployed of 4R cap
### Phase determination
- Phase 1 / 2 / 3 (justification)
### Action this quarter
- [Continue tranche-accumulate / hold / trim / close / add catalyst-triggered tranche]
### Founder gate?
- Yes/No (any phase change = founder click-through via /decisions/)
Cadence: end of each calendar quarter (March, June, September, December). Ray logs the review to ~/rdco-vault/01-projects/investing/theses/reviews/<YYYY-Q[N]>-memory-cycle-v1.md.
Sources (numbered for citation above)
- Micron Q2 fiscal 2026 prepared remarks — primary, earnings transcript
- Micron Q2 2026 Earnings: Record Revenue on AI Boom (tech-insider summary)
- TrendForce: AI Server Demand Drives Memory Contract Price Increases in 2Q26 (March 2026) — primary, sell-side capacity tracker
- TrendForce: Micron Hikes CapEx to $20B with 2026 HBM Supply Fully Booked; HBM4 Ramps 2Q26
- TechTimes: Samsung Begins HBM4 Shipments as SK Hynix Lags, 50,000-Worker Strike Threatens AI Chip Supply (May 16 2026) — bear-case operational fragility evidence
- TrendForce: AI to Consume 20% of Global DRAM Wafer Capacity in 2026
- TradingKey: SK Hynix HBM Shortage Until 2028 (May 2026)
- [[06-reference/2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]] — vault, $750B hyperscaler capex anchor
- Micron Q1 fiscal 2026 prepared remarks — primary, prior-quarter context
- [[06-reference/2026-05-15-innermostloop-singularity-optimizing-optimizer]] — vault, China H200 clearance + compute supply chain rearrangement context
- [[06-reference/2026-05-16-a16z-charts-of-the-week-memory-ai-adoption]] — vault, late-cycle attention marker + cyclicality discipline check
Related
- [[2026-05-17-memory-cycle-thesis-v1-framing]] — v2 long-horizon framing this doc upgrades
- [[2026-05-12-innermost-loop-ai-infrastructure]] — parent thesis (memory = chips-layer bottleneck sub-layer)
- [[01-projects/investing/README]] — risk frame + dual-identity framing
- [[~/.claude/skills/investing-build-thesis/SKILL.md]] — SOP this thesis follows
- [[feedback_calibrate_overconfidence]] — applies: bear case may be stronger than parent bull-case author wants to admit
Open companion deliverables
/decisions/2026-05-17-paper-trade-memory-v1-go.html— paper-trade-go decision page (4-option click-back: APPROVE / REVISE / ARCHIVE / DEFER)_decisions.jsonentry — surfaces on HQ decisions index- Quarterly review skill instrumentation — defer to v2 (manual quarterly review by Ray in interim)
Changelog
- 2026-05-17 (v1 executable) — Upgraded from framing doc per
investing-build-thesisSOP. Added source-aggregated bull case with Micron Q2 26 + TrendForce + Hynix HBM4 primary data. Bear case authored as steelman (Snowflake-shaped optimization risk, oligopoly fragility from Hynix engineer poaching + strike, China DDR5 wildcard). Cross-domain translation = Snowflake credit-glut cycle 2020-2023. Phase markers, exit rules, tranche schedule, quarterly review template all locked. Pending: founder deploy gate via /decisions/ page.