CCIR Compute Credit Index Research
Research · 2026-07-22 · filings

Breaking Down Big Tech's $1.65 Trillion "Hidden Debt"

A widely covered Nikkei study put the off-balance-sheet AI commitments of Alphabet, Amazon, Meta, Microsoft, and Oracle at $1.65 trillion — 122% of their reported debt — under the headline of "hidden debts." This note rebuilds that number from the five companies' own SEC filings and takes it apart: what kind of obligation each dollar is, what the commitment language actually binds, what sits on the other side of the same contracts, and where the number differs from debt. Every figure below carries its filing and accession number.

01 The number reconciles

Summing the latest available filings — first-quarter 2026 10-Qs for Meta, Alphabet, and Amazon; Microsoft's 10-Q for the quarter ended March 31, 2026; Oracle's fiscal-2026 10-K — the two footnote categories the study counted come to approximately $1.67 trillion: about $821 billion of leases signed but not yet commenced, and about $848 billion of purchase and other contractual commitments. In plain terms: the first bucket is rent — signed leases on data centers still under construction, with payments that begin when the buildings are delivered. The second is bills — contracts to buy chips, power, and cloud capacity, counted at the minimums the companies cannot walk away from. The aggregate is real, and none of it is hidden in any ordinary sense — every figure in this note comes from the commitments and leases notes of documents filed with the SEC, where the accounting standards require them to be disclosed. The accurate description is disclosed but unrecognized: real obligations that do not appear in the debt line the market screens.

The $1.65T, decomposed · latest filing per company · $B
Meta $421B Alphabet $408B Microsoft $339B Oracle $273B Amazon $229B leases not yet commenced purchase & other commitments
Microsoft purchase and construction commitments are as of June 30, 2025 (its fiscal 2025 10-K; the 10-Q does not update the table) — the one older window in the set. Oracle's purchase figure excludes $19B of cloud-infrastructure purchase commitments signed after its May 31 year-end, disclosed as a subsequent event.

02 Why it is not debt — and where it behaves like debt

Debt is money already received, carrying an unconditional obligation to repay, standing in leverage ratios and cross-default clauses, and surviving bankruptcy as a creditor claim. Nearly all of the $1.65 trillion is instead executory: promises to pay for facilities and goods not yet delivered. The distinction has consequences in both directions. An executory commitment generally cannot be accelerated, has no principal, and — for leases — can be rejected in bankruptcy at a capped damages claim. But a commitment with take-or-pay terms binds whether or not the capacity is used, which is precisely the property that lets other parties borrow against it. The question the aggregate cannot answer, and the filings can, is how much of the total carries that property.

Roughly half the total — $821 billion — is leases signed but not yet commenced, and this half is not permanently off-balance-sheet at all. Under the lease accounting standard, the liability is recognized when the landlord delivers the facility; the footnote is a forward calendar of lease liabilities that will surface as data centers energize. Alphabet describes its pipeline as carrying "non-cancelable lease terms primarily between one and 25 years"; Oracle's commitments run fifteen to nineteen years; Meta's reach thirty. The growth in this category is the construction pipeline made visible:

Leases signed, not yet commenced · prior disclosure → latest · $B (windows labeled per row)
Oracle $4.1B · May-22 $260B · May-26 Microsoft $92.7B · Jun-25 $197B · Mar-26 Meta $103.8B · Dec-25 $183B · Mar-26 Amazon $96.4B · Dec-25 $106B · Mar-26 Alphabet $58.5B · Dec-25 $76B · Mar-26
Windows differ by row and are labeled on each: Oracle discloses annually (fiscal 2022 → fiscal 2026, a ~63× increase); Microsoft's move is nine months; Meta, Amazon, and Alphabet are one quarter. Meta added $79B to this category in the first quarter of 2026 alone.

