Sources & Methodology

The follow-the-dollar strip

The strip on every layer page — where $1.00 of AI application spend lands — is an illustrative, directional model, and nothing more. No public dataset measures the stack's actual per-layer revenue split, so the strip states its assumptions instead of implying precision. Each slice is anchored to a named public reference:

• Data Centers + Compute ≈ $0.30 The largest slice, because compute hosting dominates the cost of serving a token in public inference-TCO teardowns — SemiAnalysis' published cost-of-inference and GPU-rental analyses are the anchor.
• Silicon ≈ $0.22 Accelerators run roughly half of a data center's build cost in public cost breakdowns, and NVIDIA's gross margin — mid-70s percent per its public quarterly filings — keeps most of that hardware dollar with the silicon makers.
• Models ≈ $0.18 API billing sits above serving cost: public inference-margin estimates (the SemiAnalysis figures cited on the Models page) put what labs bill meaningfully above the compute bill beneath it.
• Apps + Agents ≈ $0.14 Software margin above the model bill, anchored only to vendor pricing pages — thin evidence, and the strip treats it as such.
• Harness ≈ $0.10 The same software-margin logic one layer down, split with apps. This is the thinnest-evidenced slice on the strip.
• Land + Power + Shell ≈ $0.06 Power and land run a single-digit share of data-center TCO in public analyses — LBNL's DOE report to Congress and published data-center cost breakdowns are the anchor.

The slices are rounded to sum to $1.00 and are directional, not measured. Where a layer page reads a different conclusion from the same strip, the page says so in its own words.

The chart data contract

Every real chart on this site reads a hand-curated static JSON committed to the repo — data/models-prices.json, data/harness-uplift.json, data/lps-capacity-forecasts.json, data/apps-pricing-timeline.json — each carrying a sources field with full citations. Nothing is fetched live. Where no defensible public series exists yet — silicon-chain concentration, a continuous GPU-rental price series — the page ships an honest placeholder instead of a blended curve.