Big Tech's Off-Balance-Sheet AI Debt: What the $1.65 Trillion Figure Actually Means

Educational content only. This article is not investment advice and is not a recommendation to buy, sell, or hold Alphabet, Amazon, Meta, Microsoft, Oracle, any other company named, or any fund, index, or bond discussed below.

In July 2026, a Nikkei Asia investigation put a number on something analysts had been circling for months: the five largest US hyperscalers — Alphabet, Amazon, Meta, Microsoft, and Oracle — are sitting on roughly $1.65 trillion in obligations tied to their AI data-center buildout that don't appear as debt on their balance sheets, exceeding the roughly $1.35 trillion in debt those same five companies report directly (Nikkei Asia — Five US tech giants' hidden debts soar to $1.65tn on opaque AI funding). Nikkei's study found this off-balance-sheet figure has grown roughly eightfold in about four years, tracking the industry's AI infrastructure spending (Nikkei Asia — Five US tech giants' hidden debts soar to $1.65tn on opaque AI funding).

None of this is illegal, and none of it is secret in the sense of being concealed from regulators — these obligations are disclosed in each company's own SEC filings, and this article draws directly on several of them. What's true, and worth understanding if you hold or are researching any of these five stocks, is that a headline "total debt" figure on a balance sheet or a stock-screener page doesn't capture the full scale of what these companies have committed to. This piece walks through what "off-balance-sheet" actually means under US accounting rules, why the AI buildout is producing so much of it, how the five companies compare, the parallel surge in ordinary corporate bonds, and the specific risks analysts and rating agencies are pointing to — without predicting how any of it plays out.

Where these figures come from

A brief methodology note, because how a number like "$1.65 trillion" is built matters as much as the number itself.

The $1.65 trillion figure is not a line item any company reports — it's an aggregate estimate that Nikkei Asia's analysts built by pulling forward-looking lease commitments, GPU and server purchase obligations, and joint-venture/special-purpose-vehicle debt disclosed across each company's own SEC filings, then summing them across the five firms (Nikkei Asia — Five US tech giants' hidden debts soar to $1.65tn on opaque AI funding). Because it's an aggregate built from footnote disclosures rather than a single reported figure, different analysts drawing the same boundary differently get different totals. Moody's Ratings, working from a narrower definition limited specifically to data-center lease commitments that have been signed but not yet started, put the same five companies' off-balance-sheet lease exposure at roughly $662 billion — equivalent to about 113% of their combined on-balance-sheet adjusted debt — as part of nearly $1 trillion in total long-term lease obligations Moody's expected by the end of 2025 (Data Center Dynamics — Moody's: Hyperscalers understating risks of short-term AI DC lease agreements; Bisnow — Moody's: $662B In Data Center Leases Hidden From Big Tech Balance Sheets Poses Looming Risk). Moody's number is smaller than Nikkei's mainly because it excludes GPU and server purchase commitments and joint-venture debt that Nikkei's estimate includes — a good illustration that "off-balance-sheet AI obligations" is a real and growing category, but not yet a single, standardized figure the way "total debt" is. The Bank for International Settlements — the institution that coordinates central banks globally — reached a similar structural conclusion in its own March 2026 analysis, describing hyperscalers' off-balance-sheet financing as a distinct and growing channel alongside their conventional bond issuance (Bank for International Settlements — Financing the AI infrastructure boom: on- and off-balance sheet borrowing, BIS Quarterly Review). Figures in this space move quickly and are frequently revised; treat every number below as accurate as of its cited source and date, not as a live figure.

What "off-balance-sheet" actually means

A balance sheet is supposed to show what a company owns and what it owes as of a specific date. Debt that's already been borrowed — a bond that's been issued, a loan that's been drawn — has to show up there under US accounting rules. The obligations at the center of this story are different: they're commitments to pay in the future, for things that haven't happened yet, and current accounting standards genuinely don't require all of them to be recorded as liabilities today. There are three main mechanisms.

