AI Investment Is not an Accounting Variable: Evidence from Hyperscaler Disclosures

Abstract

U.S. financial statements contain no standardized “AI investment” line item, although market commentary routinely compares firms’ AI spending. This paper studies the resulting measurement problem using 40 earnings events for Amazon, Alphabet, Meta, Microsoft, and Oracle from September 2024 through July 2026. Cash capital expenditure is observable in all 40 events; numeric forward capitalexpenditure guidance appears in 14, comparable same-target-year revisions in nine, and lease-liability balances in 32. Attribution is substantially less stable than the underlying numbers. The author positively attributes all 14 guidance events to AI-related investment, whereas a label-blind independent languagemodel recoding identifies positive support in nine. Exact grade agreement is 6 of 14 (42.9 percent; Cohen’s kappa = 0.026), and high-versus-not-high agreement is 7 of 14 (50.0 percent; kappa = 0.109). The disagreement is reported as a human–model classification-stability diagnostic, not conventional human inter-coder reliability. The paper introduces an AI Investment Observability Lattice that separates recognition, attribution, horizon, and consolidation boundaries. A formal non-additivity result shows that cash flows, liability stocks, forecasts, and contractual schedules cannot be combined into a presentationinvariant scalar without explicit transformations to a common valuation date, allocation rule, and perimeter. No standardized event-level AI-specific payoff is observed. The evidence is descriptive and intensive rather than representative or causal.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.