01-projects/investing/theses

nvidia supply chain v1

2026-05-18·investing-thesis·status: framing-pending-founder-deploy-gate·! medium
ai-infrachipssupply-chainpicks-and-shovelscowoshbmfoundrypackaginglithographysemicaptsmmuavgo-not-heredruckenmiller-doctrineno-mechanical-exitsr-unitspaper-trade

NVIDIA supply-chain v1 — executable thesis (picks-and-shovels)

First of two splits of the prior "AI-infra-chips" bucket. This thesis owns the layer of names that NVIDIA itself spends money on: foundry (TSM), advanced packaging (CoWoS / OSAT — Amkor, Powertech, Unimicron, Ibiden substrate), memory pure-plays (Micron — HBM the binding constraint), and semicap equipment (ASML lithography, AMAT / KLAC / LRCX deposition + metrology + etch). Whatever NVIDIA ships, these vendors get paid first. Sister thesis [[2026-05-18-nvidia-adjacent-v1]] owns the competitive layer (AMD, Broadcom, Intel) which competes WITH NVIDIA for the same end demand.

Why split the bucket

The prior "ai-infra-chips" frame collapsed two structurally different bets into one allocation. Supply-chain names monetize NVIDIA's revenue itself — CoWoS capacity, HBM cadence, foundry market share are the binding constraints, and the same vendors fill orders for AMD / Broadcom / hyperscaler custom ASICs too. Adjacent names monetize the alternative path — capex routed to non-NVIDIA accelerators (Google TPU, Meta MTIA, Amazon Trainium, AMD MI300/MI400, Intel Gaudi). The two baskets share macro exposure (AI capex direction) but diverge on the question "is NVIDIA's pricing power sustainable, or do hyperscaler custom-silicon programs route around it?"

Founder confirmed taxonomy 2026-05-18. Supply-chain = picks-and-shovels (vendor agnosticism to which accelerator wins); adjacent = competitive (different accelerator winners reshuffle adjacent allocation). Different exit triggers (supply-chain breaks on capacity-online cadence; adjacent breaks on customer-share signals).

The thesis (structural)

Hyperscaler AI capex hit ~$750B committed for 2026 and the binding constraints are NOT NVIDIA's design cycle — they are upstream:

  1. CoWoS advanced-packaging capacity at TSMC. NVIDIA H100/H200/B200/B300 + AMD MI300/MI350 + Broadcom custom ASIC + Google TPU all queue for the same CoWoS-S/CoWoS-L lines. TSMC has guided multiple capacity expansions (3x in 2024, 2x in 2025, additional 2026 ramp). Every wafer of CoWoS that ships is monetized by TSM + the OSAT layer underneath (Amkor, Powertech, Unimicron substrate, Ibiden substrate).

  2. HBM stacking (HBM3E now, HBM4 ramping Q2 2026). The 3-vendor oligopoly (Micron, SK Hynix, Samsung) has supply pre-sold through 2026 and into 2027. Every HBM wafer crowds out ~3 commodity DRAM wafers, so the tightness cascades. Micron is the only US-listed pure-play HBM beneficiary (Hynix / Samsung Korean-listed, not Alpaca-tradeable).

  3. Litho + deposition equipment for the 3nm / 2nm / 1.4nm node migrations that gate next-gen accelerators. ASML EUV is the binding monopoly; AMAT (deposition), KLAC (metrology), LRCX (etch) are the oligopoly underneath. Capex from TSMC + Intel + Samsung foundry flows directly to these names. Tepper added LRCX 2024Q4 (later exited), Tiger Global holds LRCX in top-10 at $833M (2026Q1).

  4. OSAT / substrate layer. Amkor (AMKR) — US-listed advanced-packaging assembly. Powertech (3036.TW), Unimicron (3037.TW), Ibiden (4062.T) — substrate suppliers underneath CoWoS. Foreign-listed names are not Alpaca-tradeable; included in thesis universe for completeness and as ADR-watch.

The thesis is "vendor-agnostic infrastructure." We are not picking which accelerator wins; we are betting that whichever accelerator wins, the same upstream suppliers get paid.

