/decisions · 2026-05-18 · investing · nvidia-supply-chain-v1
Splits the prior "AI-infra-chips" bucket into picks-and-shovels (this thesis) vs competitors (sister thesis). This basket owns the layer that gets paid whichever accelerator wins: TSM (foundry + CoWoS, already 3-manager-persistent smart-money), MU (HBM pure-play, Tepper's #2 position at $562M), LRCX (semicap, Tiger Global top-10). Druckenmiller doctrine (no mechanical exits, drawdowns trigger adds, exits only on fundamental anchor break or 2+ manager exit). R-weighted: MU 1.5R + TSM 0.5R + LRCX 0.5R + cash reserve 1.5R = 4R bucket cap. TSM is double-counted with smart-money-mirror v1 — net TSM exposure across both strategies = 1.0R = $5,000. Paper only.
The prior "AI-infra-chips" frame collapsed two structurally different bets into one allocation:
Different fundamental anchors, different exit triggers, different smart-money signal lines. The split lets the founder gate each independently rather than allocating one R-budget across structurally different bets.
| Ticker | Signal | Source | Anchor | Mirror overlap? |
|---|---|---|---|---|
| TSM | 3-manager persistent | Tepper + Druckenmiller + Tiger top-10 for 2+ quarters | HIGH (CoWoS + foundry share) | YES — at 0.5R in mirror |
| MU | 1-manager high-conviction | Tepper #2 at $562M (2026Q1); +106% in 2025Q2 | HIGH (HBM3E + HBM4 ramp) | NO |
| LRCX | 1-manager + 2-manager historic | Tiger Global #10 at $833M (2026Q1) | MEDIUM (etch to all foundries) | NO |
| AMAT | weak/historic — CUT | Single Tiger NEW 2024Q2; nothing recent | Redundant with LRCX | NO |
| ASML | weak/historic — WATCHLIST | Tepper +106% in 2024Q4 (sub-top-10); no recent | HIGH but no current signal | NO |
| KLAC | none — CUT | Zero smart-money in 8-manager scan | MEDIUM | NO |
| AMKR | none — WATCHLIST | Zero smart-money; thin margins | HIGH (CoWoS-adjacent OSAT) | NO |
| 3036.TW / 3037.TW / 4062.T | not tradeable | Foreign-listed; not Alpaca | HIGH (substrate) | NO |
The bull case: Supply-chain captures hyperscaler AI capex regardless of which accelerator wins. CoWoS capacity at TSMC is the binding constraint for NVIDIA H100/H200/B200/B300 AND AMD MI300/MI350 AND Broadcom Google-TPU AND Intel Gaudi — every wafer monetizes TSM. HBM is the binding constraint at the memory layer — every shipped accelerator needs it, and Micron is the only US-listed pure-play (Hynix + Samsung Korean-listed, not Alpaca-tradeable). Smart-money signal is strongest on TSM (3-manager persistent) and Tepper's MU position (#2 holding at $562M) is the strongest single-name smart-money signal of any chip name across the 8-manager scan. The 2023-2026 sanity-check holds: TSM ~3x, MU ~2.5x from cycle bottom, LRCX ~2x — all directly tracking the NVIDIA ramp at the supply-chain's natural margin-with-a-lag discount.
The revision risk: Supply-chain margins are structurally lower than design-IP layer (TSM 55% gross margin vs NVDA 75%, MU 35-40% vs NVDA 75%), so the upside per dollar of capex is compressed compared to the design names. MU concentration is 37.5% of bucket — single-name drawdown material if HBM supply-demand inverts (CXMT or YMTC achieving HBM3E-equivalent volume production is the kill-switch). TSM has binary geopolitical tail (Taiwan strait); the 1.0R combined exposure is the bucket's largest single-stock geopolitical risk. CoWoS competitor risk: Samsung + Intel Foundry are both racing for CoWoS-equivalent capacity — if either comes online materially faster than expected, TSM's pricing power on advanced packaging compresses. Korean primaries (Hynix + Samsung) absent from basket = thesis is genuinely under-expressed by US-listed names only. N=1 sanity-check is not a backtest; multi-cycle supply-chain-follows-leader work deferred.