Leases for facilities that aren't running yet. Under the current lease accounting standard, ASC 842, most leases longer than 12 months are recorded on the balance sheet as a right-of-use asset and a matching lease liability once the lease has commenced (Visual Lease — ASC 842 Balance Sheet Guide with Examples). But a signed lease for a data center that's still under construction generally hasn't "commenced" yet in the accounting sense — it sits in a footnote as a future commitment, not a balance-sheet liability, until the facility is actually delivered and put into use (Data Center Dynamics — Moody's: Hyperscalers understating risks of short-term AI DC lease agreements). Moody's has also noted a related wrinkle: under GAAP, a lease renewal option only gets added to the recorded liability if renewal is "reasonably certain" — a high bar — so multi-year renewal options that companies may well exercise can also stay off the books (Data Center Dynamics — Moody's: Hyperscalers understating risks of short-term AI DC lease agreements).

GPU and server purchase commitments. Hyperscalers routinely sign multi-year agreements to buy specific quantities of chips, servers, or cloud capacity at agreed prices. Standard accounting treats an unconditional purchase obligation for goods not yet delivered as a disclosed future commitment, not a present liability. Oracle's most recent quarterly filing, for example, discloses roughly $11 billion of unconditional purchase obligations as of February 28, 2026, primarily related to data-center power arrangements — visible in the filing's footnotes, not on the balance sheet itself (Oracle Corporation — Form 10-Q, quarter ended February 28, 2026, SEC EDGAR). Amazon's most recent annual report discloses roughly $440 billion in total unconditional purchase obligations across all categories — spanning energy contracts, content licensing, property and equipment, and software, not solely AI infrastructure — payable over multiple years (Amazon.com, Inc. — Form 10-K, fiscal year 2025, SEC EDGAR).

Joint ventures and special-purpose vehicles (SPVs). Increasingly, hyperscalers are financing entire data-center campuses through a separate legal entity — often majority-owned by a private-credit investor rather than the tech company itself — that borrows the construction money, builds the facility, and leases it back to the hyperscaler as the anchor tenant. Because the hyperscaler typically holds a minority equity stake and isn't the primary party obligated on the debt, that debt sits on the joint venture's books, not the tech company's, even though the tech company is economically the reason the facility exists (Bank for International Settlements — Financing the AI infrastructure boom, BIS Quarterly Review). The next section walks through a real, recent example of exactly this structure.

A worked example: Meta, Blue Owl, and the Hyperion campus

In October 2025, Meta announced a joint venture with funds managed by private-credit firm Blue Owl Capital to develop "Hyperion," a large AI data-center campus in Richland Parish, Louisiana, with Meta holding a 20% stake and Blue Owl-managed funds holding the remaining 80% (Meta — Meta Announces Joint Venture with Funds Managed by Blue Owl Capital to Develop Hyperion Data Center; CNBC — Meta partners with Blue Owl Capital on $27 billion AI data center project). Morgan Stanley arranged the financing and served as sole bookrunner, with Pacific Investment Management Co. (PIMCO) as the anchor lender on debt that S&P rated A+ and that matures in 2049 (Fortune — Meta's $27 billion bet turns AI compute into Wall Street's hottest new investment). Meta is expected to be the facility's sole tenant once it's operational.

Reported figures for the exact size of the package vary by source and by date as the deal was finalized — CNBC's initial October 2025 report described a $27 billion project, while later reporting on the closed financing described a $30 billion package (CNBC — Meta partners with Blue Owl Capital on $27 billion AI data center project; Fortune — Meta's $27 billion bet turns AI compute into Wall Street's hottest new investment) — a good example of why we're flagging methodology throughout this piece: even a single, well-covered deal gets reported with different topline numbers depending on when a source is describing it (announcement versus close) and what's included (debt only, versus debt plus equity). What's consistent across sources is the structure: an SPV/joint venture, majority-owned by a private-credit sponsor, borrowing the money and owning the asset, with Meta as tenant rather than balance-sheet owner (Fortune — Meta's $27 billion bet turns AI compute into Wall Street's hottest new investment; Bank for International Settlements — Financing the AI infrastructure boom, BIS Quarterly Review). Meta's own SEC filings show the broader pattern this deal is part of: the company's disclosed leases that had not yet commenced, mostly for data centers and network infrastructure, rose from about $35 billion in its first-quarter 2025 filing to about $58 billion by its third-quarter 2025 filing (Meta Platforms, Inc. — Form 10-Q, quarter ended September 30, 2025, SEC EDGAR).

