Methodology — Hyperscaler Lens

A NON-valuation reading room for the AI-infrastructure cohort, shown beside the intrinsic value. It answers a question a DCF can't: does the price lead or lag the contracted backlog and the capital cycle? Diagnostics-by-default — every number is read from a stored filing fact, proxies are labeled, and nothing is scored. The cohort definition and any internal statistical calibration are proprietary and not enumerated here.

Why not a score: the core hyperscaler names trade so tightly together that the cohort carries only a couple of independent bets' worth of information — no cohort-scoped statistic could be significant. The lens shows and labels; it never ranks.

Four surfaces

1. Contracted backlog (RPO). Remaining Performance Obligation — the dollar value of contracted-but-unrecognized revenue that filers must disclose under the revenue-recognition standards (ASC 606 / IFRS 15) — the cleanest forward-demand number for a cloud/subscription business. It is a point-in-time balance, stored per period-end, and read within-name only: coverage (the backlog relative to revenue — years of visibility) and YoY growth. RPO is never differenced across names — one filer's backlog can be concentrated mega-deals while another's is diversified; they aren't comparable. A product company reporting ~no RPO is itself the signal.

2. Expectations vs reality. The reverse-DCF embedded implied growth (what the price requires) shown beside what demand is doing (RPO growth, capex growth). Juxtaposed, never subtracted — the legs are incommensurable (a multi-year CAGR vs a one-year backlog delta). A neutral lead/lag reading, never a score.

3. Circularity reflexivity. The AI-silicon vendor's data-center revenue set against the largest hyperscalers' cash capex — one labeled aggregate for how reflexive the AI-capital cycle is (the silicon vendor's order book is the build budget). Customer concentration and vendor financing are text-only disclosures, left as explicit no-coverage — never imputed.

4. Depreciation-life sensitivity (an inflation guard). Hyperscalers extended server useful lives even as accelerator economic life may be shorter, which lowers depreciation and flatters margins. We recompute operating margin under a shorter assumed asset life, using the standard implied-life estimate (gross plant relative to annual depreciation). The scenario only ever lowers the margin (the conservative direction) and never feeds the fair value. For IFRS filers that tag only net PP&E it falls back to a labeled "remaining book life" basis; where the implied life is implausibly short (the gross line is understated) the scenario is suppressed rather than shown.

Honest limits

RPO staleness is surfaced (a filer that stopped tagging it is shown dated, never extrapolated). The cohort is a scoped membership, not the whole market. All four surfaces are display-only and never move an intrinsic value.