Who feeds the beast? Memory, packaging, networking, cooling - the NVDA supply chain.

NVDA is the obvious AI play. But NVDA needs:

1. Memory - HBM from MU, SK Hynix
2. Packaging - TSMC CoWoS, ASE, Amkor
3. Networking - Arista, Marvell, Broadcom for AI clusters
4. Cooling - Vertiv, Modine for liquid cooling
5. Equipment - ASML, LRCX, KLAC, AMAT for chip making
6. Competitors - AMD, custom silicon (Broadcom, Marvell ASICs)

Edge: Deep understanding of NVDA's actual supply chain helps distinguish real suppliers from "AI-adjacent" hype.

• RSI ↓ = Oversold (<30) - potential buying opportunity
• RSI ↑ = Overbought (>70) - FOMO warning, may pullback
• 52wkHi = % from 52-week high (how far below peak)

• H100/H200/Blackwell all need High Bandwidth Memory
• HBM is THE bottleneck for GPU production
• Only 3 suppliers: SK Hynix (#1), Samsung (#2), Micron (#3)

MU (Micron) - The One You Missed

• What: #3 HBM supplier, direct NVDA supplier
• Why +40%: HBM thesis played out perfectly
• RSI 76 = OVERBOUGHT
• ✅ VALIDATED (2026-02-03): HBM is NOT caught up - still 2+ years constrained
• Verdict: Thesis was RIGHT, timing was WRONG. Too late now. Wait for pullback to $300-350.

SKM (SK Telecom) - Proxy for SK Hynix

• What: Korean telco that owns stake in SK Hynix (#1 HBM supplier)
• Also: Has Anthropic investment (AI double play)
• Why +35%: Anthropic deal + SK Hynix HBM momentum
• Verdict: Interesting but already ran. Research the actual SK Hynix exposure %.

• Advanced chips need advanced packaging (CoWoS, 2.5D, 3D)
• TSMC CoWoS has been THE bottleneck for H100 production
• NVDA locked 60%+ of CoWoS capacity through 2027

TSM (TSMC) - The Irreplaceable

• What: Makes NVDA's chips AND does CoWoS packaging
• Why: No TSMC = no NVDA
• RSI 56: Neutral, not overbought
• ✅ VALIDATED (2026-02-03): CoWoS constrained for "many decades" - this is THE structural bottleneck
• NEW: Samsung + TSMC building massive US fabs - US packaging capacity angle
• Verdict: Core holding. CoWoS constraint is STRUCTURAL and LONG-TERM.

ASX (ASE Technology) - OSAT Overflow Play #1 ⭐

• What: #1 OSAT globally, getting CoWoS overflow from TSMC
• Capacity expansion: CoWoS capacity jumping to 20-25K wafers/month (3x current)
• CoWoP development: Leading next-gen Chip-on-Wafer-on-PCB for NVDA
• NVDA connection: Vera CPU and automotive chips going through ASE
• Market cap: ~$25B - not as hyped as direct AI plays
• ✅ VALIDATED: CoWoS constrained "many decades" = OSATs become MORE valuable
• Verdict: HIGH PRIORITY research - benefits from validated constraint without NVDA concentration risk.

AMKR (Amkor) - OSAT Overflow Play #2 ⭐

• What: #2 OSAT globally, getting TSMC overflow alongside ASE
• NVDA capacity: ~80K wafers CoWoS production capacity (shared with ASE)
• US angle: May handle downstream packaging for TSMC Arizona
• Market cap: ~$7B - smaller than ASE, potentially more upside
• ✅ VALIDATED: Same overflow thesis as ASE
• Verdict: HIGH PRIORITY research - smaller = more levered to packaging thesis.

INTC (Intel) - EMIB/Foveros Alternative

• What: EMIB (cheaper than CoWoS) + Foveros (3D stacking, Jensen praised)
• Customers interested: Apple, Qualcomm, AWS, MediaTek approaching Intel
• US capacity: Rio Rancho, NM already positioned for US packaging
• Why interesting: CoWoS constraint pushing customers to alternatives
• Your validation: NVDA would fund INTC before AMD
• Verdict: Messy company, but packaging is bright spot. Watch for packaging-specific wins.

• AI clusters need ultra-fast interconnect (InfiniBand, high-speed Ethernet)
• NVDA owns Mellanox (InfiniBand) but still needs switches, optics
• 800GbE → 1.6TbE transition happening now

ANET (Arista) - The AI Networking Play

• What: Data center switches, AI cluster networking
• Raised AI revenue target: $1.5B → $2.75B
• 800GbE market: Tripled sequentially, Arista leads
• Meta deployed Arista for latest Ethernet AI cluster
• Key question: Does NVDA actually use Arista? Or is it all Mellanox/InfiniBand?
• RSI: Neutral
• Verdict: Need to validate - is Ethernet actually winning for AI training, or is this inference-only?

MRVL (Marvell) - Custom ASICs + Photonics

• What: Custom AI chips for hyperscalers + optical interconnect
• Data center = 73% of revenue, up 90% YoY
• Celestial AI acquisition: $5.5B for photonics (next bottleneck)
• Why -13%: Pulled back from highs
• Key question: Are custom ASICs (Google TPU, AWS Trainium) actually competitive with NVDA?
• RSI: Pullback territory
• Verdict: Interesting - both NVDA supplier AND competitor. Needs research.

