NVIDIA-adjacent v1 — executable thesis (parallel-track + competitors)
Second of two splits of the prior "AI-infra-chips" bucket. Sister to [[2026-05-18-nvidia-supply-chain-v1]] (picks-and-shovels). This thesis owns the layer of names that compete WITH NVIDIA for the same hyperscaler accelerator wallet: AMD (MI300/MI350/MI400 GPU compute), Intel (Gaudi 3 + Terafab CoWoS-equivalent foundry), Broadcom (custom ASICs for Google TPU + Meta MTIA), Marvell (custom ASICs for Amazon Trainium / Microsoft Maia). The shared exposure is "hyperscaler AI capex routes to a non-NVIDIA accelerator." The shared risk is "NVIDIA's pricing power compresses if these competitors win share."
Why split the bucket
Sister doc [[2026-05-18-nvidia-supply-chain-v1]] explains the split rationale in full. TL;DR: supply-chain monetizes NVIDIA's revenue (vendor-agnostic); adjacent monetizes the alternative path. Different exit triggers, different fundamental anchors, different smart-money signal lines. Founder confirmed taxonomy 2026-05-18.
The thesis (structural)
Three drivers underpin the adjacent layer:
Hyperscaler custom silicon is real and accelerating. Google's TPU v5/v6 (designed with Broadcom), Meta's MTIA (designed with Broadcom), Amazon's Trainium 2/3 (designed with Marvell), Microsoft's Maia (designed in part with Marvell), and AMD's MI300/MI350 deployments at Meta + Microsoft + Oracle are no longer roadmaps — they are shipping at scale in 2025-2026. Combined, these likely capture 25-40% of incremental hyperscaler AI compute spend in 2026, up from <10% in 2023. Even at the low end of that range, the dollar quantity is enormous because hyperscaler capex itself is ~$750B/yr.
AMD specifically has a credible accelerator path. MI300 shipped meaningfully in 2024-2025 (Microsoft, Meta, Oracle, Lawrence Livermore). MI325 / MI350 in 2025-2026. MI400 announced for 2026. Lisa Su has guided ~$5B AI accelerator revenue for 2025 (well exceeded) and bullish trajectory thereafter. ROCm software-stack maturation is the historic gap vs CUDA; it is closing (PyTorch 2.x native ROCm support, vLLM ROCm support, hyperscalers contributing). The competitive gap is narrowing.
Intel is a special-case turnaround. Multi-manager 2026Q1 entry (Druckenmiller + Tiger Global both opened NEW INTC positions in Q1 — first time INTC shows up on the smart-money board since 2021). The Terafab strategy (US-based 18A node, government-incentivized CoWoS-equivalent advanced packaging, Gaudi 3 + future Falcon Shores accelerators) is a credible if uncertain turnaround. The smart-money entry is the corroboration; the operational execution is the open question.
Broadcom + Marvell are the ASIC designers behind the hyperscaler custom programs. AVGO's compute-ASIC business reached >$15B run-rate in 2025 (Google TPU + Meta MTIA the visible drivers). MRVL is the smaller comp, exposed primarily through Amazon Trainium custom ASIC + Microsoft Maia adjacent design.
Smart-money signal scoring (across 8 tracked managers)
Per [[01-projects/investing/anchors/smart-money/2026-05-17-aggregate-2yr-backfill-summary]] + [[01-projects/investing/anchors/smart-money/ark-trend-2yr]]:
| Ticker | Signal strength | Source | Anchor specificity | Mirror overlap? |
|---|---|---|---|---|
| AMD | 2-manager NEW SIGNAL | ARK #2 position at 4.29% AUM (2026Q1, built from outside top-10 over 3Q); Appaloosa-Tepper top-10 hold | HIGH — MI300/MI350 shipping; MI400 announced; ROCm closing CUDA gap | NO |
| INTC | 2-manager 2026Q1 NEW | Druckenmiller-Duquesne NEW 2026Q1; Tiger Global NEW 2026Q1 (Druckenmiller had also re-entered in 2025Q2 then exited 2025Q3, now re-entered 2026Q1) | MEDIUM — Terafab + Gaudi 3; turnaround uncertainty | YES — at 0.5R in mirror |
| AVGO | 1-manager high-conviction | Tiger Global top-10 #7 at $1.11B in 2026Q1 (+912% in 2024Q3 to $318M, ramped since) | HIGH — custom ASIC for Google TPU + Meta MTIA at >$15B run-rate | NO |
| MRVL | none in tracked managers' top-10 | — | MEDIUM — Amazon Trainium + Microsoft Maia custom-ASIC adjacency | NO |
| Samsung (005930.KS) | not visible (Korean filings) | — | LOW — Korean foundry competition to TSMC; broader semis | NO; also not Alpaca-tradeable |
Names that don't earn a slot
- MRVL — fundamental story is real (Amazon Trainium ASIC, Microsoft Maia adjacency) but zero smart-money corroboration AND the AVGO position captures the same hyperscaler-custom-ASIC thesis with far stronger smart-money signal. Cut as redundant.