03 The commitment language, ranked

The purchase-commitment half is where conditionality lives, and the five companies do not write the same contract. Ranked from most binding to least, in the issuers' own words:

StrengthLanguage (verbatim)Where
Cash already fenced "we reclassified $5.00 billion of money market funds as restricted cash equivalents … restricted from general corporate use and … expected to be released upon satisfying the underlying purchase obligations" Meta 10-Q
Non-cancelable, blanket "$237.67 billion of non-cancelable contractual commitments … mostly related to third-party cloud capacity arrangements" Meta 10-Q
Take-or-pay "open purchase orders and take-or-pay contracts" (purchase commitments, "primarily … datacenters"); "take-or-pay provisions for minimum quantities of energy supply and substantive termination fees" Microsoft 10-K; Alphabet 10-Q
Unconditional "unconditional purchase obligations … enforceable and legally binding and specify … fixed or minimum quantities to be purchased" Oracle 10-K; Amazon uses the same "unconditional" table heading
Minimum-only counting "For agreements with variable terms, we do not estimate the non-cancelable obligation beyond any minimum quantities and/or pricing" Alphabet 10-Q; Amazon has parallel language
Excluded by the issuer "agreements … to purchase clean and renewable energy that do not specify a fixed or minimum volume commitment" are outside the totals Meta 10-K; Alphabet and Amazon exclude the same class

Two readings follow. The alarmed reading — that the totals hide soft, walk-away spending — is backwards: the issuers already filter their totals to the non-cancelable portion or the minimum commitment, and exclude volume-based energy agreements entirely. The disclosed figures sit closer to a floor on true commitment than an exaggeration. The complacent reading — that executory means escapable — fails on the same language: take-or-pay and termination-fee terms are the reason these promises can serve as collateral for other people's borrowing, which is where this note is ultimately headed.

04 The instruments at the edge: backstops, RVGs, guarantees

The smallest category by notional is the most interesting by structure: credit support written around other companies' financings. The filings now show two different accounting treatments of the same economic act. Alphabet accounts for its data-center payment backstops as credit derivatives, carried at fair value, with disclosed notionals: zero at the end of 2024, $16.9 billion at the end of 2025, $28.4 billion at March 31, 2026, approximately $15.3 billion more in agreements entered in April, and a framework "to provide up to $33.3 billion of future backstops … subject to finalization of terms with data center providers." The agreements run up to fifteen years; on a default, Alphabet "retain[s] the right to assume the underlying leases for internal use or to sublease to third parties." No counterparty is named in the filings.

Alphabet data-center backstops, credit-derivative notional · $B
$0.0B Dec-24 $16.9B Dec-25 $28.4B Mar-26 $43.7B Apr-26 quarter-end notional (10-K/10-Q) incl. April agreements (subsequent event)
Alphabet 10-K FY2025 and 10-Q Q1 2026. The April figure adds the ~$15.3B of backstop agreements disclosed as a subsequent event to the March 31 notional; the further "up to $33.3 billion" framework is not charted because terms were not final. Separately, Alphabet reports financial guarantees supporting energy-infrastructure procurement of $9.0B maximum potential payments.

The same function appears in two other forms: Meta's Louisiana data-center venture — unconsolidated, 20%-owned — carries residual value guarantees with an aggregate threshold of approximately $28 billion at zero recorded liability (disclosed maximum exposure to loss: $46.0 billion), and Oracle's lease pipeline includes a facility for which it "guaranteed up to $3.3 billion of the lessor's borrowing." Same economic act, three treatments — a fair-valued derivative, an RVG at zero, a guarantee inside a lease footnote. These notionals are small next to the $1.65 trillion; they matter because they are the purest form of the pattern the next section documents at scale.

05 The other side — and who owes it

A purchase commitment is a two-sided object, and four of the five companies also sell compute at scale. Their disclosed revenue backlogs — remaining performance obligations, the mirror image of somebody's commitment footnote — now total roughly $2.1 trillion: Oracle $638 billion (from $138 billion a year earlier), Microsoft $633 billion, Alphabet $467.6 billion, Amazon $364 billion. In aggregate, the five companies' contracted future revenues exceed their contracted future spending — the arithmetic that makes the commitments look covered:

Commitments vs revenue backlog (RPO) · latest filings · $B
commitments revenue backlog (RPO) Meta $421B no compute revenue backlog — coverage runs through advertising cash flow Alphabet $408B $468B Microsoft $339B $633B Oracle $273B $638B Amazon $229B $364B
RPO definitions differ by issuer (Alphabet's excludes cancellable contracts; Amazon's covers contracts over one year, weighted-average life 5.5 years; Microsoft's commercial RPO has a weighted-average duration of ~2.5 years) — comparable in kind, not in duration.

