Goldman Sachs puts Big Tech’s off-balance-sheet AI commitments at $3.2 trillion

Alphabet, Amazon, Meta, Microsoft and Oracle have roughly $3.2 trillion in commitments outside their balance sheets tied to the AI buildout, Goldman Sachs estimates. The finding, reported by Yahoo Finance on November 4, puts future payments alongside today’s infrastructure spending—and underpins the bank’s view that borrowing will remain a key source of funding.

The precise total needs clarification. Yahoo Finance separately gives $1.7 trillion in lease commitments and $1.9 trillion in other purchase commitments. Those figures add to $3.6 trillion, not $3.2 trillion. Its lease total also includes about $400 billion for leases already underway. The article does not reconcile these categories with its stated off-balance-sheet aggregate, so the figures should not be treated as a clean, additive measure of unrecorded liabilities.

Signed leases, future balance-sheet entries

The lease commitments cover data centers, computing infrastructure and other facilities. Roughly $1.3 trillion of the $1.7 trillion total relates to leases that have not yet begun, according to Goldman Sachs credit analyst Amanda Lynam. That separates a substantial future payment obligation from facilities whose leases are already running.

The accounting distinction is timing. When a lease starts, a company typically records both a lease liability and a corresponding asset on its balance sheet. Before that point, future payments generally are not recognized as lease liabilities—even if the company has signed a binding agreement to pay. Off-balance-sheet therefore does not mean the commitment is unsigned or optional.

$685 billionPurchase commitments due by year-end 2027

Goldman Sachs identifies $685 billion due by year-end 2027 within the companies’ $1.9 trillion in purchase commitments beyond leases.

The contracts extend beyond buildings

The $1.9 trillion purchase category captures commitments beyond lease obligations. It can include several parts of the infrastructure supply chain, rather than a single type of equipment or service:

• Contracts to buy computing capacity from third-party cloud providers.
• Server and network investments, plus power supply arrangements for data centers.
• Other expenditures connected to data-center construction.

That purchase figure excludes what Goldman calls contingent commitments. These include backstop obligations, where a payment becomes necessary only if a specified event occurs. They are a different kind of exposure from a scheduled purchase payment: the obligation depends on a trigger, rather than simply the arrival of a payment date.

Funding access is not an investment return

Lynam’s November 3 client note cautions against treating every commitment as spending missing from existing forecasts. She believes some of these obligations are already captured in consensus capital-expenditure forecasts—the market’s estimates of spending on long-lived assets. Still, she describes the fast-growing category as an additional use of cash, supporting her expectation that debt financing across syndicated and private markets will play a key role.

For most firms in the AI ecosystem, Lynam sees strong credit ratings preserving access to borrowing even after multiple rating downgrades. That is an argument about the ability to raise money, not proof that the investments will earn adequate returns. She identifies the payoff as an unresolved issue while saying recent earnings commentary gives no indication that infrastructure investment is about to slow.