Big Tech's AI Debt Is Real. Most of It Just Hasn't Landed Yet

Buried in the announcement of the largest data-center financing ever completed is a clause that explains the AI boom's balance sheets better than any earnings call. When Meta unveiled its $27 billion Hyperion campus in Louisiana, a joint venture in which funds managed by Blue Owl Capital own 80 percent and Meta owns 20, the company disclosed that it would lease the finished facilities back for an initial term of just four years⁸. Hyperscale data-center leases have historically run a decade or longer, as Bisnow noted⁹ when the deal closed. To compensate investors for that flexibility, Meta attached a residual value guarantee lasting sixteen years: if Meta ever declines to renew and the campus sells below an agreed threshold, Meta covers the shortfall.

Keep that clause in mind while reading this week's financial news. Alphabet just sold $25 billion of investment-grade bonds¹ in ten tranches stretching to 2066, drawing roughly $115 billion in peak demand, the third-largest order book of the year behind Oracle's and Amazon's. The sale came weeks after Alphabet raised its 2026 capital spending forecast to as much as $205 billion² and reported negative free cash flow¹² for the quarter. The hyperscalers, the handful of giant cloud platforms comprising Amazon, Microsoft, Alphabet, Meta and Oracle, issued about $194 billion of bonds in the first half of 2026, up 79 percent from a year earlier³. And Moody's now counts $1.2 trillion in data-center lease commitments⁴ across the group, more than $820 billion of it tied to facilities still under construction. All of which raises the question of whether we are watching prudent financing of contracted demand or the early innings of a telecom-style capacity bust.

Having worked through the numbers, I think the answer is neither, and the second framing misses what is actually risky here. Alphabet is not WorldCom; a company with Aa2/AA+ ratings, a search monopoly and $115 billion of bond orders is in no meaningful danger of default. The real story is the architecture: a financing structure engineered, quite deliberately and quite legally, to push obligations into the future and disperse the downside onto other balance sheets. This is not a solvency problem. It is a recognition problem, and recognition problems have a habit of resolving all at once.

Start with the accounting. Under US GAAP, a lease that has been signed but has not commenced does not appear as a balance-sheet liability, no matter how binding it is. Moody's found that as of year-end 2025, the top five hyperscalers had amassed $969 billion in future lease commitments, of which $662 billion had not yet commenced, a sum equal to 113 percent of their most recent adjusted debt⁵. By July the totals had grown to $1.2 trillion and $820 billion. Moody's calls the effect a "reporting deferral"⁶ that keeps official disclosures from showing the full scale of long-term economic risk, especially because short leases are often backstopped by exactly the kind of guarantee Meta gave Blue Owl. None of this is hidden from a determined reader of footnotes. It is simply debt that arrives on a delay, timed to land when the facilities switch on, which is also precisely when the revenue those facilities were built to earn has to show up.

The case that this is all fine is genuinely strong, and the bond market makes it daily. Demand today is contracted rather than imagined: Google Cloud's backlog of signed-but-unrecognized revenue reached $514 billion in the second quarter¹¹, up from about $106 billion a year earlier¹². Microsoft closed its fiscal year with $678 billion in commercial remaining performance obligations, up 84 percent¹³. Every major platform says it cannot build capacity fast enough. The 1999 comparison flatters nobody who reads the history: Brookings' post-mortem found carriers poured more than $500 billion into networks while no more than 2 percent of long-distance capacity was actually being used¹⁷. Fiber was built for traffic that never came. Today's compute is sold before the concrete is poured. That distinction is real, and it is the main reason I do not expect a hyperscaler credit event.

But the backlog comfort has two cracks. The first is concentration. Microsoft's CFO told analysts that backlog grew 25 percent excluding OpenAI, which implies roughly a third of that $678 billion traces to a single customer¹⁴, a company that loses money and whose own funding depends on the same AI capital cycle it anchors. Contracted revenue is only as durable as the counterparty signing the contract. The second crack is that backlog measures demand rather than returns. It bundles multi-year cloud, software and marketplace commitments, and says nothing about whether the marginal data center earns its keep after depreciation, power, networking and the two-to-three-year lag Moody's sees between capital spending and AI revenue⁷. Alphabet's negative free cash flow quarter, its first since 2004, is what that lag looks like at the strongest company in the cohort.

You can already watch what happens when this risk cannot shelter behind a fortress parent. CoreWeave, the leveraged GPU cloud provider, saw its credit default swaps blow past 855 basis points in late July, implying about a 50 percent five-year default probability¹⁵. Days later, a loan funding its GPU purchases cleared at a 10.44 percent yield, with amortization and covenant protections lenders had not demanded in years¹⁶. The optimists read this as discipline working, weak credits priced differently from strong ones, and that reading is fair as far as it goes. It just ignores that the same private-credit complex, Blue Owl, PIMCO and their peers, sits on both sides of the market, underwriting hyperscaler lease vehicles and neocloud GPU loans against the same correlated assumption about AI utilization. Capital cycles crack at the periphery first. They rarely stop there when the core assumption is shared.

Which brings me back to the Hyperion lease, because it is the most honest disclosure in this entire boom. A four-year initial term on a campus designed for decades of AI work only makes sense if the tenant wants an exit. S&P's Naveen Sarma put it plainly to PitchBook¹⁰: "Meta wants the optionality to be able to walk away from this facility" if strategy changes or the bubble cracks. The structure is a priced confession that nobody, including the people spending the trillion dollars, knows what AI's unit economics will be in 2030. And the sixteen-year guarantee means walking away is not free; the risk boomerangs back to Meta, capped but real.

Meanwhile the downside of the underlying bet is already being priced somewhere: labor markets. Concentrix, a cornerstone employer of the Philippine outsourcing industry, cut its revenue outlook in June and sold off alongside Teleperformance¹⁸ as Bloomberg Intelligence concluded that "AI is shrinking demand for its core customer-experience outsourcing business faster than higher-value AI services are expanding." Note the asymmetry. Substitution is showing up in BPO revenue before durable AI profits have shown up in anyone's cloud income statement, and there is no guarantee the hyperscalers capture enough of those labor savings to service the infrastructure, since enterprise customers may keep most of them and token prices keep getting competed down.

The test of all this has a schedule anyone can read. Moody's projects hyperscaler capital spending will approach $1 trillion in 2027, and the $820 billion of not-yet-commenced leases will migrate onto balance sheets as buildings power up through 2027 and 2028. If cloud margins and backlog conversion hold through that migration, the architects of these structures will have financed the largest private build-out in history without breaking anything, and this column will read as excessive worry. If they slip, the losses will surface exactly where the architecture directed them: at the lessors, the private-credit funds, the chip suppliers and the call-center floors that absorbed the risk Big Tech's pristine bond spreads never had to carry.