US AI Value Chain · Investment Map

From Power to Applications:
An Investment Decision Map of the US AI Value Chain

Mapping 38 leaders along "infrastructure → compute services & models → application monetization": core products, moats, investment thesis, key metrics and main risks, alongside ETF tools and a signal calendar. This map complements the "US Semiconductor Value Chain Map" — the semiconductor chain is a zoomed-in view of this map's "compute chips" segment.

Data as of: early June 2026 · Market figures are approximate; verify live data before placing any order
Global data-center investment (next 5 years)
≈$3T
Moody's forecast, early 2026
Top-5 hyperscaler 2026E capex
≈$700B
Analyst consensus; the funding source for the whole chain
US data-center power demand
41→66GW
Goldman: 2026E→2027E, doubling in two years
CoreWeave revenue backlog
≈$100B
2026 capacity already pre-sold out
OpenAI's latest funding round
$100B-scale
A yardstick for funding intensity at the model layer
01

Value Chain Map: Compute Flows Right, Money Flows Left

The AI chain has one extra judgment versus the semiconductor chain: money must not only flow (hyperscaler capex → infrastructure orders) but ultimately come back (application monetization → funding the next round of capex). So this chain has only one ultimate question — can the money earned at the application layer pay for the investment at the infrastructure layer. Click any segment and the company list below filters automatically.

Upstream · Compute & InfrastructureMidstream · Compute Services & ModelsDownstream · Applications & Monetization
UPSTREAM INFRASTRUCTURE
→
MIDSTREAM COMPUTE & MODELS
Foundation Models (mostly private)OpenAI · Anthropic · xAI ↓ see note below
→
DOWNSTREAM APPLICATIONS
Compute flow → power / chips / clusters / models / apps ← Money flow: app monetization / cloud capex / infra orders (the cash-return loop is the core question)
Why the model layer has no company cards: the leading foundation-model companies are largely private — OpenAI (latest funding at the $100B scale, with huge compute contracts signed with Microsoft/Oracle/CoreWeave), Anthropic (Amazon and Google as principal investors), xAI, and others. Public-market exposure is indirect: MSFT (OpenAI stake and revenue share), AMZN / GOOGL (Anthropic equity), NVDA (strategic investment portfolio), ORCL / CRWV (compute-contract revenue from model companies). Note: private-market valuations lack liquid pricing, transmission to public markets carries discounts and lags, and any cooling in model-layer funding hits compute-rental firms' order expectations first.
Zoom-in · Breaking down the compute-chip upstream
The four gates a Blackwell-class AI chip must pass through to be born — any one of them can choke the entire AI chain. Click each stage to filter companies:
→ → → →
The three physical bottlenecks of AI compute today all sit on this sub-chain: EUV lithography capacity (ASML's monopoly), CoWoS advanced packaging (TSMC), and HBM supply (Micron and two others). For a fuller breakdown of equipment, materials, analog chips and more, use this alongside the "US Semiconductor Value Chain Investment Map."
02

Leading Company Profiles

Moat ratings are qualitative judgments (5-point scale: a composite of market structure, switching costs, technology generation gap, customer lock-in, and margin durability). Click a card for the full profile: investment thesis, key metrics, main risks, and corresponding ETF exposure. For a fuller breakdown of the compute-chip segment (equipment, foundry, memory, packaging/test), use this alongside the semiconductor map.
03

ETF Tools: How to Allocate Beta to the AI Theme

The common flaw of AI-theme ETFs is "a bit of everything": before buying, check the top-10 holdings to confirm which segment of the chain it's actually betting on. Unlike semiconductors, the AI chain has no widely accepted high-purity index, and a broad index (QQQ) is in fact many people's de facto AI position. Expense ratios are approximate; refer to the issuer's official site.

Ticker Theme Focus Expense ≈ Structure Best For
SMH Semiconductors (upstream AI compute) 0.35% Highly concentrated in NVIDIA + TSMC; the purest beta on "compute supply" Want to bet only on the compute-hardware segment
AIQ Global X Artificial Intelligence & Technology 0.68% A large-cap tech basket covering chips + cloud + apps; diversified but dilutes purity Want a whole-chain basket without weighting it yourself
IGV Software industry (proxy for the AI app layer) 0.40% Software heavyweights like Microsoft/Salesforce/Palantir; rides the "app monetization" logic Bullish that AI pays off in software rather than hardware
BOTZ Robotics & automation (physical AI) 0.68% Global robotics/automation names, incl. Japanese industrial control; highly correlated with the AI narrative A long-term option on embodied/physical AI
QQQImplicit Exposure Nasdaq 100 broad index 0.20% Semiconductor weight already ~33%, the Magnificent Seven even higher — already heavily AI Calculate its implicit AI exposure first, then decide whether to layer on the theme ETFs above
04

Decision Signal Calendar: What to Watch, and When

The AI chain has two unique signal types beyond the semiconductor chain: model-capability leaps (re-pricing the app layer) and credit markets (re-pricing the highly leveraged infrastructure layer). Ordered by importance.

