Burry Warns AI Big Five Harbor Over $3 Trillion in Off-Balance-Sheet Risk
nashnova research
Michael Burry — the investor who called the 2008 subprime crash — warns that Amazon, Meta, Alphabet, Microsoft, and Oracle have quietly accumulated over $3 trillion in off-balance-sheet commitments. This means → the market may be drastically underpricing the real liabilities behind the AI infrastructure boom.
What exactly is this $3 trillion "invisible tab"?
Burry estimates the five hyperscalers' off-balance-sheet commitments include nearly $1.2 trillion in uncommenced leases, over $1.5 trillion in purchase commitments, plus special-purpose entities and guarantees — totaling more than $3 trillion.
In plain terms = the money is contractually locked in, but because the data centers haven't started operating, accounting rules don't count it as "debt." It sits buried in footnotes.
The mechanism: U.S. GAAP rule ASC 842 — a lease-accounting standard — says lease obligations hit the balance sheet only when the lease "commences." This means → GPU purchase contracts and signed-but-unopened data center leases create real cash obligations, yet reported debt shows zero.
Which two companies have the most extreme off-balance-sheet exposure?
Meta's total off-balance-sheet obligations reach roughly $420 billion — nearly three times its disclosed debt. Much of this ties to a joint venture with Blue Owl Capital building a mega data center in Louisiana: Meta holds a minority stake and "buys back" compute via long-term leases, keeping the facility off its own books.
Oracle's hidden liabilities have ballooned roughly thirtyfold in four years, driven by build-out commitments for OpenAI's compute needs. Its disclosed uncommenced lease commitments alone hit $260 billion — nearly seven times its recognized lease liabilities of $37.89 billion.
Burry himself has been shorting Oracle from around $145 and previously held put options on the stock. This reflects a bet backed by real money, not just commentary.
What is the core risk logic — duration mismatch and understated depreciation?
Data center leases run 13 to 20 years, but AI accelerator power density and cooling requirements can shift dramatically within 12 to 18 months. In plain terms = the building is leased for 20 years, but the gear inside could be obsolete in a year and a half.
Microsoft CEO Satya Nadella himself has said: "I don't want to be locked into massive depreciation of one generation of hardware." This means → even insiders acknowledge that hardware refresh cycles and lease tenors are severely mismatched.
On depreciation: the five companies hold over $400 billion in construction-in-progress assets not yet in service — GAAP depreciation is zero. Burry estimates that by stretching GPU useful life to 5–6 years, hyperscalers will understate depreciation by roughly $176 billion between 2026 and 2028.
How are companies already gaming depreciation assumptions?
Amazon shortened some server useful lives from 6 years back to 5 in 2025, citing faster-than-expected AI technology turnover — the first of the Big Five to walk a depreciation assumption backward.
Meta moved the other way, extending server depreciation from 4–5 years up to 5.5 years. That single change added roughly $2.9 billion to reported profit that year — 4% of pre-tax income.
This means → the same class of asset, two opposite useful-life judgments from two companies. Depreciation assumptions have become a profit-management lever, not just an accounting technicality.
Where does Wall Street stand — forward planning or real liability?
Analysts project the five companies' combined 2026 capex at $769.2 billion — nearly double 2025 — with 2027 potentially exceeding $1 trillion. Amazon's free cash flow already turned negative in Q2 2026; Alphabet posted its first negative quarterly free cash flow in decades of being public.
Fundstrat's Tom Lee publicly pushed back on Burry, arguing that $3 trillion in off-balance-sheet commitments "can hardly reflect actual risk for investors" given the high-tech profit base behind them. The bull case rests on one assumption: AI demand keeps growing fast enough for cloud revenue to cover these costs.
Burry himself was measured — he explicitly said he is not alleging fraud — but warned: "When the music stops, these off-balance-sheet commitments turn into real liabilities fast." This means → the real disagreement is not about accounting rules but about whether AI demand growth can keep outrunning the long-term costs already locked in.
Have regulators started paying attention?
The Bank for International Settlements flagged "shadow lending" in AI infrastructure buildout in a March report this year.
Moody's separately quantified $662 billion in uncommenced data center lease commitments across the five companies in February — equivalent to 113% of the group's adjusted on-balance-sheet debt.
This reflects that the concern is not Burry's alone — international regulators and rating agencies have begun independently tracking this exposure.
市场有风险，内容仅供研究参考，不构成投资建议。