- Samsung (005930.KS) — not Alpaca-tradeable; documented in universe for future brokerage-expansion conversation; zero deployment in v1.
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 |
|---|---|---|---|---|
| AMD | 1.5R | $7,500 | Strongest signal in basket. 2-manager NEW (ARK + Tepper). ARK built from zero to #2 position over 3 quarters — Cathie Wood was historically Nvidia-skeptical, the AMD pivot is meaningful. MI300/MI350 shipping with named hyperscaler customers; MI400 announced for 2026. No mirror overlap → full weight. | None |
| AVGO | 1.0R | $5,000 | Tiger Global top-10 at $1.11B (~$1B+ position — among Tiger's largest). $15B+ run-rate compute-ASIC business with Google TPU + Meta MTIA as named anchor customers. Single-manager signal but the position size and position trajectory (+912% in 2024Q3) is convincing. No mirror overlap → full weight. | None |
| INTC | 0.25R | $1,250 | 2-manager 2026Q1 NEW (Druckenmiller + Tiger), HIGH smart-money corroboration. But ALREADY in smart-money-mirror at 0.5R → this is an INCREMENTAL 0.25R bringing total INTC exposure to 0.75R across both strategies. Sized lower than the raw signal would suggest specifically because of the duplicate-coverage adjustment + the Druckenmiller entry-exit-re-entry pattern reads as opportunistic rather than conviction. | DOUBLE-COUNTED — net exposure 0.75R |
| Cash reserve | 1.25R | $6,250 | Held for (a) MRVL if a smart-money signal materializes in Q2/Q3 2026 13F (currently zero — would re-evaluate), (b) anchor-strength add to AMD on >20% drawdown with bullish anchors per Druckenmiller doctrine, (c) Samsung if brokerage expands to Korean ADRs | — |
| Total deployed | 2.75R | $13,750 | ||
| Bucket cap | 4R | $20,000 |
Druckenmiller doctrine (same as memory v1.1 + power v1.1 + nvidia-supply-chain v1)
- No mechanical price-based stops. Drawdowns are not exit triggers. Drawdowns with bullish anchors are BUYING signals.
- Tranche entry permitted. Initial 0.5R per name; add 0.5R on -5% from t1 avg with thesis-confirming catalyst; add 1R on -10% from t1 avg with earnings catalyst.
- Exit triggers are fundamental, not price-based:
| Trigger | Severity | Action on first signal | Action on confirmation |
|---|---|---|---|
| NVIDIA gross margins re-expand >78% for 2+ quarters (signals competitors NOT taking share) | High | Founder review via /decisions/ | Close ENTIRE bucket |
| AMD AI revenue MISSES Lisa Su guidance by >20% for 2+ consecutive quarters | High | Founder review | Trim AMD specifically; consider bucket close if structural |
| Broadcom compute-ASIC revenue decelerates >30% QoQ (Google / Meta program slowdown) | High | Founder review | Trim AVGO specifically; consider bucket close if confirmed |
| Hyperscaler combined capex revised DOWN >10% across 2+ consecutive quarters | High | Founder review | Close ENTIRE bucket (same as supply-chain trigger) |
| 2+ tracked smart-money managers EXIT same adjacent name same quarter | Medium | Founder review | Trim affected name to 0.5R |
| ARK exits AMD top-10 (specific to AMD; ARK's build is the load-bearing 2nd manager signal) | Medium | Founder review | Trim AMD to 0.5R, await second signal |
| INTC turnaround signals fail (Terafab delays >12mo, Gaudi 3 customer adoption stalls) | Medium | Founder review | Trim INTC specifically |
| Single hyperscaler announces re-commitment to NVIDIA-only at scale | Low-Medium | Watch | Note in review |
| Founder kill-switch (channel: "halt adjacent" / "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-adjacent names track or diverge from NVDA's 2023-2026 ramp?
