Skip to content
⌘K
Special Topics ▾
AI InfrastructureDefenseQuantum

Methodology — AI Study (the value-chain map)

A diagnostics-only map of the AI-infrastructure value chain. It answers which stack layer each name sits in and compares the cohort on currency-neutral ratios — never a score, never a ranking, no predictive claim. The cohort is selected for being AI-exposed, so any "strong fundamentals" reading is mechanically correlated with realized winners; sorting it by attractiveness would be spurious alpha by eyeball. The exact membership rule and its lexicons are proprietary and not enumerated here.

Membership — a deterministic rule, not a hand-list

A name is assigned to layers by a deterministic classification over its sector and business description, with false-positive guards — not by manual curation. Names can belong to multiple layers (a diversified semiconductor firm can span silicon, networking and software); the "primary" layer is just the most-upstream match — a sort key, not a verdict. Capex/R&D intensity are shown as descriptive attributes, never as a membership gate (gating on them would misclassify names with data gaps). The layers run EDA/design → silicon → foundry → equipment → memory → optical → datacenter-infra → neocloud → datacenter-operators → hyperscale-cloud → AI-power → platform/apps.

The derivations — currency-neutral only

Absolute-dollar levels (market cap, revenue, capex) are not comparable across a cohort spanning KRW/JPY/TWD/EUR/USD reporters, so the map uses ratios only (an absolute cross-currency total would be a wrong number). Each is a labeled diagnostic, none is a composite:

  • Self-fund coverage — operating cash flow relative to capex, the capital-cycle fault line (who funds the build internally). Suppressed for asset-light layers where capex is near zero and the ratio explodes.
  • FCF margin → true FCF margin (FCF after deducting stock-based compensation) + SBC intensity — earnings-quality reads.
  • Gross-dilution rate — stock-based compensation relative to market cap, a split-immune dilution proxy (a raw share-count CAGR conflates splits and would be a wrong number).
  • Gross / operating margin — the profit-pool-by-layer profile.
  • RPO coverage — the contracted backlog relative to revenue and to capex: contracted demand vs the buildout (capex-heavy layers only).
  • Incremental revenue per unit of capex — a labeled build-phase conversion read (low = pre-revenue build, not necessarily fading).
  • Implied depreciation life — an earnings-quality read (see the Hyperscaler Lens).

What the adversarial review deliberately dropped

Rule-of-40 (a composite score whose margin half our own backtests found to be a dead anti-signal), a "profit migrating into silicon" trend (a single-name artifact), share-count-CAGR dilution (split-conflated), and all leaderboard ranking/coloring on a selected winner cohort. Sorting the table is a neutral lookup tool; no column is a good/bad verdict. Our own multi-year backtest found almost nothing here robustly predicted returns.