/decisions · 2026-05-17 · investing · memory-cycle-v1
Memory thesis v1 is filed (source-aggregated + bear-cased + cross-domain-translated). Strategy: tranche-accumulate MU + SMH, no per-trade stops, exit on phase-marker flip. Initial deploy = 1R ($5k) split across the two names. This is paper only — strategy validation, not real capital deployment.
Full thesis: 2026-05-17-memory-cycle-v1.md. Parent: Innermost Loop chips-layer thesis.
Thesis in one paragraph: HBM supply structurally short through 2027. All 2026 HBM capacity pre-sold to hyperscalers (MU + Hynix + Samsung). MU Q2 fiscal 2026 revenue +196% YoY at 81% gross margins, Q3 guide $33.5B / $19.15 EPS, 2026 capex raised to $20B with HBM4 ramping Q2 2026. DRAM contract prices +90-95% Q1 2026 QoQ, +58-63% Q2 forecast. Three demand-side anchors: hyperscaler capex direction, DRAM spot 30d trend, HBM capacity-online cadence. Exit on 2+ anchors flipping for 2 consecutive quarters.
Cross-domain handle: structurally identical to the Snowflake credit-consumption cycle 2020-2022 — 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 (MoE routing, KV-cache compression) faster than supply catches up. If you've seen Snowflake's 2021 GM curve vs its 2023 curve, you've seen this thesis. My read: memory 2026 ≈ Snowflake mid-2021. The bear case is that we're actually at Snowflake late-2022 already.
Bear case strength: medium-strong. The "fully sold out" signal is exactly the late-cycle marker memory has always had at peaks (2017 supercycle ended this way). Operational fragility at Hynix (engineer poaching + 50k-worker strike) + China DDR5 wildcard (CXMT/YMTC) are real near-term risks. Single fact that would force capitulation to bull case: any hyperscaler signing a 3+ year HBM offtake contract at fixed price.
| Param | Value |
|---|---|
| Mode | Paper only (Alpaca paper sandbox) |
| 1R unit | $5,000 |
| Initial deploy | 1R total — MU 0.5R ($2.5k) + SMH 0.5R ($2.5k) |
| Max per name | 2R ($10k) |
| Max bucket (memory total) | 4R ($20k) |
| Tranche 2 trigger | -5% from t1 avg OR thesis-confirming catalyst |
| Tranche 3 trigger | -10% from t1 avg OR earnings catalyst |
| Per-trade stop | None (thesis-stop, not price-stop) |
| Profit trims | +50% trim 0.5R; +100% trim another 0.5R |
| Kill switch | 2+ anchors bearish for 2Q OR founder channel "halt memory" |
| Review cadence | Quarterly (Ray runs, surfaces to founder via channel + /decisions/ if any anchor flips) |
Ray fires initial 1R tranche (MU 0.5R + SMH 0.5R) into Alpaca paper. Logs to positions/. Sets up quarterly anchor watch. Use the field for any context Ray should know going in.
Approve + sendThesis is close but a parameter needs adjusting. Name what (sizing, ticker, tranche trigger, exit rule, etc.) and Ray re-files a v2.
Revise + sendThesis fundamentally doesn't pencil. Ray files as status: archived-pre-deploy with reason. No paper trade.
Wait. Re-evaluate at the named date (next quarterly check or specific catalyst).
Defer + send