TL;DR
- Journal Estimate: The Wall Street Journal estimates that nine technology companies have roughly $3 trillion in future contracts tied largely to AI infrastructure.
- Reproducible Base: Finterm’s filing-based five-company reconstruction totals $1.651 trillion; public sources do not show the company figures behind the Journal’s broader estimate.
- Accounting Timing: Signed future leases and purchase contracts can be disclosed in filing footnotes before they become balance-sheet liabilities or current capital spending.
- Project Financing: Long contracts reduce spending flexibility and allow data-center developers to borrow against technology buyers’ promised payments.
- Assets and Demand: The contracts also secure productive assets and services, while cloud revenue and contracted demand can help cover the future costs.
A Wall Street Journal analysis estimates that nine technology companies have roughly $3 trillion in future leases and purchase commitments tied largely to AI infrastructure. The total spans years of promised payments rather than cash spent on capital projects in one reporting period. The contracts can lock the technology companies into payments for data centers, chips, power and computing capacity years into the future. Data-center developers can borrow to build facilities against those promised payments. If the developers later need new loans or the facilities lose value, project lenders may recover less than they expected.
The estimate covers Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, AMD and SpaceX. The reported estimate put about $1.2 trillion in leases that have been signed but have not started and about $1.9 trillion in purchase commitments. Rounded separately, those two buckets can display as $1.2 trillion and $1.9 trillion even when their unrounded combined value displays as $3.0 trillion. The cited filings were mostly current through June 2026, but public sources do not provide the company-by-company rows needed to produce a matched-date snapshot or test possible overlaps.
What the Estimate Counts
Capital expenditure measures money paid during a reporting period for property and equipment. It is a flow, and the acquired assets appear on the balance sheet. Cash capital spending at five hyperscalers has been rising much faster than operating cash flow through the first quarter of 2026, but even that spending did not include the entire pipeline of future contracts.
The new estimate looks farther ahead. An uncommenced lease is a signed agreement for a facility or other asset that is not yet available to the customer. A purchase or construction commitment requires future payment for goods or services such as servers, wafers, data-center construction, electricity or computing capacity. The amounts are generally nominal payments across the life of the contracts, not one year’s cash outlay and not their discounted present value.
Adding future leases and purchase contracts captures genuine fixed exposure, but it also combines different businesses and economic objects. At May 3, 2026, Broadcom had $128.11 billion of enforceable purchase commitments and $4.105 billion of other contractual commitments. Its purchases primarily cover inventory inputs rather than the data-center leases carried by a cloud provider. At June 27, 2026, AMD had $30.276 billion of unconditional commitments for wafers, substrates, cloud services, licenses and guarantees. These obligations cover broader supplies and services, not only AI-specific data centers.
Baseline for Five Big-Tech Leaders
A filing-based reconstruction published in July provides a narrower comparison whose arithmetic can be checked. It counted signed leases that had not commenced and disclosed purchase, construction and other commitments for Alphabet, Amazon, Meta, Microsoft and Oracle. It excluded lease liabilities already recognized on balance sheets, avoiding an obvious source of double counting.
Future commitments in a July 2026 reconstruction
The total is $1.650676 trillion: $821.427 billion of uncommenced leases plus $829.249 billion of purchase, construction and other commitments. It is reproducible within those definitions, but it is not a same-day snapshot. Microsoft’s row combines leases reported at March 31, 2026 with purchase and construction obligations from June 30, 2025. The five-company population also differs from the Journal’s nine-company set.
Disclosed commitments can change sharply as companies sign contracts or update their scope. By June 30, Alphabet alone disclosed roughly $902 billion across the same broad categories, up from the $408 billion March subtotal in the table. Substituting only that newer Alphabet figure raises the five-company arithmetic to about $2.145 trillion, but doing so leaves the other four companies frozen at older dates. It is a sensitivity check, not a current cohort total.
Disclosure Is Not the Same as Debt
The phrase off balance sheet describes recognition timing, not secrecy. These obligations are visible because companies disclose them in filing footnotes. Under US lease-accounting rules, a customer generally records a right-of-use asset and a discounted lease liability when the leased property becomes available and the lease begins. A signed lease for a data center still under construction can remain an uncommenced commitment until that point.
Purchase and construction agreements are different again. Before delivery, both sides still owe performance: the supplier must provide the chips, building, power or service, and the customer must pay under the contract’s terms. Accounting recognition follows the delivery and the nature of the transaction. A corporate bond, by contrast, is funded borrowing already received and recognized as debt.
Because leases, purchase contracts and bonds are recognized differently, a single gross figure can measure the contract pipeline but not corporate leverage. The total does not reveal present value, annual cash burden, cancellation rights, delivery risk or the share already reflected elsewhere in a company’s accounts. It also cannot be compared cleanly with one year’s capex, because the two measures cover different time spans and stages of the buildout.
Who Finances the Buildout
Long commitments matter even before accounting recognition because they can make enormous projects financeable. A developer or special-purpose vehicle can own and build a data center, while a technology company promises lease or capacity payments. Banks, private-credit funds, insurers and other investors supply debt or equity against those expected cash flows. The technology buyer receives future capacity without funding the entire project upfront.
The Bank for International Settlements describes this structure as a form of shadow borrowing when leases, offtake contracts or guarantees support debt raised outside the technology company’s consolidated balance sheet. That does not turn the vehicle’s borrowing into the buyer’s corporate debt. It shows how the buyer’s contract can carry credit risk through the financing chain.
Meta’s Hyperion development illustrates the roles. Meta holds a minority interest in a venture, made an initial lease commitment and provided residual-value guarantees. The venture and its lenders finance infrastructure that Meta expects to use. The lease payments and guarantees expose Meta to future conditions, while refinancing and the eventual value of the facilities remain material to the vehicle and its capital providers. Adding every contract, guarantee and project loan together without testing overlap would count the same economic chain more than once.
One Total, Different Risks
The five large cloud companies do not occupy identical positions. Meta is primarily buying capacity for its own services. Alphabet, Amazon, Microsoft and Oracle both buy infrastructure and sell cloud capacity to customers, so contracted cloud revenue can help cover some future costs. Oracle’s lease pipeline runs as long as 15 to 19 years and is large relative to its revenue and assets, which makes timing and customer concentration especially important. Alphabet’s much larger June commitment total shows how quickly the disclosed pipeline can expand.
Semiconductor suppliers introduce another mismatch. Broadcom and AMD commit to inventory and production inputs needed to serve future demand, while Nvidia can act as supplier, buyer and financing participant. SpaceX has a different business model again, and the public reports relaying the estimate do not identify its contribution to the Journal’s total. A nine-name sum can describe scale, but it cannot describe one common exposure.
The contracts secure productive assets and services for companies that generate substantial operating cash flow, while cloud backlogs provide evidence of customer demand. Moody’s measured $969 billion of future lease payments for the five hyperscalers at the end of 2025, including $662 billion for leases not yet commenced, and expected earnings growth to mitigate the effect for most of the group.
Backlog is not a dollar-for-dollar offset, however. Revenue may arrive on a different schedule, at uncertain margins and from concentrated counterparties; some contracts can be changed or cancelled. Each company’s exposure therefore depends on whether its assets, customer contracts and cash generation arrive soon enough to meet the fixed payments it has already promised.