The picks: APPROVE locks in the picks-and-shovels paper-trade shape with the R-weighted basket. REVISE walks back (e.g. drop TSM to zero in this layer to avoid double-count, or reduce MU concentration to 1R, or add ASML / AMKR at smaller weights). ARCHIVE concedes the picks-and-shovels argument doesn't pencil (e.g. supply-chain margin compression dominates the cycle). DEFER waits for Q2/Q3 2026 13F to confirm Tiger Global LRCX hold and Tepper MU position size.
| Name | R | $ | Convergent signals |
|---|---|---|---|
| MU (Micron) | 1.5R | $7,500 | Tepper #2 position $562M (2026Q1) = strongest single-name smart-money signal of any chip name; HBM3E shipping; HBM4 Q2 2026 ramp; 2026 supply pre-sold. Only US-listed HBM pure-play. |
| TSM (Taiwan Semi) | 0.5R | $2,500 | 3-manager persistent (Tepper + Druckenmiller + Tiger). CoWoS capacity is the binding constraint for the entire AI-accelerator stack. DOUBLE-COUNTED — mirror already holds 0.5R; net TSM = 1.0R combined. |
| LRCX (Lam Research) | 0.5R | $2,500 | Tiger Global #10 at $833M (2026Q1); semicap exposure adds different shape from foundry/memory pure-plays. Single-manager signal so sized at half of MU. |
| Cash reserve | 1.5R | $7,500 | Reserved for: (a) anchor-strength add to MU on >20% drawdown with bullish anchors, (b) ASML/AMKR add if smart-money signal lights up in Q2/Q3 2026 13F, (c) CoWoS-shortage news-cycle opportunity. |
| Total deployed | 2.5R | $12,500 | Initial deploy; remaining 1.5R held in reserve. |
| Bucket cap | 4R | $20,000 | Paper, Alpaca sandbox. |
| Param | Value |
|---|---|
| Mode | Paper only (Alpaca paper sandbox) |
| 1R unit | $5,000 |
| Initial deploy | 2.5R total = $12,500 (MU 1.5R + TSM 0.5R + LRCX 0.5R) |
| Cash reserve | 1.5R = $7,500 held for anchor-strength adds + signal-confirmation new names |
| Bucket cap (supply-chain total) | 4R ($20k) |
| Per-trade stop | None (thesis-stop, not price-stop) |
| Profit trims | None — let winners run (Druckenmiller doctrine) |
| Exit trigger (single signal) | Founder review via /decisions/ page within 7 days |
| Exit trigger (confirmation) | 2+ HIGH severity anchors confirmed 1-2Q = close ENTIRE bucket |
| Kill switch | Founder channel "halt supply-chain" or "pause investing" = immediate close |
| Review cadence | Quarterly (Ray runs, surfaces to founder only on anchor flip) |
| TSM double-count decision | Surfaced for founder explicit OK — net TSM = 1.0R across mirror + supply-chain |
Ray fires 2.5R initial deploy ($12,500 paper) into Alpaca per the R-weighted table — MU 1.5R + TSM 0.5R + LRCX 0.5R. Holds 1.5R cash reserve. Logs to positions/nvidia-supply-chain-v1/. Sets up quarterly anchor watch. TSM net exposure across mirror + supply-chain = 1.0R = $5k accepted. Use the field for any deploy notes or sizing overrides.
Approve + sendv1 is close but a parameter needs adjusting (e.g. drop TSM from supply-chain to avoid double-count, reduce MU concentration to 1R, add ASML at 0.25R, equal-weight instead of R-weighted, smaller initial deploy). Name what; Ray re-files v1.1.
Revise + sendIf the picks-and-shovels argument doesn't pencil — not just the strategy. Ray files as status: archived-pre-deploy with reason. No paper trade. Supply-chain bucket closes (mirror still holds TSM independently).
Hold for Q2 2026 13F (re-confirm Tiger Global LRCX hold, Tepper MU sizing direction) before deploying. Q2 13F filings land mid-August.
Defer + send