AI Stack Deep Dive
Dissecting the AI Supply Chain
Where value accrues when intelligence becomes cheap but deployment stays constrained.
3 May 2026 · YK Research
Contents
The Takeaway
🔍
“AI is deflationary at the model layer and inflationary everywhere below it. The trade is not simply 'buy AI apps'. The trade is to own the toll roads every token, robot and agent must cross.”
YK Research
Stack layers
6
Most scarce layer
Power
Most copied layer
Apps
Core question
Who owns chokepoints?
Chamath's framing is useful because it forces us to stop treating AI as one sector. AI is a stack. Each layer has a different supply chain, capital requirement, moat, margin profile and failure mode. The market will overpay for visible apps when cycles get hot, but the durable value should accrue to constrained layers: power, cooling, lithography, memory, advanced packaging, proprietary data, runtime orchestration and physical-world actuation.
Source: Chamath Palihapitiya / Social Capital article, “Deep Dive: Where Value Accrues in the AI Stack”, 1 May 2026; YK Research synthesis.
Defensibility by Layer
Scored directionally. Lower layers are harder to disrupt because they require physical assets, supply-chain control, scarce expertise and balance-sheet scale. Application companies can still win, but only if they turn usage into proprietary workflow data or distribution lock-in.
Bottleneck Severity
The strongest moats are not always in the most exciting product demos. They are often in boring constraints: grid interconnects, chillers, EUV tools, HBM, CoWoS slots and proprietary datasets.
Hyperscaler Capex: The Stack Is Becoming Industrial
The AI stack is no longer asset-light software. Amazon, Microsoft, Google and Meta are guiding toward roughly $725B of 2026 capex. Add OpenAI, xAI, Oracle, CoreWeave and sovereign AI programs, and the spending cycle starts to look like railroads, electrification or telecom buildout — not classic SaaS.
Source: All-In discussion of 2026 hyperscaler capex guidance; company earnings commentary; YK Research directional chart.
1. Infrastructure: The Real Bottom of the AI Stack
- Supply chain: land → grid interconnect → generation → transformers/switchgear → cooling → data-center shell → operations.
- Bottleneck: power availability and time-to-energize. GPUs can be ordered faster than substations, turbines and grid capacity can be built.
- Why value accrues: AI labs need guaranteed megawatts/gigawatts. The owner of reliable, near-term power becomes the gatekeeper.
- Watch: gas turbines, nuclear restarts, behind-the-meter power, liquid cooling, transformers, high-voltage equipment and modular data-center construction.
Public-market hooks
GEV, ETN, PWR, VRT, TT, CARR, CAT, BE, CEG, NEE, DUK, LNG, copper/uranium exposure
Main risk
Permitting, cyclic overbuild, power-price politics, capex delays, utility interconnect queues.
2. Chips: Compute Is a System, Not Just a GPU
- Supply chain: EDA/IP → lithography/equipment → foundry → HBM → advanced packaging → networking → racks → cluster software.
- Core chokepoints: ASML EUV, TSMC leading-edge capacity, HBM3E/HBM4 supply, CoWoS/advanced packaging, high-speed networking.
- NVIDIA moat: CUDA + systems + networking + software ecosystem. The moat is not just silicon performance; it is time-to-working-cluster.
- ASIC fork: Google TPU, Amazon Trainium, Microsoft Maia and Meta MTIA reduce hyperscaler dependence on NVIDIA, but often expand total silicon demand through cheaper inference.
Public-market hooks
NVDA, AVGO, AMD, TSM, ASML, AMAT, LRCX, KLAC, MU, Hynix, Samsung, ARM, MRVL, ANET
Main risk
Export controls, HBM shortage, CoWoS bottlenecks, custom ASIC cannibalization, China/Taiwan geopolitics.
3. Data: From Internet Scraping to Proprietary Signal
- Digital AI data: text, code, images, video, enterprise records, user behavior and workflow traces.
