Goldman finds $3.2 trillion of AI commitments outside Big Tech balance sheets
Five hyperscalers have signed large lease and purchase obligations for AI infrastructure. Goldman's estimate points to future cash needs, not a new debt total.
Alphabet, Amazon, Meta, Microsoft and Oracle have roughly $3.2 trillion in AI-related contractual commitments that are not currently presented as ordinary balance-sheet debt, according to Goldman Sachs credit analyst Amanda Lynam. The estimate captures future cash obligations across leases, computing capacity, equipment, power and construction. It raises questions about how the companies will finance those commitments and how much flexibility they retain if AI demand falls short of expectations.
The distinction matters. The $3.2 trillion is a stock of contractual commitments extending across multiple years. It is not new borrowing, a single-year spending forecast, the present value of debt or evidence of immediate financial distress. Some of the spending is already reflected in analysts' capital expenditure forecasts. The financing risk lies in when payments become due, how they are funded and whether the resulting infrastructure earns an adequate return.
What Goldman's $3.2 trillion calculation measures
Goldman's estimate combines two disclosure groups. The five hyperscalers reported approximately $1.7 trillion of lease commitments covering data centers, computing infrastructure and other facilities. Roughly 75% of those commitments, about $1.3 trillion, relates to leases that have not yet commenced.
Those uncommenced leases generally do not appear as lease liabilities on current balance sheets, even when companies have signed binding agreements. Once a lease begins, the company typically recognizes a corresponding asset and liability.
The remaining approximately $400 billion of the $1.7 trillion lease total relates to leases that have already commenced. That amount should not be added again when calculating the obligations outside current lease liabilities.
Goldman identified another $1.9 trillion in purchase commitments. This category can include third-party cloud computing capacity, servers, networking equipment, power supply arrangements and data-center construction expenditures.
Combining the approximately $1.3 trillion of uncommenced leases with $1.9 trillion of other purchase commitments produces the $3.2 trillion estimate.
Approximate amounts in billions of dollars. The first bar is the uncommenced portion of reported lease commitments; the second covers other purchase commitments.
View data
| Commitment type | Approximate amount, billions of dollars |
|---|---|
| Uncommenced lease commitments | 1300 |
| Other purchase commitments | 1900 |
The estimate does not include all possible AI financing exposures. Goldman separately identified contingent commitments, including backstop obligations that become payable only when specified events occur.
The calculation also does not establish how much of the commitments represents spending beyond existing capital expenditure forecasts. The two disclosure categories measure contractual exposure, not an incremental financing requirement.
Cash timing matters more than the headline
The commitments have different start dates, counterparties, payment schedules and cancellation rights. Goldman said approximately $685 billion of the $1.9 trillion in other purchase commitments was due by the end of 2027.
That near-term component is particularly relevant to credit investors. Companies may need to pay equipment suppliers, begin lease payments or fund construction before the infrastructure produces sufficient customer revenue.
Lynam wrote that "a portion of these off-balance-sheet liabilities are captured in consensus capex forecasts," while emphasizing that the commitments still represent an additional use of cash.
She also argued that debt financing across syndicated and private markets would play a significant role in funding AI infrastructure.
The financing challenge is not necessarily whether hyperscalers can raise the entire amount immediately. It is whether operating cash flow, debt issuance, equity financing and customer receipts can support payments as contracts become active.
A signed capacity commitment may reduce flexibility even before it becomes a recognized liability. If customers delay deployment or demand proves weaker than anticipated, the buyer may still face contractual payments.
Conversely, capacity that quickly attracts paying workloads can generate the cash needed to meet those obligations. The same contractual structure can therefore support profitable expansion or amplify the cost of a disappointing investment.
Alphabet and Oracle show the accounting boundaries
Corporate filings provide examples of why the reported commitments deserve attention, while also demonstrating why different accounting categories should not be treated as interchangeable.
Alphabet disclosed $811.0 billion of material purchase commitments and other contractual obligations as of June 30, 2026. Of that amount, $200.7 billion was classified as short-term.
Its second-quarter filing also reported $707.0 billion of expected future fixed or guaranteed commitments under particular long-term agreements.
Those agreements included arrangements to secure technical infrastructure, inventory and energy supply. Certain energy agreements contained minimum purchase requirements and substantive termination provisions.
The $707.0 billion figure describes a particular group of commitments within Alphabet's broader disclosures. It should not be added mechanically to the $811.0 billion total.
Alphabet separately reported existing backstops involving financial guarantees with maximum potential payments of $7.6 billion and credit derivatives with maximum potential payments of $43.8 billion. It also disclosed an estimated $24.1 billion of future backstops, subject to finalizing arrangements with infrastructure providers.
These exposures differ from unconditional purchase commitments because payment can depend on specified events.
Oracle presents another important distinction.
Its quarterly filing disclosed $288 billion of additional lease commitments as of August 31, 2026. Substantially all related to data centers, with leases generally expected to commence between the second quarter of fiscal 2027 and fiscal 2029.
Those agreements have terms of 15 to 19 years and were not reflected in Oracle's balance-sheet lease liabilities at the reporting date.
Oracle also reported $664 billion in remaining performance obligations.
That second figure represents contracted customer performance obligations associated with expected future revenue, not another borrowing liability. It provides information about potential revenue supporting Oracle's commitments but does not establish when cash will be collected or whether every related infrastructure investment will be profitable.
Neither company's individual disclosure maps directly onto Goldman's combined $3.2 trillion. Reporting dates, contractual classifications and accounting treatments differ.
How contractual obligations become financing risk
The financial exposure develops through a sequence of commitments, construction, capacity delivery and customer utilization.