Diverged meaningfully — adjacent names lagged for 2 years, then re-rated in late 2024 / 2025. AMD spent 2023-2024 underperforming NVDA badly: while NVDA went ~10x, AMD went sideways from $100→$130 area then re-rated in 2024H2 when MI300 began shipping at scale. AVGO tracked NVDA more closely (the custom-ASIC business is the right shape) — roughly 3-4x from 2023 lows. INTC was the worst performer through 2023-2025, down 30%+ from 2023 highs, only stabilizing in late 2025 / early 2026 as the smart-money entries surfaced.
The honest read: adjacent names are a derivative bet on "NVIDIA's pricing power eventually compresses." They underperform in the early innings of an AI ramp (when NVIDIA's monopoly pricing IS the story) and outperform in the maturation phase (when hyperscalers materially route around it). We are deploying in 2026, which by the smart-money signal (ARK pivot + Druckenmiller INTC entry + Tepper AMD position) reads as the start of the maturation phase. The thesis is empirically the right shape for THIS phase; it would have been the wrong shape in 2023.
Caveat: this is a 5-min sanity check, not a backtest. Multi-cycle adjacent-vs-leader analysis (e.g., AMD vs Intel during 2003-2006, AMD vs Intel during 2017-2020) would strengthen confidence. Defer.
Overlap with smart-money-mirror v1 (explicit)
INTC is the only direct overlap. Mirror has INTC at 0.5R; adjacent v1 has INTC at 0.25R; net INTC exposure across both strategies = 0.75R = $3,750.
This is intentional double-counting at a reduced rate:
- The mirror layer says "Druckenmiller + Tiger opened INTC in 2026Q1, so we follow at the strategy's 0.5R-per-name default."
- The thesis layer says "INTC is the Terafab + Gaudi turnaround story; the smart-money entry IS the entry trigger but the operational execution is uncertain — size at 0.25R reflecting the higher uncertainty vs AMD."
- Both signals are independently valid; the reduced incremental sizing reflects that the mirror already captures the signal at a default weight.
Position-tracking stays distinct per strategy. Founder needs to know total INTC notional is 1.5x what each ledger shows individually.
Decision point for founder: is 0.75R = $3,750 net exposure to INTC acceptable, or should the adjacent v1 layer drop INTC entirely (rely solely on mirror coverage) and redeploy the 0.25R into AMD (bringing AMD to 1.75R)?
Cross-domain translation
Shape here is competitive-displacement bet on a monopoly-pricing cycle. Same shape as: when a dominant cloud provider's pricing power compresses, the "second cloud" + "private cloud + hybrid" beneficiaries re-rate (e.g., the AWS-to-multi-cloud rotation in 2019-2022 benefited Azure + GCP + on-prem hybrid vendors). Or: when Intel's x86 server monopoly compressed in 2017-2022, AMD EPYC re-rated 20x. The NVIDIA-adjacent thesis applies the same template: when hyperscalers route around the dominant supplier (NVIDIA), the next-best alternatives (AMD GPU, custom ASIC designers) capture the incremental spend.
In data-engineering analogy: this is the "the data-warehouse Snowflake monoculture cracked, look at the multi-tool Databricks + DuckDB + Postgres-OLAP beneficiaries" bet. AMD is Databricks (credible alternative with momentum); AVGO is the platform-vendor enabling-the-alternatives (Snowflake-on-Iceberg-style); INTC is the legacy vendor with the comeback story.
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: NVDA margin re-expansion AND AMD AI revenue miss AND AVGO ASIC decel firing together = full bucket exit. Single-name softness = trim, not close.
