$662 billion in AI data-center leases isn't on any balance sheet yet

The capex number is the one everyone quotes

Every earnings season, the same number gets repeated across the AI-bubble debate: combined 2026 capital-expenditure guidance from Alphabet, Amazon, Microsoft and Meta runs to roughly $700 billion, by analyst tallies compiled across multiple research notes this year. It is a genuinely large number. It is also, by a wide margin, the smallest of the real figures in play, because it leaves out the hundreds of billions of dollars in AI data-center leases that haven't hit a balance sheet yet.

Capex guidance is what a company tells investors it plans to spend, on its own balance sheet, in the year the spending happens. It is the easiest number to find, because it sits on a single slide of every earnings deck. It is also the number that two specific accounting choices were built to keep looking smaller than the underlying commitment actually is.

Two-thirds of a trillion dollars in data-center leases that haven't started yet

In February 2026, Moody's Ratings published a figure that drew far less attention than any single earnings call: the five largest US hyperscalers, Amazon, Meta, Alphabet, Microsoft and Oracle, had amassed $969 billion in total undiscounted future data-center lease commitments by the end of 2025. Of that, $662 billion was for leases that had not yet commenced.

That distinction matters because of how lease accounting works. Under GAAP, a company only puts a lease on its balance sheet once the lease term actually begins, meaning once the building is delivered and the company takes possession. A signed commitment to lease a data center still under construction is real, in the sense that the company is contractually on the hook for it, but it doesn't have to appear as a liability yet. Moody's analysts David Gonzales and Alastair Drake noted that the $662 billion figure equals roughly 113% of the five companies' combined adjusted debt: more obligation sitting outside the balance sheet than inside it.

None of this is unusual on its own. Companies have always signed multi-year leases before occupying a building, and the accounting standard has always treated not-yet-commenced leases this way, for every industry, not just hyperscalers. What's new is the scale, and what the leases are actually for: a category of asset whose economic value depends entirely on a hardware cycle measured in months, not decades.

Inside one deal: Meta, Blue Owl, and a campus in Louisiana

The clearest single example of how this works is Meta's Hyperion data-center campus, a 2,250-acre site in rural Louisiana. In October 2025, Meta arranged what reporting at the time called the largest private-credit data-center financing on record: roughly $27 billion in debt plus $2.5 billion in equity, structured through a special purpose vehicle reportedly named Beignet.

Blue Owl Capital owns 80% of that vehicle. Meta owns the remaining 20%. Because Meta isn't deemed the SPV's primary beneficiary under consolidation accounting rules, Hyperion's debt doesn't appear on Meta's balance sheet. Meta instead leases the capacity it uses from the vehicle on a long-term operating basis, with those leases reportedly broken into four-year chunks specifically so rating agencies wouldn't treat them as debt.

Meta didn't walk away from the downside, though. It gave the joint venture a residual-value guarantee covering the first 16 years of operation: if a lease isn't renewed or is terminated early, Meta has committed to a cash payment covering the shortfall. The risk is real, and it still sits with Meta. What changed is when it shows up in Meta's reported numbers, not whether it exists. Reporting on the deal noted that Meta's own auditors had raised questions about keeping a project this size off the balance sheet, which is a reasonable signal that the structure sits closer to the edge of what's defensible than to the center of it.

PIMCO and BlackRock anchored the debt portion at $18 billion and $3 billion respectively, arranged by Morgan Stanley. Blue Owl has since pursued similar structures with other hyperscalers, and reporting through 2026 points to Meta preparing a second SPV, around $13 billion, for a Texas data center. Once one bank and one private-credit fund prove a template works, it replicates fast.

The other lever: how long does a GPU actually last

The lease structure controls when spending shows up. A second, quieter accounting choice controls how big it looks once it does: the assumed useful life of the GPUs themselves.

Investor Michael Burry, writing through his fund Cassandra Unchained, has argued publicly that hyperscalers depreciate Nvidia-based data-center hardware over five to six years, while Nvidia's actual product cadence, Hopper in 2022, Blackwell in 2024, Rubin slated for 2026, gives the newest chips a competitive life closer to two to three years before a faster, more efficient generation makes them uneconomic for frontier model training. Burry's estimate: roughly $176 billion in understated depreciation, and correspondingly overstated profit, across the industry between 2026 and 2028.

The mechanism is the same shape as the lease structure, just applied to an expense line instead of a balance-sheet entry. Spread a fixed cost over more years and the quarterly depreciation charge shrinks, which makes reported profit look larger than the hardware's real economic decay would suggest.

The hyperscalers' counter-argument isn't unreasonable on its face. A GPU that's no longer competitive for training frontier models doesn't become worthless. It gets reassigned to inference, smaller-model serving, or rendering workloads, where it keeps generating revenue years past its frontier-relevant lifespan. Companies have defended their depreciation schedules to auditors using utilization data and hardware-failure analysis, and those schedules have, so far, kept passing review. Whether a six-year schedule reflects genuine cascading utility or an assumption that won't survive the next hardware cycle isn't something anyone outside the finance team has the data to settle. What's certain is that the assumption is doing a lot of work either way.

Same shape, different lever

Put the two mechanisms next to each other and the pattern is hard to miss. Both take a cost that would otherwise hit the current quarter hard, and either move it off the statement entirely or spread it thin enough that no single quarter looks alarming.

Oracle runs a version of the third row at meaningful scale. Roughly $75 billion of its AI infrastructure backlog is now structured as prepaid or customer-supplied-hardware deals, in which the customer either pays upfront for capacity or brings its own chips. That doesn't just move debt off Oracle's balance sheet, it moves the entire purchasing decision, and the depreciation question, onto whoever signed the contract.

What changes if AI demand growth slows

None of these structures are a problem on their own, as long as demand for the underlying compute keeps growing at something close to the rate these commitments assume. The risk shows up if it doesn't.

Moody's own framing is the more useful one here: more than half a trillion dollars of data-center assets are going to start appearing on these companies' balance sheets over the next few years, simply because the leases behind them are scheduled to commence. That's true regardless of what AI demand does between now and then. If utilization comes in lower than assumed, the same leases that look like operating expenses today convert into fixed obligations against revenue that didn't show up, at roughly the same moment that write-downs on under-depreciated hardware would also be landing.

The Bank for International Settlements flagged the systemic version of this risk in its first-quarter 2026 review: these SPV structures deepen the links between hyperscalers and non-bank private-credit lenders, while banks carry their own exposure through funding lines extended to the same vehicles. A slowdown wouldn't just hit hyperscaler earnings. It would test refinancing capacity across a chain of lenders that built their own balance sheets assuming AI compute demand keeps compounding.

What to watch instead of the capex headline

The quoted capex figure is real, but it's the easiest number for a reason: it's already been shrunk by every choice described above. A more honest read of where the AI buildout actually stands means tracking a handful of less convenient numbers instead.

• Not-yet-commenced lease commitments, disclosed in 10-K footnotes, not the capex guidance slide.
• Any change in stated useful-life assumptions for servers and networking equipment, which usually shows up as a one-line footnote, not a headline.
• Credit-default-swap spreads on hyperscaler bonds and on the private-credit vehicles behind their SPVs, which move faster than earnings calls do.
• Impairment or write-down language on earnings calls, the first place a depreciation assumption that didn’t hold actually becomes visible.

The AI infrastructure being built right now might turn out to be exactly as valuable as its backers expect. But the number argued over every earnings season was never built to answer that question. It was built to make this quarter's filing look as small as the accounting rules allow.