Smart-money signal scoring (across 8 tracked managers)

Per [[01-projects/investing/anchors/smart-money/2026-05-17-aggregate-2yr-backfill-summary]] + per-manager backfills:

Ticker Signal strength Source Anchor specificity Mirror overlap?
TSM 3-manager persistent Tepper + Druckenmiller + Tiger Global top-10 for 2+ quarters (#3 Druckenmiller $167M; #5 Tepper $448M; persistent Tiger Global hold) HIGH — direct CoWoS capacity + foundry share YES — at 0.5R in mirror
MU 1-manager high-conviction Tepper #2 position 2026Q1 at $562M (+106% in 2025Q2 to $101M, then ramped to $562M by 2026Q1) HIGH — HBM3E shipping, HBM4 Q2 2026 ramp, supply pre-sold 2026 NO
LRCX 1-manager + 2-manager historic Tiger Global top-10 #10 at $833M in 2026Q1; 2024Q4 NEW from Tepper + Tiger (both later exited); re-entered by Tiger MEDIUM — etch equipment to all foundries; cyclical NO
AMAT weak/historic Tiger Global NEW 2024Q2; not in 2026Q1 visible top-10 MEDIUM — deposition to all foundries NO
ASML weak/historic Tepper +106% in 2024Q4 (to $110M, sub-top-10); no recent HIGH — EUV monopoly NO
KLAC none in tracked managers' top-10 MEDIUM NO
AMKR none in tracked managers' top-10 HIGH — direct CoWoS-adjacent OSAT NO
3036.TW / 3037.TW / 4062.T not visible (foreign filings) HIGH — substrate supply NO; also not Alpaca-tradeable

Names that don't earn a slot

Position sizing (R-unit, v1)

Per parent README: 1R = $5,000. Bucket cap = 4R ($20k). The R-weight reflects signal-strength + anchor-specificity + duplicate-coverage adjustment.

Initial deploy (paper)

Ticker R $ Why this weight Mirror overlap adjustment
MU 1.5R $7,500 Strongest single-name fundamental anchor (HBM supply pre-sold through 2026, +106% YoY revenue Q2-fiscal-2026, only US-listed HBM pure-play). Tepper sized to #2 position at $562M = his strongest semis-conviction. No mirror overlap → full weight. None
TSM 0.5R $2,500 3-manager persistent (the highest smart-money corroboration of any chip name). But ALREADY in smart-money-mirror at 0.5R → this is an INCREMENTAL 0.5R on top, bringing total TSM exposure to 1.0R across both strategies. Acknowledged double-count: each strategy tracks separately, founder needs to be aware net TSM = 1.0R combined. DOUBLE-COUNTED — net exposure 1.0R
LRCX 0.5R $2,500 Tiger Global top-10 at $833M (2026Q1); cyclical semicap exposure adds different shape from foundry/memory pure-plays. Single-manager so sized at half of MU. None
Cash reserve 1.5R $7,500 Held for (a) ASML / AMKR add if smart-money signal lights up in Q2/Q3 2026 13F, (b) anchor-strength add to MU on >20% drawdown with bullish anchors per Druckenmiller doctrine, (c) AMKR if CoWoS-shortage news cycle materializes
Total deployed 2.5R $12,500
Bucket cap 4R $20,000

Korean primaries explicitly excluded

Samsung (005930.KS) and SK Hynix (000660.KS) are excluded from v1 due to Alpaca paper-sandbox tradeability. They ARE the largest HBM beneficiaries by absolute output and would belong in this basket if our brokerage supported them. Re-evaluate when brokerage expands.

Druckenmiller doctrine (same as memory v1.1 + power v1.1)

Trigger Severity Action on first signal Action on confirmation
TSMC CoWoS capacity expansion announcement materially exceeds projected demand (cycle-peak signal) High Founder review via /decisions/ Close ENTIRE bucket
HBM capacity-online cadence catches/exceeds hyperscaler capex (the "memory thesis" broke) High Founder review Close ENTIRE bucket
Hyperscaler combined capex revised DOWN >10% across 2+ consecutive quarters High Founder review Close ENTIRE bucket
2+ tracked smart-money managers EXIT same supply-chain name same quarter Medium Founder review Trim affected name to 0.5R
Major foundry / memory vendor signals 2027 capex cut Medium Founder review Trim to 0.5R
CXMT or YMTC achieves volume HBM3E-equivalent production Medium Founder review Trim MU specifically
Single hyperscaler announces in-house foundry / packaging strategy that bypasses TSMC Medium Founder review Trim TSM specifically
Founder kill-switch (channel: "halt supply-chain" / "pause investing") n/a Close immediately n/a

Single signal of any severity = Ray surfaces via /decisions/ within 7 days. No autonomous exit on a single anchor flip.