The company breakdown — and where the public reporting is thinner

Nikkei's reporting gave detailed figures for two of the five companies. Meta's off-balance-sheet obligations were put at roughly $420 billion, described as around three times the debt Meta reports directly on its balance sheet (Nikkei Asia — Five US tech giants' hidden debts soar to $1.65tn on opaque AI funding). Oracle's figure was put at roughly $273.3 billion, an increase of about 2,900% — roughly 30-fold — since 2022, as the company has taken on enormous data-center commitments tied to its cloud contracts, including with OpenAI (Nikkei Asia — Five US tech giants' hidden debts soar to $1.65tn on opaque AI funding). Available secondary coverage of the Nikkei study emphasized Meta's and Oracle's figures specifically; comparably detailed individual figures for Alphabet, Amazon, and Microsoft were not consistently reported in the coverage we reviewed, so this article doesn't assert specific numbers for those three companies beyond noting they're included in the five-company, $1.65 trillion total.

Oracle's growth is corroborated independently: S&P Global Ratings downgraded Oracle's long-term credit rating one notch to BBB- on July 9, 2026 — one step above speculative grade — citing roughly $260 billion of additional lease commitments expected to begin between fiscal 2027 and 2029, on top of a $638 billion remaining performance-obligation backlog of which OpenAI accounts for roughly half (S&P Global Ratings — Research Update: Oracle Corp. Downgraded To 'BBB-/A-3'). S&P's own figure for Oracle's future lease commitments (~$260 billion) lands close to, though not identical to, Nikkei's $273.3 billion estimate — consistent with the broader point that different institutions counting the same underlying disclosures arrive at similar but not identical totals.

The parallel story: an ordinary corporate bond boom

Off-balance-sheet financing is only part of how this buildout is being funded. The same five hyperscalers have also become unusually large, unusually frequent borrowers in the ordinary corporate bond market — debt that does show up on their balance sheets. Meta priced a $30 billion, six-tranche bond offering on October 30, 2025, after drawing a record $125 billion in investor orders — Bloomberg described it as the largest public high-grade corporate bond sale of that year and the largest non-acquisition-related investment-grade deal on record (Bloomberg — Meta Sets Bond Sale at $30 Billion, Biggest Offering of the Year). Oracle raised $18 billion in a six-part bond sale in late September 2025, including a rare 40-year tranche, to help fund its data-center commitments (Data Center Dynamics — Oracle takes on $18bn in debt ahead of AI data center build-out), then returned to the market in February 2026 with a $25 billion, eight-part offering alongside a separate $20 billion equity distribution agreement (IFR — Oracle raises US$25bn in bond offering as AI funding spree continues). Alphabet and Amazon both issued bonds in November 2025 (roughly $17.5 billion and $15 billion, respectively), and Alphabet returned in February 2026 with a $20 billion, seven-part offering that included a rare 100-year, pound-sterling-denominated tranche — the first century bond from a technology company since Motorola in 1997 — filed with the SEC as part of a formal prospectus (Alphabet Inc. — Form 424B5 prospectus supplement, SEC EDGAR; CNBC — Why Alphabet's 100-year sterling bond is raising new fears over debt-fuelled AI arms race). Amazon added another leg to this in July 2026, raising at least $25 billion in a fresh bond sale while telling investors it did not plan further debt issuance for the remainder of that year (CNBC — Amazon raising at least $25 billion in bond sale, won't issue more debt in 2026).

