Moody's Puts a Rating-Agency Number on the Committed-vs-Started Gap: $969B Committed to AI Infrastructure, $662B in Leases Not Yet Started
The deployment-gap debate this note tracks (the bear chain: Zitron's $50–100B warehoused-GPU estimate at Microsoft alone, ~$234B complex-wide, and the claim that less than half of $1.2T+ of hyperscaler capex has converted to operational capacity; the demand-side rebuttal: CoreWeave seeing "absolutely not any reduction in orders") got its first rating-agency canonical anchor this cycle.
The number (Passport to Wall Street, citing Moody's September estimate): "Moody's estimated in September that Alphabet, Amazon, Meta, Microsoft and Oracle have committed $969 billion to AI infrastructure, with $662 billion tied to data-center leases that have not yet started1. Those leases only appear as liabilities once the facilities go live."
That is the same committed-vs-operational distinction Zitron's audit argued from blogger data — now with a rating agency's byline and a bigger denominator: 68% of the five hyperscalers' $969B commitment sits in leases that have not begun, and the accounting treatment (off-balance-sheet until go-live) is exactly the structure the financing layer's SPV/lease machinery (Financing Layer: The "Darwin Moment" Arrives — CoreWeave Bonds Below 75 Cents, Oracle IG Paper at 88.5, Gartner's $4T Writedown Warning) is built to keep off the leverage ratios. S&P Global's July warning, quoted in the same piece: "market participants are growing leery of quickly rising leverage" among issuers once valued for steady cash flow.
The counterweight in the same source: Morgan Stanley Q1 data put hyperscaler total leverage at 1.3x earnings and net leverage at 0.5x — against 2.4x for the non-financial investment-grade universe — and JPMAM strategist Jorn Veeneman estimates the group "could add about $1.5 trillion of debt before their lease-adjusted leverage reaches the average for U.S. investment-grade industrial companies." Oracle is the outlier that proves the rule: 71.6% debt-to-capital ($169.1B debt vs $67.2B equity), $43B raised in FY2026 and ~$40B more planned for FY2027 — and it is Oracle whose bonds (51 issues above 6% yield) and CDS (record 261bps) are taking the repricing.
Read against the brief: the capex story's demand pipeline is still being committed at record pace, but two-thirds of the committed base is not yet delivering compute — which is why the tripwires that matter are deployment disclosures, not order announcements (Supply, Not Demand, Is the Binding Constraint: Huang Says Nvidia Can Fill Only ~70% of Demand as Taiwan Plans $4B Texas Supplier Parks). The Moody's $662B is now the number to track against the first GPU write-down or uninstalled-GPU disclosure watch: if go-lives slip, the gap converts from an accounting subtlety into a credit event.
-
An instance of AI capex now buys silicon faster than the physical world can plug it in. — The rating-agency number quantifies capex outrunning deployment — two-thirds of commitments sit in leases invisible on balance sheets until go-live, so installation lag rather than money sets what becomes operating compute. ↩︎