Quarterly · end of Jan/Apr/Jul/Oct
Hyperscaler capex guidance + AI revenue disclosure
The chain's most important dual indicator: capex guidance determines upstream orders, while AI revenue growth at Azure AI / AWS / Google Cloud determines whether the spending is sustainable. The "scissors gap" between the two is the core variable the market prices.
Quarterly · late Feb/May/Aug/Nov
NVIDIA earnings and conference call
Still the market's bellwether, but AI-chain investors should watch one extra detail: the changing share of "neoclouds / sovereign AI / model companies" in the customer mix — it reflects whether demand is broadening or concentrating.
Ad hoc · re-prices on release
Major model releases and capability leaps
Major versions of GPT / Gemini / Claude are "re-pricing events" for the app layer: each capability leap cannibalizes the moat of a batch of thin-shell apps (wrappers) while spawning new compute consumption. The impact is direct on app stocks like PLTR/NOW/APP.
Quarterly + real-time bond market
Compute-rental backlog and credit metrics
CoreWeave's backlog (≈$100B) and Oracle's RPO and leverage (total debt already at $124B, negative rating outlook) are the thermometer of AI infrastructure credit risk. Widening CDS spreads often lead the share price.
Monthly/quarterly · policy windows
Power supply and data-center permitting/grid connection
Goldman expects US data-center power demand to double in two years (41→66GW), and power is becoming a harder bottleneck than chips. Watch each state's grid-connection queue data and PPA contract prices — they determine power stocks' earnings leverage and the physical ceiling on compute expansion.
Daily
Rate path + regulatory developments
The AI chain has extremely long duration (investment first, returns later) and is highly rate-sensitive; export controls affect the compute-supply side, and AI regulation (EU AI Act enforcement, copyright litigation) affects the app layer's cost structure.
05

Risk Matrix: Systemic vs. Segment-Specific

The AI chain adds two risk types the semiconductor chain lacks: circular dealing/credit leverage (systemic) and "models cannibalizing apps" (segment-specific). Segment-specific risk can be hedged structurally: hold both the GPU and ASIC chains, pair infrastructure with applications, and balance highly leveraged rental firms with cash-rich cloud giants.

Systemic Risks (Chain-Wide Resonance)

  1. Monetization scissors gap — $3T of infrastructure investment over five years vs. AI revenue still ramping. If app monetization keeps lagging, capex downgrades would amplify along the whole chain "cloud → chips → equipment → power." This is Wall Street's number-one current disagreement.
  2. Circular dealing and credit leverage — chipmakers invest in model companies, model companies order from cloud providers, cloud providers borrow to buy chips: Oracle's debt is $124B with a negative rating outlook, CoreWeave expands on heavy debt, and a funding break in any link would transmit in a chain reaction.
  3. Power and physical bottlenecks — if power demand doubling in two years runs into grid-connection/permitting/transformer-supply constraints, compute deployment will be rate-limited by the physical world, and chain-wide growth expectations downgrade in sync.
  4. Valuation and rates — long-duration assets are rate-sensitive; under stretched valuations, a reversal in risk appetite drags everything down (the SOX falling over 5% in one day on May 12 is a sample).
  5. Regulation and geopolitics — export controls limit the supply side, while AI safety regulation and copyright litigation raise costs at the model and app layers.

Segment-Specific Risks (Structurally Hedgeable)

  1. GPU vs. ASIC share battle — more and more buyers turn to AMD or to Broadcom-co-designed in-house TPU-class chips, hitting NVIDIA and the rental firms within its ecosystem (CoreWeave is suppressed by this narrative).
  2. Compute-rental model validation — orders are full but losses are widening (CoreWeave's Q1 profit plunged, with guidance below expectations): the "high growth, high debt, high capex" triple-high model has yet to prove it can make money.
  3. Models cannibalizing apps (wrapper risk) — each model capability leap renders a batch of apps lacking data/distribution/workflow moats worthless; app stocks must be checked one by one for "does it get stronger or weaker when the model gets stronger."
  4. The AI double-edged sword for SaaS — AI may bring new pricing (outcome-based billing) to software stocks, or it may dismantle the old "per-seat" model; the bull-bear split is large.
  5. Overextended expectations for power stocks — the AI premium for nuclear/independent power producers is already heavily priced in, with sharp drawdowns if PPA signings disappoint or power prices fall back.