Scrutica
capxy field into incremental quarterly capex: Q1 = capxy, Q2–Q4 = capxy[fqtr] − capxy[fqtr−1] within the same fiscal year. Where the prior quarter sits outside the extract window, the incremental cell is NULL (never a back-fill guess; DQP §1). Sum-of-4-quarters reconciles within ±$2M of the source's annual capx across 180 / 180 closed fiscal years; the trailing-4-quarter cohort line aggregates absolute capex across the six hyperscalers (MSFT, AMZN, GOOGL, META, ORCL, AAPL). Companion view: Compute Cost Index for the $/petaFLOP-day price layer.Source. Quarterly corporate fundamentals from a licensed financial-fundamentals database, held under academic subscription and not redistributed. Underlying primary documents are the 10-Q and 10-K filings each issuer makes with the SEC; the database re-serves these with standardised column codes (capxy, revtq, xrdq, ppentq). The licensed re-serve is authority tier 2 — structurally identical to SEC XBRL primary measurement, one step removed by the subscription layer.
YTD decomposition. The capxy field is YEAR-TO-DATE capital expenditure that resets to zero at the start of each fiscal year. A naïve quarter-over-quarter difference at the fiscal- year boundary produces a large negative number; the correct decomposition is Q1 = capxy, Q2–Q4 = capxy[fqtr] − capxy[fqtr−1] within the same fiscal year. Where the prior-quarter row falls outside the extract window (typically the first fiscal year for late-listed issuers like ARM and GFS), the incremental cell is NULL — never a back-fill guess (DQP §1).
Reconciliation.Sum of four quarterly incrementals reconciles within ±$2M of the source's annual capx across 180 / 180 closed fiscal years. 14 partial-FY buckets (one or more NULL quarters) and 11 still-open FY2025 / FY2026 buckets are honestly unreconcilable rather than guessed. The full reconciliation audit is retained in the build artifacts under data/processed/; see licensed-database provenance.
Cohort selection. The 25 anchor tickers are fixed by the fundamentals-extract pipeline — the AI-infrastructure-relevant cohort spanning hyperscalers, chip designers, foundries, semiconductor equipment, memory, networking, colocation REITs, and AI-DC power infrastructure. The chart aggregates the six hyperscalers (MSFT, AMZN, GOOGL, META, ORCL, AAPL); the per-issuer table below carries the latest-quarter snapshot for the full cohort.
X-axis alignment. Bars stack by calendar-quarter end (the source's datadate), not by fiscal-quarter label, so MSFT FY24Q4 (2024-06-30) stacks alongside GOOGL FY24Q2 (2024-06-30) on the same bar even though their fiscal-quarter labels differ. Apple's fiscal year ends in late September; AAPL Q4 rows are at calendar Q3 dates and stack accordingly.
Resolution chain. The vendor-key → canonical org_id mapping is materialised in a cross-reference table per DQP §10 (lookup data in DB, not Python dicts) and DQP §11 (write-time resolution). Future fundamentals ingests resolve via this table, not a hand-curated mapping.
29 of 781rows have a NULL incremental-capex cell — the prior quarter's YTD ledger is outside the extract window for that issuer's first fiscal year, so the Q-over-Q difference is not computable. Per DQP §1 these are not back-filled with guesses; the missing quarters surface as gaps in the chart and are excluded from cohort sums for those periods.
| FY2025 Q4 · 2025-12-31 |
| $27.9B |
| $113.8B |
| 24.5% |
| MetaMETA | FY2025 Q4 · 2025-12-31 | $21.4B | $59.9B | 35.7% |
| OracleORCL | FY2025 Q3 · 2026-02-28 | $18.6B | $17.2B | 108.4% |
| AppleAAPL | FY2026 Q1 · 2025-12-31 | $2.4B | $143.8B | 1.7% |