The AI trade is priced today with revenue multiples, yet the three forms being priced (IDMs that build and operate for themselves, Foundries that build compute and sell it, Fabless labs that rent compute and sell intelligence) have entirely different cash structures. One company's news can move the whole sector because the market lacks unit data it can observe directly. Research #17 proposes two things in response: a valuation KPI matrix spanning a financial layer, a signal layer and a physical layer (Table 11), and a tiered disclosure standard that would make those metrics readable straight from the filings (Table 12). Until that disclosure arrives, this page splices and computes the readings each quarter from public filings and publishes them as a public reference for the transition. Whoever discloses first earns the computability premium first. Definitions, derivations and all sources are in Research #17; a version graded against alert thresholds, with scenario testing, is the KPI dashboard.
A unified physical-layer measure already has parallel work in academia: Stanford's intelligence per watt project (Hazy Research and the Scaling Intelligence Lab; PIs include Christopher Ré, Azalia Mirhoseini and John Hennessy) proposes task accuracy per watt as the north-star metric for model architecture, hardware design and national strategy, and attributes the roughly 5.3× efficiency gain from 2023 to 2025 to algorithm-side and hardware-side contributions (as reported by the project). This page extends the same chain of measurement to the capital-markets end: engineers choose architectures by intelligence per watt, investors price assets by revenue per GW, and the two ends are joined by Equation 1 (revenue = GW × utilization × tokens/joule × $/token).
ConsolidatedIndustry P&L after eliminations
Coverage on three bases against the reference line of 1
Coverage = external revenue ÷ the denominator named on each row. Readings are spliced and computed for the 2026Q3 release and are InAI estimates; every bar starts at zero, and for a range the lighter segment runs from the low end to the high end.
FinancialWhat the filings show
SignalsClosed-loop and structural signals
Vendor financing ratio Lu against the telecom red line
Same basis throughout: financial support to a single customer ÷ that customer's purchases from the supplier. NVIDIA–OpenAI is the largest current case (PORTS-Pike campus guarantee basis: grade two, not yet in effect; 15–17% on the funded-equity basis; the earlier letter-of-intent basis is void). The three ANT points are on the facility basis as filed in the S-1/A (supplier note facilities ÷ purchase commitments to that supplier, all undrawn as of 2026-09-30). The red line is the Nortel case of 1999–2001.
NVIDIA → OpenAI (guarantee basis)Red line 130%62–71%
Broadcom → ANT (facility basis)≈ 26%
Google → ANT (facility basis)≈ 16%
Amazon → ANT (facility basis)≈ 14%
Telecom endgame, 1999–2001130%
PhysicalWhat the equations run on
MediaWorking with the press
Working with reporters and editors
Every reading on this page may be quoted with attribution (InAI Capital).
For reporting on the economics of AI infrastructure, InAI can provide reporters and editors with: complete data charts at print resolution; filing-based analysis worksheets (SPCX 8-K/10-Q, Anthropic S-1/A); a demonstration of the interactive KPI dashboard (graded automatically against alert thresholds, with scenario testing); the output of the three-scenario model with reproducible scripts; and written answers and background interviews in English and Chinese. Contact: [email protected]; please name your outlet and deadline.