Hyperscalers have piled up roughly $2.8 trillion of leases, purchase commitments and guarantees, and four of the five now spend about as much as, or more than, their operating cash flow on capital equipment. The useful question is what stands behind that gap: contracted revenue and a balance sheet that can carry it.
The balance-sheet risk in the AI buildout is a timing risk. It appears when a company’s fixed payments for buildings, power and chips arrive before the revenue meant to cover them. The key measure is how well each company matches cash calls to contracted revenue, and the five largest spenders differ sharply on it.
Moody’s research published this week puts the off-balance-sheet obligations of Amazon, Microsoft, Alphabet, Meta and Oracle at about $2.8 trillion, up from about $350 billion in 2023. The total includes more than $1 trillion of leases that have not yet started, about $1.57 trillion of purchase commitments and about $137 billion of guarantees. Morgan Stanley reached a similar $2.8 trillion across seven companies, including Nvidia and Broadcom.
That total combines instruments with very different degrees of rigidity. A bond creates a contractual schedule of principal and interest. A data-centre lease generally becomes a balance-sheet liability only when the facility is handed over. A chip purchase agreement commits cash in exchange for hardware. A power contract secures electricity. A residual-value guarantee may require payment only after a default or a shortfall in asset value. Adding them together discards the timing and trigger information that equity and credit investors need.
Three layers of obligation
The first layer is leases that have not yet started. Reuters reported in August that Amazon, Alphabet, Meta, Microsoft and Oracle had about $1.09 trillion of future lease payments attached to leases that had not commenced, much of it for data centres. As facilities come online, those commitments move onto reported balance sheets and start to affect depreciation, lease expense and cash flow.
The second layer is ordinary debt, which is rising fast. The five companies had issued about $220 billion of bonds by August 10, according to BNP Paribas data reported by Reuters, against about $108 billion in all of 2025. UBS expects $270 billion to $290 billion for the full year, and Goldman Sachs has estimated $400 billion in 2027.
The third layer moves financing further from the hyperscaler’s own balance sheet. AI campuses can be funded by infrastructure investors and lenders, backed by long leases or residual-value guarantees from technology companies. The Financial Times has reported roughly $300 billion of debt capacity associated with such guarantees across the buildout. Meta’s Hyperion campus in Louisiana is the best-known example. It is financed through a roughly $27 billion joint venture in which Blue Owl funds own 80%, backed by a capped Meta residual-value guarantee covering the first 16 years of operation. The real exposure depends on each guarantee’s trigger, cap and collateral, not its face value.
A non-cancellable lease or purchase commitment consumes cash as surely as debt service once its payment date arrives. Keeping it off the balance sheet does not make it optional.
The four-number test
Investors need four figures for each company: committed cash payments by year, debt-equivalent guarantees, lease liabilities as facilities commence, and customer revenue contracted against that capacity. The latest quarter shows why the test now applies to everyone. Four of the five spent about as much as, or more than, their operating cash flow on capital equipment.
| Company | Capex as share of operating cash flow, latest quarter | Contracted backlog |
|---|---|---|
| Microsoft | about 75% | $678 billion commercial RPO |
| Meta | about 98% | none (builds for its own products) |
| Alphabet | about 115% | $514 billion Google Cloud backlog |
| Amazon | about 119% | $496 billion AWS backlog |
| Oracle | about 123% | $664 billion RPO |
Spending more than operating cash flow is not a failure in itself. The question is what stands behind the gap: contracted revenue, and a balance sheet that can carry the shortfall if that revenue slips. On those two measures the five sort into four positions.
Microsoft is the strongest. It is the only one still funding its buildout comfortably from cash flow, and its backlog is the largest. It has an OpenAI concentration, since OpenAI was about 45% of commercial RPO at the end of 2025. But Microsoft says RPO excluding OpenAI grew 25% in the latest year, and if OpenAI stopped paying, Microsoft could carry the capacity from its own cash while it resold it.
Alphabet and Amazon pass on backlog, not on cash. Both now spend more than they generate, with free cash flow negative in the latest quarter. But each has about half a trillion dollars of contracted cloud revenue spread across thousands of customers, and each has one of the strongest balance sheets in corporate America. They can borrow for years against that backlog.
Meta cannot be scored on contracted revenue. Its capex nearly matches its operating cash flow, and it builds for its own advertising and assistant products, so no customer contract stands behind the spending. Its test is whether AI-driven gains in ad pricing and engagement keep growing faster than depreciation. Its use of joint-venture structures and guarantees also makes it the company whose obligations are hardest to see.
Oracle is the first test case. Its backlog is the largest relative to its size: remaining performance obligations reached $664 billion at the end of August, and OpenAI’s roughly $300 billion five-year contract is the largest single piece. But it has the highest capex-to-cash-flow ratio of the five, negative free cash flow of $23.7 billion in its last fiscal year and the lowest credit rating of the group. Its biggest customer is itself funding its commitments by raising external capital. Alphabet and Amazon also spend beyond their cash flow, but against diversified backlogs and strong balance sheets. Oracle is doing it against a concentrated backlog and a balance sheet that is already stretched. If AI demand disappoints, Oracle’s cash calls will outrun its revenue first.
Depreciation is the hidden variable
Matching also has an asset-life dimension. If GPUs lose economic value faster than the depreciation schedules assume, reported earnings will overstate the returns on the buildout. The resulting cash shortfall will show up only when the hardware has to be replaced. Companies that lease accelerators to others, and therefore bear re-leasing risk, are more exposed than those that use the chips for their own services.
Apollo estimates hyperscaler capex has nearly tripled since 2023. The step-up has pushed the buildout from retained earnings into capital markets. That is healthy when contracted revenue grows alongside it and dangerous when it does not.
The verdict
The AI capex cycle turns dangerous when fixed infrastructure cash calls outrun the cash earned by the workloads they serve. On today’s numbers, four of the five already spend beyond or near their operating cash flow, so the difference lies in what backs the gap. Microsoft funds its buildout itself. Alphabet and Amazon borrow against diversified backlogs. Meta bets its own cash on its own products. Oracle borrows against one customer’s ability to keep raising money. That is why Oracle’s December quarter, and how much of its backlog growth comes from outside OpenAI, is the first real test of the AI financing boom.