Bank of America's credit strategists initially projected roughly $140 billion a year of hyperscaler bond issuance over several years, before raising that full-year 2026 forecast to $175 billion after Amazon's issuance came in larger than expected — a forecast that, if it holds, would put the five hyperscalers' annual borrowing on a similar scale to that of the largest US banks (Bloomberg — Bank of America Increases Hyperscaler Issuance Forecast to $175 Billion). The Bank for International Settlements independently noted that combined hyperscaler bond issuance topped $100 billion in 2025 and that credit-default-swap spreads on this group of borrowers — a market-based measure of how expensive it is to insure against their default — have risen, particularly for the hyperscalers carrying lower credit ratings (Bank for International Settlements — Financing the AI infrastructure boom, BIS Quarterly Review).

The "Enron" comparison — and why it needs a caveat

The following reports named commentators' published opinions, quoted accurately and linked to their original source. It does not reflect Zydhra's own assessment of any company, and it is not an allegation of wrongdoing, concealment, or fraud against Alphabet, Amazon, Meta, Microsoft, Oracle, or any other company named in this article.

Some commentary on the Nikkei findings has drawn a direct parallel to Enron, the energy trading company whose 2001 collapse became a landmark case of off-balance-sheet special-purpose entities concealing debt from investors. A Bloomberg Tax analysis on the current story quoted Gil Luria, head of technology research at D.A. Davidson & Co., drawing a specific distinction worth repeating in full: "Enron's crime wasn't having special purpose vehicles. Enron's crime was hiding them" — noting that today's hyperscaler structures comply with GAAP and IFRS accounting rules and are disclosed in each company's public SEC filings (Bloomberg Tax — Big Tech AI Spree Revives Accounting Devices That Toppled Enron). Separately, media commentator Ed Zitron has argued the scale of the disclosed-but-easy-to-miss figure "should be front page news everywhere" and "should be a shareholder riot," in comments made on the Prof G Markets podcast (Yahoo Finance UK — Big Tech's Off-Balance-Sheet AI Debt Reportedly Tops $1.65 Trillion: Ed Zitron Says It 'Should Be a Shareholder Riot').

Both points can be true at once, and this article isn't taking a side: the accounting mechanisms genuinely echo tools involved in the Enron case, and separately, the current use of those mechanisms is lawful, GAAP-compliant, and disclosed in filings that are publicly available on SEC EDGAR — which is precisely how the sourcing in this article was gathered. Whether the current scale of off-balance-sheet AI financing represents a reasonable, well-disclosed way to fund a capital-intensive buildout, or a transparency problem that under-informs investors relying on headline balance-sheet figures, is an actively debated question among analysts and rating agencies — not a settled fact, and not one this article is resolving.

Risks worth understanding, in plain terms

The following reports specific, cited concerns raised by named rating agencies, central-bank researchers, and analysts about the financing structures described above. It is not Zydhra's own risk assessment of any company, bond, or security, and it is not a recommendation to buy, sell, or hold Alphabet, Amazon, Meta, Microsoft, Oracle, or any other security discussed in this article.

A few structural features of this financing wave are worth understanding on their own terms, independent of any prediction about what happens next.

