$1.65T Hidden AI Debt at Big Tech: What Nikkei Found

The most important number in Big Tech’s AI buildout may be the one that doesn’t appear in a standard debt-to-equity ratio. A new Nikkei analysis estimates that five hyperscalers — Alphabet, Microsoft, Amazon, Meta, and Oracle — now carry roughly $1.65 trillion in off-balance-sheet debt tied to AI, a figure that has grown about eightfold since 2022 and now exceeds the same companies’ combined on-balance-sheet debt of roughly $1.35 trillion.

Put plainly: the liabilities investors can’t easily see are now larger than the ones they can.

Where the debt hides

The mechanism is not fraud; it’s financial engineering that keeps future obligations off the formal liabilities line. Instead of borrowing directly to build data centers — which would add debt to the balance sheet and pressure credit ratings — the hyperscalers increasingly lock in the same economic commitments through structures that don’t register as debt today:

• Long-term data-center leases, especially operating leases and deals signed before construction begins, which commit years of cash outflows without appearing as borrowings.
• GPU and compute supply commitments — multi-year purchase obligations for chips and capacity that function like debt but are booked as commitments.
• Joint ventures with private-credit funds, where a fund or special-purpose vehicle raises the capital and the hyperscaler contracts to pay for the capacity, moving the borrowing off the tech company’s books entirely.

Each structure locks in future cash outflows. None of them lands in the debt figure a casual reading of the balance sheet would surface — which is precisely why the Nikkei estimate is so much larger than the reported totals.

Meta is the starkest case

Among the five, Meta stands out. Its off-balance-sheet debt is estimated at around $420 billion — nearly three times its transparent, on-balance-sheet debt load. That ratio captures how far the financing has shifted: for Meta, the visible borrowing is now the minority of its actual commitments. The company’s aggressive data-center expansion, including deals like its $10 billion compute arrangement with Anthropic, runs through exactly the kinds of leases and supply contracts the study flags.

Wall Street’s own numbers agree

This is not just a Nikkei framing. Morgan Stanley’s estimate of the broader AI financing picture puts total off-balance-sheet exposure across the industry even higher — around $1.8 trillion — broken down as:

• Nearly $1 trillion in purchase commitments,
• More than $800 billion in leases that haven’t yet begun, and
• About $110 billion in accounts-payable financing.

The convergence of an independent study and a major bank’s own tally on a figure in the mid-trillions is the point. This is a structural feature of how the AI buildout is being funded, not a one-off accounting quirk at a single company.

Why the structure exists

The appeal is straightforward. Direct borrowing at this scale would balloon reported leverage, potentially threaten credit ratings, and force uncomfortable conversations about free cash flow as capex consumes an ever-larger share of it. Off-balance-sheet structures let a company secure the same physical capacity — the racks, the chips, the power — while keeping the headline leverage ratios that ratings agencies and index funds watch looking healthier than the underlying commitments imply.

That instinct is amplified by the sheer velocity of spending. The AI capex cycle has pushed infrastructure outlays to sums with no real precedent, a dynamic we’ve traced in the AI capex boom. When you’re committing hundreds of billions a year, the difference between “debt” and “commitment” on the balance sheet becomes a strategic lever, not an afterthought.

The risk: obsolescence meets leverage

The concern isn’t just that the debt is hidden — it’s what the debt is buying. A traditional lease finances a building that depreciates slowly over decades. A large slice of these AI commitments finances GPUs and specialized hardware that can be functionally obsolete in a few years as new chip generations land. Locking in multi-year cash outflows against fast-depreciating assets is a different risk profile than the lease accounting was designed for.

Layer on the financing structure and the exposure compounds. As more of the buildout runs through private-credit funds and joint ventures, risk migrates from the hyperscalers’ rated balance sheets toward private lenders — a market that is less transparent and less tested in a downturn. Bond investors have already started pushing back on the volume of AI-related debt hitting public markets, even as issuance like Amazon’s $25 billion bond sale shows how much appetite the buildout still requires. The off-balance-sheet layer sits on top of all of that.

What the reported numbers miss

For anyone reading these companies through a standard leverage screen, the takeaway is that the screen is incomplete. A debt-to-equity ratio built on reported borrowings understates the real forward commitments by more than half in aggregate — and by roughly 3x in Meta’s case. The economics of the AI data-center buildout, which we’ve examined in AI data center economics, only make sense once these commitments are counted, because they represent the true cost of the capacity the revenue projections depend on.

What it means

The headline isn’t that Big Tech is doing something illegal — lease and commitment accounting is standard and disclosed in the footnotes. The headline is scale and visibility: the AI buildout has grown so large, so fast, that the most economically meaningful liabilities have migrated to the parts of the financial statements most investors never read. When the invisible debt exceeds the visible debt across five of the most valuable companies on earth, the risk picture that flows from reported leverage ratios is materially wrong.

The winners of this structure, for now, are the hyperscalers themselves — they get the capacity while keeping reported leverage manageable — and the private-credit funds collecting yield to warehouse the risk. The losers, if the cycle turns, would be whoever holds that risk when demand softens or a chip generation renders committed hardware uncompetitive faster than the contracts amortize.

Watch three things. First, the footnotes, not the headline debt line, in the Q2 earnings that Big Tech is reporting this week and next — purchase commitments and future-lease disclosures are where the real number lives. Second, private-credit health, since that market is increasingly the counterparty to these deals and the least visible link in the chain. Third, the depreciation debate: how quickly these companies write down AI hardware will tell you whether they’re treating GPUs like long-lived buildings or the faster-aging assets they actually are. The buildout’s financing has been the quiet story beneath the model launches — and $1.65 trillion is a hard number to keep quiet.