The AI Value Chain
The single most important conceptual note in the vault. It explains why these specific companies coexist. PCA SOF is not a basket of AI stocks; it is a deliberate ownership of every layer of the AI value chain, so the fund captures the spend wherever it lands.
The Chain, Top To Bottom#
┌─────────────────────────────┐
LAYER 0 EQUIPMENT │ Disco Corporation (dicing/grinding for HBM & packaging)
│ (ecosystem: ASML, Applied Materials, Lam, not held)
└──────────────┬──────────────┘
▼
LAYER 1 FOUNDRY │ TSMC fabricates ALL leading-edge AI silicon
└──────────────┬──────────────┘
▼
LAYER 2 SILICON │ COMPUTE NVIDIA MEMORY Micron Technology
│ NETWORK Marvell Technology (optical DSP, custom ASIC)
└──────────────┬──────────────┘
▼
LAYER 3 POWER │ Constellation Energy Vistra (electricity for datacentres)
└──────────────┬──────────────┘
▼
LAYER 4 CLOUD / │ Microsoft Amazon Alphabet Meta Platforms
HYPERSCALE │ (buy the silicon, build the datacentres, rent the compute)
└──────────────┬──────────────┘
▼
LAYER 5 DATA & │ DATA Snowflake OBSERVABILITY Datadog
OPERATIONS │ SECURITY CrowdStrike Zscaler
└──────────────┬──────────────┘
▼
LAYER 6 ENTERPRISE │ WORKFLOW ServiceNow CRM Salesforce
APPLICATIONS │ (embed AI agents into the enterprise)
└──────────────┬──────────────┘
▼
LAYER 7 CONSUMER │ Meta Platforms Netflix Roblox Alphabet (Search/YouTube)
ENDPOINT │ (AI-driven engagement, recommendation, content, monetisation)
└─────────────────────────────┘
Layer-By-Layer: The Holding, The Role, The Chokepoint#
| Layer | Holding | Role in the fund | Why it's a chokepoint / why owned | |, -|, -|, -|, -| | 0 Equipment | Disco Corporation | AI manufacturing tools | Near-monopoly in precision dicing/grinding, required for HBM stacking & advanced packaging. Picks-and-shovels of the picks-and-shovels. | | 1 Foundry | TSMC | AI manufacturing | The only foundry able to volume-produce leading-edge (3nm/2nm) AI chips. Single point of the entire chain. | | 2 Compute | NVIDIA | AI compute | GPU + CUDA software lock-in; the default AI training/inference platform. | | 2 Memory | Micron Technology | AI memory | HBM is the gating input for every accelerator; only 3 suppliers globally. | | 2 Networking | Marvell Technology | AI networking | Optical DSPs + custom ASICs (incl. hyperscaler in-house chips) tie GPUs into clusters. | | 3 Power | Constellation Energy | AI power demand | Largest US carbon-free (nuclear) generator; datacentres need 24/7 clean baseload. | | 3 Power | Vistra | AI power demand | Generation + nuclear; merchant power leverage to datacentre demand. | | 4 Cloud | Microsoft | AI infra + enterprise AI | Azure + OpenAI; sells AI as a service across the enterprise. | | 4 Cloud | Amazon | AI infra + cloud | AWS scale + custom silicon (Trainium/Inferentia); retail logistics flywheel. | | 4 Cloud | Alphabet | AI apps + search + cloud | Full-stack: TPUs, DeepMind, Gemini, Search, YouTube, Google Cloud. | | 4 Cloud | Meta Platforms | AI infra + consumer AI | Largest open-model (Llama) effort; AI-driven ad engine; massive GPU buyer. | | 5 Data | Snowflake | AI data layer | The governed data foundation models train and run on. | | 5 Observability | Datadog | AI monitoring | Monitors the cloud + AI estate; usage scales with cloud workloads. | | 5 Security | CrowdStrike | AI security (endpoint) | AI-native endpoint/SIEM platform protecting the AI estate. | | 5 Security | Zscaler | AI security (network) | Zero-Trust network access; secures cloud-first/AI traffic. | | 6 Workflow | ServiceNow | AI workflow automation | The system-of-action embedding agentic AI into enterprise processes. | | 6 CRM | Salesforce | AI CRM | Agentforce; embeds AI into the customer-facing workflow. | | 7 Consumer | Meta Platforms | AI consumer apps | 3B+ users; AI recommendation = ad dollars. | | 7 Consumer | Netflix | Digital media | AI personalisation/content; streaming scale. | | 7 Consumer | Roblox | Future digital consumer platform | UGC + AI creation tools; the next-gen engagement platform. |
The Key Insight, One Transaction, Many Holdings#
When Microsoft signs a $100m Azure AI deal:
- Microsoft books the revenue (Layer 4). ✅ held
- It orders GPUs from NVIDIA (Layer 2). ✅ held
- NVIDIA's chips are made by TSMC (Layer 1). ✅ held
- TSMC uses Disco tools (Layer 0). ✅ held
- The GPUs need HBM from Micron (Layer 2). ✅ held
- The cluster is networked with Marvell silicon (Layer 2). ✅ held
- The datacentre draws power from Constellation/Vistra (Layer 3). ✅ held
- The workload is secured by CrowdStrike/Zscaler, monitored by Datadog, and runs on data in Snowflake (Layer 5). ✅ all held
- The customer surfaces AI via ServiceNow/Salesforce (Layer 6). ✅ held
A single AI deal lights up ~12 holdings. This is portfolio design, not coincidence. It is why the fund reads hyperscaler capex as a leading indicator for the entire spine → Cross-Holding Read-Throughs.
Where The Value Migrates Over Time#
The fund's edge is owning the whole chain because value migrates between layers as the cycle matures:
- 2023-2025 (build-out): value concentrates in compute, memory, foundry, power (scarcity rents). NVIDIA/TSMC/Micron/CEG lead.
- 2025-2027 (deployment): value shifts toward cloud + software as enterprises monetise AI (Microsoft, ServiceNow, Snowflake, CrowdStrike).
- 2027+ (consumption): value shifts toward the consumer/application endpoint (Meta, Alphabet, Roblox, Netflix) where AI drives engagement & ad/subscription dollars.
By owning all layers, PCA SOF doesn't have to time the migration, it rides it.
Linked Notes#
- Related Holdings: every holding listed above.
- Maps: Knowledge Graph · AI Supply Chain Map · AI Ecosystem Map · Competitive Landscape Map
- Themes: Artificial Intelligence · AI Power Demand · Cloud Computing · Semiconductors
- Construction: Portfolio Construction Overview · Return Drivers vs Hedges
- Risk: AI Capex Cycle Risk · Taiwan Strait Risk