- Physical AI data: cameras, LiDAR, torque, friction, collisions, warehouse movement, driving miles, factory telemetry and simulation.
- Shift: generic internet data is saturated. Scarcity moves to proprietary workflow data and embodied physical-world data.
- Moat test: does each customer interaction improve the product in a way competitors cannot copy? If yes, data compounds. If no, data is just a commodity input.
Public-market hooks
GOOG, META, AAPL, TSLA, AMZN, MSFT, PLTR, Bloomberg/ICE/SPGI/MCO-type data networks
Main risk
Data saturation, copyright/regulatory constraints, synthetic-data collapse, privacy and customer-data lockup.
4. Models: Powerful, But Margin Compression Is Real
- Supply chain: data → training compute → model architecture → alignment/evals → inference serving → API distribution.
- Deflation: model capability gets cheaper over time. DeepSeek-style releases accelerate price pressure and reduce the standalone model margin pool.
- Where value survives: frontier reliability, tool-use quality, enterprise security, domain-specific post-training, distribution bundles and compute availability.
- Investment read: model labs can be huge, but they are capex-hungry and exposed to margin compression unless tied to cloud, data, distribution or agent execution.
Public-market hooks
OpenAI/Microsoft, Anthropic/Google/Amazon, xAI/Tesla, Meta, Google, Mistral, DeepSeek ecosystem
Main risk
Price compression, open-source catch-up, compute burn, regulatory liability, customer multi-homing.
5. Execution: The Agent and Robot Control Plane
- Digital execution: memory, tool routing, API calls, browser/computer use, code execution, workflow orchestration and permissioning.
- Physical execution: sensor fusion, planning, kinematics, batteries, motors, end effectors, safety systems and real-time edge inference.
- Why this layer matters: a smart model that cannot take safe action is a demo. Execution turns intelligence into economic output.
- Moat: deep workflow integration, audit logs, trust, deterministic guardrails and access to the operating system of the enterprise or machine.
Public-market hooks
MSFT, GOOG, AMZN, PLTR, NOW, CRWD, PANW, TSLA, ISRG, SYM, Teradyne, Rockwell
Main risk
Reliability failures, permissions/safety, low switching costs if orchestration is generic, liability in physical-world actions.
6. Applications: Highest Visibility, Weakest Default Moat
- Supply chain: model/API → workflow UX → customer data → distribution → retention loop.
- Good apps: own a painful workflow, have distribution, create proprietary data, and sit close to budget authority.
- Bad apps: wrappers around commodity models with no data flywheel and no switching cost.
- Portfolio implication: app winners exist, but the underwriting hurdle is higher. Demand proof of workflow ownership, not just impressive demos.
Public-market hooks
ADBE, CRM, NOW, INTU, DUOL, Figma/Canva-style private comps, vertical AI apps, Harvey-like legal AI
Main risk
Feature-copying by model providers, seat-price pressure, customer churn, app-store/platform dependency.
Portfolio Read: Follow the Forced Flow of Dollars
Highest conviction buckets
- Power equipment, grid gear, cooling and data-center infrastructure.
- HBM, advanced packaging, leading-edge foundry and semi equipment.
- Vertically integrated platforms with data + distribution + compute.
Selective buckets
- Cloud providers that can convert capex into sticky workloads.
- Security platforms using AI to harden code and infrastructure.
- Physical AI names with real embodied data and manufacturing scale.
Avoid / short-list risk
- AI app wrappers with no proprietary workflow data.
- Model-only exposure without distribution or compute advantage.
- Hardware beneficiaries priced as if no cycle risk exists.
🔍
“If a company does not control a bottleneck, a workflow, proprietary data or distribution, assume its AI margin gets competed away.”
Process rule
This is a framework, not a buy list. Position sizing still matters. The correct trade can still lose money if bought at the wrong price or sized like a prediction instead of a probability.