Phase markers (informational, not exit triggers)
| Phase | Definition | Position posture |
|---|---|---|
| Phase 1 — Monopoly pricing intact (2023-2024) | NVIDIA gross margins 75%+, customers NVIDIA-only at scale | Adjacent underperforms; do not deploy |
| Phase 2 — Customer diversification starts (current, 2025-2026) | Hyperscaler custom-silicon programs reach material scale; AMD MI300+ shipping; ARK + Druckenmiller signal turn | Deploy adjacent basket; ride the maturation phase |
| Phase 3 — Competitive pricing pressure | NVIDIA gross margins compress materially; share gains visible in quarterly reports | Hold; add on drawdowns per anchor-strength logic |
| Phase 4 — Cycle peak / consolidation | All accelerators commoditized; pricing collapses across the board | Close per disqualifying-condition logic |
Caveats / open risks
- N=1 sanity check, not a backtest. Adjacent-vs-leader thesis empirically held through 2025-2026 but that's a partial window. Multi-cycle backtest deferred.
- AMD concentration (1.5R = 37.5% of bucket). Single-name drawdown risk material. Acceptable in paper given anchor specificity + 2-manager signal; revisit before live-capital migration.
- ARK is the load-bearing 2nd manager for AMD. ARK is thematic-disruption, not quality-compounder — its 2+ signal is "idea-generation grade" not "conviction-stacking grade" per the ARK manager-note. If ARK exits or trims AMD next quarter, the 2-manager threshold for AMD breaks and we're back to single-manager (Tepper).
- INTC turnaround risk is binary. Terafab + Gaudi 3 either delivers or doesn't; the multi-year position will work or won't based on execution. Smart-money entry is a credible vote of confidence but is not insider information about execution.
- AVGO single-manager signal. Tiger Global is the only persistent holder. If Tiger trims aggressively, the smart-money corroboration weakens significantly.
- Custom-silicon thesis depends on hyperscaler software-stack maturity. ROCm (AMD), JAX-on-TPU (Google), in-house Trainium toolchain (Amazon), Maia toolchain (Microsoft) all need to keep maturing. If software stalls, the hardware doesn't ship at scale and the adjacent thesis underperforms.
- The thesis is BEARISH on NVIDIA at the margin. We are not paired-trading (no NVDA short), but the adjacent bet is structurally a partial-bearish-on-NVIDIA bet. If NVIDIA's pricing power proves more durable than the thesis implies (e.g., Blackwell Ultra commands an even larger premium), the adjacent basket lags.
Related
- [[2026-05-18-nvidia-supply-chain-v1]] — sister thesis (picks-and-shovels layer)
- [[2026-05-18-memory-cycle-v1.1]] — sibling thesis (memory bucket); INTC also lives there
- [[2026-05-18-power-cycle-v1.1]] — sibling executed thesis; same Druckenmiller-doctrine shape
- [[2026-05-12-innermost-loop-ai-infrastructure]] — parent thesis (chips-layer)
- [[01-projects/investing/strategies/2026-05-18-smart-money-mirror-v1]] — mirror layer; INTC double-count documented
- [[01-projects/investing/anchors/smart-money/2026-05-17-aggregate-2yr-backfill-summary]] — selection-source brief
- [[01-projects/investing/anchors/smart-money/ark-trend-2yr]] — ARK 2yr backfill; AMD signal originates here
- [[feedback_calibrate_overconfidence]] — applies: N=1 sanity-check is not a backtest; ARK's signal is idea-generation grade not conviction-stacking grade
Open companion deliverables
/decisions/2026-05-18-paper-trade-nvidia-adjacent-v1-go.html— paper-deploy decision page (4-option click-back: APPROVE / REVISE / ARCHIVE / DEFER)_decisions.jsonentry — surfaces on HQ decisions indexpositions/nvidia-adjacent-v1/ledger folder — created on first executed trade post-approval- Next quarterly review queued for ~2026-08-22 (post 2026 Q2 13F window)
Changelog
- 2026-05-18 (v1) — Initial split of the prior "ai-infra-chips" bucket. Adjacent basket: AMD 1.5R + AVGO 1.0R + INTC 0.25R + cash reserve 1.25R = 4R bucket cap. MRVL cut (redundant with AVGO + zero smart-money). Samsung excluded for Alpaca tradeability. INTC double-counted with smart-money-mirror v1 (net 0.75R) at REDUCED incremental sizing; flagged for founder decision (drop INTC entirely from adjacent → redeploy to AMD?). Druckenmiller doctrine applied. Pending founder deploy gate via /decisions/ page.