The AI buildout has not universally crossed over to the capital markets — Microsoft's FY2026 cash capex of $115.9bn sat inside $182.9bn of operating cash flow, with a $678bn contracted commercial book behind it (SEC filing, 29 Jul 2026)

Why it matters (spine). The 28 July sweep carded Alphabet's Q2 2026 exhibit (dated 22 July) as the cleanest evidence yet that the AI buildout had outgrown operating cash flow (alphabet-q2-2026-ai-capex-exceeds-cash-flow: $44.9bn of quarterly capex against $39.1bn of operating cash flow, free cash flow of negative $5.9bn, buybacks at zero). One week later Microsoft filed the counterexample, and RDKG's discipline requires carding it rather than keeping the tidier half of the picture. For the fiscal year ended 30 June 2026, Microsoft's cash outflow for additions to property and equipment was $115,948 million against $182,935 million of net cash from operations. In the June quarter alone, capex was $35,802 million against $55,441 million of operating cash flow, and the company still returned $10.2 billion to shareholders in dividends and repurchases. The crossover Alphabet showed is real, and it is company-specific, not a law of the sector.

A definitional caution on that ratio. The cash capex line puts the year's spend at roughly 63% of operating cash flow, but that understates the buildout. Gross property and equipment — net plus accumulated depreciation — rose from $298,619 million to $431,767 million over the year, an increase of $133,148 million, and more than that if anything was retired. So at least ~$17 billion of additions arrived outside the cash capex line, by finance leases or similar. On a gross-asset basis the year's additions consumed 73% or more of operating cash flow. It was still funded from operations; the margin is thinner than 63% suggests, and any comparison against Alphabet's cash-capex figure is comparing the narrow line on both sides.

What sits behind the spend. Commercial remaining performance obligation rose 84% to $678 billion — contracted, not-yet-recognised customer commitments across the whole commercial book, not compute alone. Microsoft Cloud revenue was $59.3 billion, up 27%, with Azure and other cloud services up 43%. Read together: capacity is being built against a large contracted book, which is the behaviour of a utility being provisioned rather than a purely speculative land grab. Money angle for RDKG: this is a sharper version of the moat argument, not a weaker one. If frontier capability is a contracted utility bought on multi-year terms and delivered on assets that depreciate, then capability is an operating expense every competitor can rent at a published price — and the thing that cannot be rented is a proprietary, non-toxic, hard-to-synthesise, freshly-dated corpus. Metacrisis lever: property and equipment net rising from $204,966 million to $313,076 million in twelve months is a physical claim on power, water, land and materials; an auditable record of its true scale is a precondition for any honest public argument about whether it is worth it.

The depreciation clock (the part that does the work). Accumulated depreciation on that asset base reached $118,691 million, up from $93,653 million a year earlier, and full-year depreciation, amortisation and other was $38,534 million against $29,433 million the year before. A young, fast-growing asset base means the depreciation charge is still climbing. That is the honest mechanism by which "capability is a rented, depreciating input" holds even for a company funding it from operations — the cost does not disappear, it arrives later and on a schedule.

What would prove it wrong (falsifier). The facts fail if Microsoft's FY2026 Exhibit 99.1 does not show additions to property and equipment of $115,948 million, net cash from operations of $182,935 million, or commercial RPO up 84% to $678 billion. The interpretation — "funded from operations, against a contracted book" — fails if subsequent quarters show capex exceeding operating cash flow (the Alphabet pattern arriving late), if RPO growth decelerates sharply while capex keeps climbing, or if the backlog proves concentrated and long-dated enough that "contracted" overstates near-term cash conversion. On that last point the release is silent by design: it discloses no weighted-average duration for the $678 billion, no split between compute and software, and no counterparty concentration. This card therefore asserts nothing about how much of the backlog relates to any single customer, and nothing about how much of it is Azure.

The honest caveat on headline profitability. GAAP net income was $35,766 million for the quarter (+31%) and $133,749 million for the year (+31%), but Microsoft itself strips investment effects out: on a non-GAAP basis excluding the impact of investments in OpenAI, quarterly net income was $35,286 million (+22%) and full-year $128,786 million (+22%). The release names OpenAI's contribution as +$480 million of net income and +$0.07 of diluted EPS in the quarter, and +$4,963 million and +$0.67 for the fiscal year. It separately discloses a $3.2 billion gain on its investment in Anthropic among discrete items that together benefited diluted EPS by $0.27 versus prior guidance, partly offset by "severance expense and impairment charges in XBOX." Equity and other investments on the balance sheet more than doubled, from $15,405 million to $36,348 million. Anyone quoting 31% earnings growth as an operating result is quoting equity marks, exactly as with Alphabet's Q2.

Company-stated metrics (flagged, not audited). Satya Nadella's "This year, Azure revenue surpassed $100 billion for the first time, and Microsoft 365 Copilot reached over 30 million paid seats" is a first-party assertion inside a filed exhibit — usable with attribution, but neither figure is broken out in the financial statements and no methodology is disclosed. "For the first time" is Microsoft's own characterisation. Also first-party and unaudited: "We are advancing the frontier on the cost-to-outcome curve, ensuring every customer can turn tokens into business results." Amy Hood is quoted on "Microsoft Cloud revenue of $59.3 billion, up 27% year-over-year."

Provenance (what I actually read). Fetched Exhibit 99.1 to Microsoft Corporation's Form 8-K for the quarter and fiscal year ended 30 June 2026 from SEC EDGAR in session on 31 July 2026; CIK 0000789019, accession 0001193125-26-323632. The document is datelined "REDMOND, Wash. — July 29, 2026" and states "All information in this release is as of June 30, 2026." Every figure above is read directly from its income statement, balance sheet, cash-flow statement, segment tables and business-highlights list; the two gross-property figures are derived by adding the balance sheet's net property and equipment to its stated accumulated depreciation, and that derivation is labelled as such in the text. Note on naming: this is Microsoft's fiscal fourth quarter of FY2026, not a calendar Q2 — the fiscal framing is used throughout to avoid the confusion.

Relationship to the Alphabet card, and one conflict to resolve. These cards make different, separately falsifiable claims from different primaries, and each should be read with the other: Alphabet's shows a quarter in which capex exceeded operating cash flow; Microsoft's shows a year in which it did not, with an 84% increase in contracted backlog behind it. The synthesis, and the only sector claim this card makes, is narrow: the financing stress is real but uneven, and the universal fact is the depreciation, not the borrowing. Flagged for the founder: the Alphabet card calls Alphabet "the most cash-generative advertising business ever built" and "the best-capitalised company in the industry." Neither superlative is grounded in a primary, and the second is in tension with this filing. Recommend striking both from that card rather than adding a competing superlative here.