The comfort in that arithmetic depends entirely on who owes the $2.1 trillion — and here the two sides of the ledger are not symmetric. The commitments are owed by five of the largest, most creditworthy balance sheets in existence, in language they themselves describe as non-cancelable, unconditional, or take-or-pay. The backlog is owed to them, and the filings say less about by whom — but what they do say points the same direction. Amazon is the only one that names a counterparty: its backlog growth is driven by an OpenAI arrangement expanded by $100 billion to $138 billion — while Amazon separately reports a $15 billion investment in OpenAI, a $35 billion equity commitment letter, and a financing facility of up to $20 billion for Anthropic that unlocks "as we reach certain delivery milestones of compute capacity." Oracle attributes its $500 billion backlog jump to "certain significant cloud contracts," unnamed, expects only 12% of the total to convert to revenue within twelve months, and states that returns on its infrastructure "are dependent on customer demand and the ability of our key customers to meet their contractual obligations." Microsoft's commercial backlog grew 99% in the year it recapitalized its OpenAI relationship — an equity-method stake with continuing revenue-sharing — and Alphabet's includes "a limited number of agreements" to supply multiple gigawatts of TPU hardware, around which Alphabet itself provides the credit backstops of the previous section.

One disclosure mechanic keeps this concentration off the standard screens: customer-concentration tests run on recognized revenue. Oracle can state, accurately, that no customer accounted for 10% or more of its total revenues in fiscal 2026 — while attributing a $500 billion backlog increase to a small number of contracts that have not begun converting. Backlog concentration only surfaces in the concentration disclosures as the contracts turn into revenue. So the honest restatement of the "hidden debt" number is this: the obligations are disclosed, unconditional, and owed by the strongest credits in the market; the receivables covering them are concentrated in a limited number of counterparties that are private, unrated, and — in the cases the filings themselves disclose — financed in part by the sellers. What the aggregate measures is not concealment but credit transformation, and its scale is set by the weakest link in that chain, not the strongest.

Commitments as a multiple of operating cash flow · trailing twelve months
Oracle 8.5× Meta 3.4× Alphabet 2.3× Microsoft 2.0× Amazon 1.5× operating cash flow: Oracle FY26 $32.0B · Meta TTM $124.0B · Alphabet TTM $174.4B · Microsoft TTM $170.2B · Amazon TTM $148.5B
A duration-free coverage sketch, not a solvency measure: the commitments pay out over as much as thirty years while the cash flow is one year's. Its use is the cross-section — the two ends of the range are five times apart.

The cross-section, then, is the honest summary. Meta is the one pure buyer: the strongest commitment language of the five and no contracted revenue against it — its coverage is advertising cash flow, and its exposure is a return-on-investment question rather than a counterparty one. Oracle is the concentrated transformation trade: the largest lease pipeline, the longest terms, and the highest commitment-to-cash-flow multiple, covered by the largest and least-seasoned backlog. Microsoft, Alphabet, and Amazon hold commitments matched or exceeded by contracted revenue plus the largest operating cash flows in corporate history — with their exposure concentrated in how much of the backlog traces to the same few counterparties. One aggregate number describes five different balance sheets. These same commitments are the collateral of the compute-credit market — the facilities in the compute credit tracker lend against exactly these promises, and price their quality — which is why this record keeps them in view. This note is dated; the quantities it decomposes are maintained, filing by filing, on the AI commitments monitor.

Sources — all figures from SEC filings, verified against the documents: Meta Platforms 10-Q Q1 2026 (acc. 0001628280-26-028526) and 10-K FY2025 (0001628280-26-003942); Alphabet 10-Q Q1 2026 (0001652044-26-000048) and 10-K FY2025 (0001652044-26-000018); Amazon 10-Q Q1 2026 (0001018724-26-000014) and 10-K FY2025 (0001018724-26-000004); Microsoft 10-Q for the quarter ended March 31, 2026 (0001193125-26-191507) and 10-K FY2025 (0000950170-25-100235); Oracle 10-K FY2026 (0001193125-26-277521) and 10-K FY2022 (0001564590-22-023675). The $1.65 trillion aggregate and "hidden debts" characterization: Nikkei Asia, "Five US tech giants' hidden debts soar to $1.65tn on opaque AI funding" (July 2026). Quotations are verbatim from the filings. Nothing here is a solvency opinion, a forecast, or investment advice; CCIR publishes reference data and records.