A hyperscaler may sign a long-term facility lease, commit to equipment purchases or reserve third-party computing capacity before the associated infrastructure generates meaningful revenue.
As contracts become active, previously uncommenced leases can create recognized lease liabilities. Equipment purchases and construction obligations can require cash payments or capital investment.
If internal cash generation does not cover those requirements, companies may choose to issue debt, raise equity or adjust other spending.
Credit risk increases when contracted payments continue but expected revenue arrives late. Borrowers can then face higher financing costs, reduced liquidity or a need to refinance existing obligations while supporting unfinished projects.
Equity investors face a related but different exposure. Even if contractual payments remain manageable, disappointing utilization or returns can reduce free cash flow, constrain shareholder distributions and weaken the economic justification for the investment.
Goldman's wider credit research illustrates how financing exposure extends beyond the five hyperscalers.
In an August 5, 2026 discussion, Goldman estimated that nearly $500 billion of AI-related debt had been issued globally so far that year. Lynam said hyperscalers accounted for only 40% of that issuance.
The remaining borrowing involved the broader AI ecosystem. That distinction matters because financing arrangements can distribute exposure among infrastructure owners, suppliers and other borrowers rather than concentrating every funding obligation on hyperscaler balance sheets.
The issuance figures do not, by themselves, establish how much risk sits with banks, private-credit lenders or infrastructure investors. They show that analyzing only the five technology companies would leave out an important part of the financing structure.
Why off-balance-sheet does not mean hidden debt
The strongest argument against treating $3.2 trillion as hidden borrowing is that the commitments are disclosed, their accounting classifications differ and the headline aggregates obligations across many years.
An uncommenced lease is not identical to a drawn loan. A purchase commitment may be fulfilled through an ordinary operating or capital expenditure. A contingent backstop may never require payment.
Companies also have different contractual protections, termination rights and potential opportunities to renegotiate delivery schedules.
Some commitments may already be incorporated into consensus capital expenditure estimates. Counting the entire $3.2 trillion as additional spending on top of those forecasts would risk double counting.
Nor does the estimate establish that the five companies face immediate difficulty accessing credit.
In Goldman's August credit-market discussion, Lynam said the firm was "not concerned about access to capital" for the AI financing theme. She pointed to the strong credit quality and relatively low leverage of major technology borrowers.
The question for lenders was increasingly where financing could be raised and at what price, rather than whether financing would necessarily become unavailable.
That view does not eliminate risk. It places the risk in the timing, cost and returns of financing rather than assuming that large contractual commitments automatically imply distress.
Who ultimately carries the exposure
For Alphabet, Amazon, Meta, Microsoft and Oracle, contractual commitments can secure future computing capacity while reducing flexibility to cut spending if market conditions change.
The direct financial effects depend on each company's operating cash flow, existing debt, payment schedules and ability to turn infrastructure into customer revenue.
For equipment and networking suppliers, long-term purchase agreements can improve demand visibility. They may also increase dependence on a relatively concentrated group of customers whose investment plans determine the timing of major orders.
For data-center owners, a long lease with a highly rated tenant can support construction financing and debt repayment assumptions.
But a facility that is delayed or lacks a firm customer commitment may leave its owner or financing partners exposed to construction costs before reliable rental income begins.
Banks and private lenders face counterparty, completion and refinancing risks. Those risks depend on the structure of individual transactions, not simply the aggregate value of hyperscaler commitments.
Shareholders face a further distinction: infrastructure spending can remain financeable while delivering disappointing returns on invested capital.
The balance of these risks will depend on contract terms and the ultimate distribution of losses if projects are delayed, reduced or canceled. Goldman's headline estimate does not identify which party would absorb every potential shortfall.
Three possible paths for the commitments
The most likely path is continued investment, with more obligations becoming recognized lease liabilities, capital spending and scheduled cash payments.
The commitments were made in anticipation of future computing demand. If capacity utilization increases and customers pay for the resulting services, companies may be able to support much of the spending through operating cash flow, supplemented by borrowing where appropriate.
The second possibility is selective renegotiation.
Hyperscalers could seek to delay lease commencements, resize facilities, adjust procurement schedules or shift capacity toward customers with stronger demand.
Such changes could reduce immediate cash requirements. They could also indicate that some original infrastructure demand assumptions were too aggressive.
The risk case is a more substantial financing squeeze.
If capacity utilization disappoints while credit spreads widen, companies could face higher borrowing costs precisely when contractual payments are increasing.
That combination could crowd out other investment, constrain financial flexibility or place pressure on credit ratings.
The effects would not necessarily be uniform. Financing structures, counterparty quality and existing liquidity would influence which companies and infrastructure owners face the greatest pressure.
What the 2027 filings will reveal
CTOL expects quarterly disclosures during 2027 to provide the clearest test of whether the $3.2 trillion estimate primarily represents manageable long-term financing exposure or an emerging credit problem.
Three developments will matter most: how much of the uncommenced lease portfolio becomes active, how much additional debt companies raise to support contractual spending, and whether customer revenue develops alongside the commissioned capacity.
Investors should also examine changes in purchase commitments, termination provisions, contingent guarantees and the timing of cash payments.
If the commitments increasingly become productive assets supported by operating cash flow, the headline will look more like a measure of long-term financing scale than evidence of financial strain.
If contractual payments accelerate while revenue generation and financing flexibility deteriorate, the same disclosures could signal growing pressure on creditors and shareholders.
The $3.2 trillion figure is significant because infrastructure is being contracted years before its economic returns are fully established.
Its ultimate importance will be determined not by the headline alone, but by which obligations become payable, who finances them and how quickly the capacity produces cash.