AVGO (Broadcom) - The Everything Company

• What: Networking ASICs, custom AI chips, VMware
• AI networking: Supplies switches, NICs
• Custom silicon: Building chips for Google, others
• RSI: Neutral
• Verdict: Diversified play. Less pure NVDA exposure.

• H100/H200/Blackwell run HOT (700W+ per GPU)
• AI racks now 120-150kW, air cooling impossible
• Liquid cooling required for dense AI clusters

VRT (Vertiv) - The Cooling King

• What: #1 in data center liquid cooling
• Orders +60% YoY, liquid cooling revenue doubled
• $9.5B backlog
• Why +250%: AI data center thesis played out massively
• RSI 70+ = OVERBOUGHT
• Key question: Is liquid cooling actually deployed at scale, or still ramping?
• Verdict: Thesis was right, but +250% = you missed it. Wait for -20% pullback.

MOD (Modine) - The #2

• What: Thermal management, EV cooling, data center cooling
• Less pure play than Vertiv but smaller
• Verdict: Research - may have more room to run if you missed VRT.

• Every NVDA chip needs ASML, LRCX, AMAT, KLAC tools to be made
• BUT these are TSMC's suppliers, not NVDA's directly
• More of a derivative play

ASML - The EUV Monopoly

• What: Only company making EUV lithography machines
• Why +17%: AI chip demand = more TSMC orders = more ASML
• RSI 70 = approaching overbought
• Verdict: Essential but already ran. Wait for pullback.

LRCX/AMAT/KLAC - The Tool Ecosystem

• What: Deposition, etch, inspection equipment
• Less direct NVDA play - more TSMC/Samsung supplier
• Verdict: More cyclical, less pure AI thesis.

Your assessment (2026-02-03):

• AMD MI300: "lmao"
• ROCm (AMD's CUDA competitor): "absolute shitshow"
• Personal test with AMD AI 360 chip: "it is sad"
• NVDA would fund INTC before putting money in AMD

Key insight: The CUDA moat isn't hardware specs (MI300 looks competitive on paper). It's the SOFTWARE ECOSYSTEM. ROCm is broken. This is what retail/analysts miss.

Implication: Don't buy AMD on "catching up to NVDA" narrative. The gap is software, not hardware.

                              NVDA GPUs
                                  │
        ┌─────────────┬───────────┼───────────┬─────────────┐
        │             │           │           │             │
        ▼             ▼           ▼           ▼             ▼
     MEMORY       PACKAGING   NETWORKING   COOLING      EQUIPMENT
        │             │           │           │             │
        ▼             ▼           ▼           ▼             ▼
   MU (+40%)      TSM (+6%)   ANET (+1%)  VRT (+250%)  ASML (+17%)
   SKM (+35%)     ASX ⭐       MRVL (-13%)  MOD (?)     LRCX (?)
                  AMKR ⭐      AVGO (-4%)
                  INTC (alt)
        │             │           │           │             │
        ▼             ▼           ▼           ▼             ▼
   🔴 MISSED     ⭐ RESEARCH   🟢 RESEARCH  🔴 MISSED    🟡 WATCH
              (VALIDATED!)

1. Is HBM still THE bottleneck, or has supply caught up? ✅ ANSWERED: NOT caught up, 2+ years constrained
2. Which supplier is NVDA most dependent on - MU or SK Hynix? Still open

3. Is CoWoS still constraining production, or has TSMC scaled enough? ✅ ANSWERED: Constrained "many decades" - structural
4. Are OSATs (ASE, Amkor) actually getting meaningful NVDA overflow? Still open

5. Critical: Does NVDA use Arista gear? Or is it all Mellanox InfiniBand? Still open
6. Is Ethernet actually winning for AI training, or just inference? Still open
7. Is Marvell's custom silicon business competing with or complementing NVDA? Still open

8. Is liquid cooling actually deployed at scale in NVDA clusters? Still open
9. Who is the actual cooling vendor for NVDA reference designs? Still open

10. Is AMD MI300 actually competitive? What's the real CUDA moat? ✅ ANSWERED: "lmao" - ROCm is shitshow, CUDA moat is SOFTWARE
11. Are hyperscalers actually reducing NVDA orders for custom silicon? Still open

{
  "name": "nvda-ecosystem",
  "description": "NVDA supply chain - memory, packaging, networking, cooling, equipment",
  "symbols": [
    "MU", "SKM",
    "TSM", "ASX", "AMKR",
    "ANET", "MRVL", "AVGO",
    "VRT", "MOD",
    "ASML", "LRCX", "AMAT", "KLAC",
    "AMD", "INTC"
  ]
}

The narrative difference:

• AI Infrastructure = "AI needs power and buildings"
• NVDA Ecosystem = "What does NVDA specifically need to make GPUs?"

• Deep dive ASX - financials, packaging revenue %, CoWoS capacity breakdown
• Deep dive AMKR - financials, TSMC relationship, US packaging angle
• Compare ASX vs AMKR - which is better OSAT play?
• Watch INTC packaging news - customer wins for EMIB/Foveros

• Get your validation on networking (Arista vs Mellanox question)
• Research MRVL deeper (custom ASICs + Celestial AI acquisition)

• Set MU alert at $350 for pullback entry
• Set VRT alert at $120 for pullback entry

cd stonks
the desk's own tooling
the desk's own tooling