Tenant and customer concentration. S&P's Oracle downgrade specifically cited the concentration of Oracle's contracted backlog in a single customer, OpenAI, alongside a projected free-cash-flow deficit of nearly $42 billion for fiscal 2027 (S&P Global Ratings — Research Update: Oracle Corp. Downgraded To 'BBB-/A-3'). Structures like the Meta-Blue Owl joint venture similarly depend on a single anchor tenant for a facility financed with billions in third-party debt.
Leases converting to balance-sheet liabilities all at once. Moody's has flagged that as data centers currently under construction go live over the next several years, roughly $662 billion (by its narrower count) of currently off-balance-sheet lease commitments will convert to on-balance-sheet liabilities, potentially in concentrated waves rather than gradually (Data Center Dynamics — Moody's: Hyperscalers understating risks of short-term AI DC lease agreements).
Interest-rate and refinancing exposure. Much of this financing — both the on-balance-sheet bonds and the off-balance-sheet private-credit structures — was priced in a specific interest-rate environment; BIS researchers have noted that credit spreads on lower-rated hyperscaler debt have already widened somewhat as issuance volumes have grown (Bank for International Settlements — Financing the AI infrastructure boom, BIS Quarterly Review).
"Circular financing" concerns. Analysts, including those at Bernstein Research, have raised questions about deals in which a chipmaker invests in an AI company, which then commits large sums to a cloud provider, which in turn spends heavily with the original chipmaker — a closed loop that can make reported revenue and commitments harder to interpret independently (Business Standard — Nvidia-OpenAI deal sparks concerns over circular financing in AI boom).
GPU depreciation assumptions. Separately from the debt question, investor Michael Burry has publicly argued that hyperscalers depreciate AI chips over five-to-six-year schedules that may be longer than the chips' actual useful life given the pace of new hardware generations — a claim Nvidia has publicly disputed — and CNBC has reported that some hyperscalers changed their own useful-life estimates for servers in different directions during 2025 (CNBC — The question everyone in AI is asking: How long before a GPU depreciates?). This is a distinct, actively disputed accounting question from the off-balance-sheet debt topic above, and it's included here because both affect how much a hyperscaler's reported profitability tells you about its underlying AI economics.

None of these points are predictions that any company's debt structure will or won't cause problems — they're the specific, named concerns that credit-rating agencies, central-bank researchers, and financial journalists have raised, cited to where each one was raised.

How this connects to the AI hardware demand story

This off-balance-sheet financing exists to fund a very physical thing: the servers, chips, and buildings that make up an AI data center. Zydhra's explainer on memory chips and the semiconductor stock market covers a related piece of that same buildout — the DRAM, NAND, and HBM memory that sits inside those servers, and why that corner of the chip industry has its own long history of boom-and-bust cycles. Understanding the financing side covered here and the hardware-demand side covered there together gives a fuller picture of what's actually driving the capital intensity behind current AI-sector headlines.

What this means if you're reading a hyperscaler's financial statements

The practical takeaway isn't that any of these five companies has done anything improper — the obligations described here are disclosed, GAAP-compliant, and sitting in plain view in SEC filings that are free to read on EDGAR. The takeaway is narrower: a company's reported "total debt" line is not the complete picture of what it has committed to pay, particularly for a business investing this heavily in long-lived physical infrastructure. Reading the commitments and contingencies footnotes — not just the balance sheet — is where a more complete picture of a company's forward obligations actually lives, and that applies to reading any capital-intensive company's filings, not just these five.

This article hasn't taken a position on whether any of these companies' AI investments will pay off, whether their credit ratings will change further, or whether the current financing structures represent good or poor capital allocation. Those are judgment calls that depend on assumptions about future AI demand, interest rates, and each company's execution — questions this article isn't positioned to answer, and that depend on an individual investor's own research, risk tolerance, and circumstances. If you're researching any of these companies, consider reading their SEC filings directly, comparing how different analysts and rating agencies define "off-balance-sheet" obligations given how much the boundary varies, and, if appropriate, speaking with a licensed financial professional about how any of this fits your own situation.

Disclaimer: This article is for general educational and informational purposes only. It is not investment advice, and nothing in it should be construed as a recommendation to buy, sell, or hold Alphabet, Amazon, Meta, Microsoft, Oracle, Blue Owl Capital, any other company named, or any bond, fund, index, or security discussed above. Zydhra is a portfolio-tracking and analytics tool — it is not a broker-dealer and is not a registered investment adviser. Debt figures, lease commitments, credit ratings, bond-issuance totals, and deal terms discussed above change frequently, were current as of the dates and sources cited, and may no longer be accurate by the time you read this — different research providers and rating agencies have also reported different figures for overlapping periods, as noted throughout this article. Past performance, including past credit-rating actions, bond pricing, and market reception, does not guarantee or predict future results. Before making any investment decision, consider doing independent research, reviewing companies' primary SEC filings, and consulting a licensed financial professional.