When Oracle signs a long-term contract to buy Nvidia chips, Nvidia records the sale as revenue. Oracle records nothing.
Not a liability, not a footnote you'd notice, not a line on the balance sheet the market prices off of. Under current accounting rules, a purchase commitment doesn't become a liability until the goods are delivered — so nearly a trillion dollars in already-contracted cash outflows exist, per Morgan Stanley, as disclosure rather than debt.
Todd Castagno of Morgan Stanley's global valuation, accounting and tax team puts the asymmetry plainly: vendors book the revenue, buyers book nothing. The asymmetry is deliberate, and it is enormous.
Vendors book the revenue, buyers book nothing. The asymmetry is deliberate, and it is enormous — nearly a trillion dollars of contracted outflows sitting as disclosure rather than debt.
That single mechanic reframes the entire capex debate. Everyone is quoting Goldman Sachs' projection that hyperscaler capital expenditure hits $1.1 to $1.4 trillion by 2027 and asking whether AI will earn it back. That is the wrong number and the wrong question.
Last week we walked through the systemic-risk thread and the Amazon-and-Google capex escalation. This week the picture sharpens, because Morgan Stanley has done the work of adding up what isn't on the books — and the answer is roughly $1.8 trillion.
Here is the arithmetic. Around $982 billion in purchase commitments. Over $822 billion in leases that have been signed but not yet commenced, so they aren't liabilities yet. Another $110 billion in unpaid capex sitting in accounts payable. Add it up: about $1.8 trillion of true commitment engineered off the balance sheet.
These aren't contingent obligations. They lock in future cash outflows as firmly as any bond — they simply don't look like debt.
There's a mirror image worth naming. Where the China story reads Alibaba's AI backlog as a promissory note denominated in RMB — a disclosed number investors can at least argue over — the US version is a promissory note denominated in dollars that never made it onto the balance sheet at all. This is not a returns story. It's a debt-plumbing story.
The disguise becomes visible the moment you re-lever the ratios. Morgan Stanley estimates that counting finance leases would push Microsoft's capex-to-sales ratio from 33%/50% to 44%/64% for fiscal 2026 and 2027.
Oracle is the extreme case. Its capex-to-sales ratio would go from 76%/115% to 101%/189% once signed-but-unstarted leases are counted. Read that again: capex exceeding sales, on spending already contracted. The company is effectively levered at twice its reported load.
Aggregate it and the sector's gross leverage doubled from 0.9x to 1.8x in just two quarters — roughly 0.3x per quarter. That now exceeds the leverage of the entire energy sector, an industry whose assets are pipelines and reserves rather than data centers that don't fully exist yet.
Follow the money down the chain and it gets stranger. Private credit giants Apollo and Blackstone are moving leverage into the supply chain through special-purpose vehicles — the $35 billion "chip-backed" SPV deal for Anthropic being the headline example.
The structure matters more than the number. Leverage that would look alarming on a hyperscaler's balance sheet gets relocated to an SPV, then into the private credit system, where it is harder to see and harder to trace. Circular, opaque, and by design outside the reporting perimeter.
Then there's the vendor financing, which is where the plumbing starts to hum. Days payable outstanding — how long these companies take to actually pay suppliers — has stretched dramatically. Oracle's DPO is up 370% year over year. Meta's up 73%. Microsoft's up 69%.
Translated: the vendors are financing the buildout. Suppliers ship the equipment and wait to get paid, effectively extending credit to their own customers. When your accounts payable becomes a funding source, you are borrowing from the supply chain without calling it borrowing.
Meanwhile the visible debt is piling up too. Amazon is borrowing another $25 billion to support data-center expansion, and it now spends more on capex than it generates in cash flow from operations — the shortfall has to come from investors.
That is the tell. Morgan Stanley projects Amazon and Meta free cash flow approaches zero or goes negative in 2026, which means incremental spending gets funded almost entirely by new debt. The companies most identified with AI's commercial promise are the least able to generate the cash to pay for it.
Zoom out to the bond market and the scale registers. Global AI-related bond issuance hit $236 billion by the end of May 2026 — a 357% jump year over year — with Morgan Stanley expecting the full year to clear $570 billion. April alone saw over $74 billion.
Five names — Amazon, Meta, Google, Microsoft, Oracle — now make up 4% of the entire investment-grade bond index. Project-finance structures for data-center construction account for 85% of high-yield supply and 40% of investment-grade supply. AI is no longer a tech-sector story; it is a fixed-income story.
And the bond market is already sniffing out the strain. Under supply pressure, spreads on these names have drifted from AA territory into A. Meta's credit spread now sits wider than the CDX IG benchmark — the broad investment-grade index.
That is a quiet but real signal. Bond investors, who get paid to worry about return of capital rather than return on capital, are starting to price Meta as riskier than the average investment-grade credit. They are not waiting for the demand debate to resolve.
Which is the whole point, and the reframe worth holding onto. The consensus treats this as a demand-risk story: what if AI monetization disappoints, what if enterprises pivot to cheaper models. That is a real risk, but it is a 2027-and-beyond risk.
The fragility described here is structural and pre-loaded. The $1.8 trillion is already committed. The leases are already signed. The payables are already stretched. The SPVs are already levered. None of that unwinds if demand shows up — the cash outflows are locked either way.
So the danger isn't that the AI bet fails to pay off. The danger is that the financing architecture is brittle before we ever find out whether it does — a plumbing problem, not a demand problem, sitting mostly in footnotes while the market debates the wrong trillion-dollar number.