Lightweight backtest read (5-min sanity check)

Did NVDA supply-chain names track NVDA's 2023-2026 ramp?

Yes, clearly. TSM appreciated ~3x from early 2023 ($75 area) to 2026Q1 ($200+) — directly tracking NVDA's ~10x in the same window, with the natural derating because supply-chain captures margin-with-a-lag and isn't the design-IP layer. MU returned 2.5x from the 2022 cyclical bottom ($50) to 2026 ($120+) driven specifically by HBM3E mix-shift in 2024-2025. LRCX returned ~2x from 2023 lows tracking foundry capex. The supply-chain didn't capture NVDA's pricing-power premium (NVDA gross margins 75%+ vs TSM 55% vs MU 35-40%) but it absolutely tracked the ramp directionally. The thesis is empirically the right shape for the past cycle.

Honest caveat: 2023-2026 is a single cycle, and this is a sanity check not a backtest. A full multi-cycle backtest of "supply-chain follows leading-edge accelerator" would need to backstop into 2017-2018 (NVDA Pascal/Volta cycle) and 2020-2021 (Ampere). Defer to next backtest cycle.

Overlap with smart-money-mirror v1 (explicit)

TSM is the only direct overlap. Mirror has TSM at 0.5R; supply-chain v1 has TSM at 0.5R; net TSM exposure across both strategies = 1.0R = $5,000.

This is intentional double-counting:

Position-tracking stays distinct per strategy (smart-money-mirror-v1 logs TSM separately from nvidia-supply-chain-v1 logs). Founder needs to know that total notional in TSM is 2x what each ledger shows individually.

Decision point for founder: is 1.0R = $5k net exposure to TSM acceptable, or should the supply-chain v1 layer reduce TSM to 0.25R (net 0.75R) to free up R for AMKR or ASML adds?

Cross-domain translation

Shape here is upstream-supplier capture of a downstream pricing-power cycle. Same shape as: oil-services (Halliburton / Schlumberger) when E&P capex booms — the picks-and-shovels names monetize the cycle without taking the discovery-risk that the operator names take. Or: SaaS-vendor capture of the 2018-2022 cloud-native build-out — Datadog/Snowflake/Confluent grew alongside AWS without competing with AWS. The NVIDIA supply-chain thesis applies the same template: TSM/MU/LRCX/AMKR monetize NVIDIA's hyperscaler-capex windfall without taking the design-IP risk that NVIDIA itself takes.

In data-engineering analogy: this is the dbt-vendor bet on the modern-data-stack cycle — pick the workflow surface every operator passes through, not the operator itself. TSM is dbt for the AI-chip stack; whichever accelerator wins, the workflow passes through TSM's fabs.

Disqualifying conditions (thesis-archive triggers)

Two-or-more of the HIGH severity exit triggers confirmed across 1-2 quarters = thesis archived, bucket closed, post-mortem filed. Specifically: any of (CoWoS oversupply, HBM cadence catches demand, hyperscaler capex DOWN >10% for 2Q) firing together = full bucket exit.

Phase markers (informational, not exit triggers)

Phase Definition Position posture
Phase 1 — Tight upstream, demand outrunning capacity (current) TSMC CoWoS sold out, HBM pre-sold, hyperscaler capex flat-or-up Initial 2.5R deploy + cash reserve for anchor-strength adds
Phase 2 — Capacity expansions announced but not yet online Material CoWoS or HBM capacity expansion in the pipeline Hold; consider trim if expansion materially exceeds projected demand
Phase 3 — Capacity online catches/exceeds demand Cycle peak signal — supply chain margins compress Close per disqualifying-condition logic

Caveats / open risks

Related

Open companion deliverables

Changelog