Michael Burry, the investor famous for profiting from the 2008 subprime mortgage crisis, is once again taking a contrarian stance against market consensus. This time, his focus is on the heavy financial obligations hidden beneath the artificial intelligence arms race. In a recent blog post, Burry warned that Amazon, Meta, Alphabet, Microsoft, and Oracle, the five major hyperscalers, have quietly accumulated over $3 trillion in off-balance-sheet commitments, a risk largely ignored by the market.

Burry estimates that unstarted lease commitments approach $1.2 trillion, while purchase commitments exceed $1.5 trillion. When including special purpose entities, guarantees, and contingent support agreements, the total surpasses $3 trillion. "These liabilities, in my opinion, are essentially growing at an extremely rapid pace," Burry wrote, emphasizing that the expansion of these obligations far outpaces the companies' own growth.

To grasp the scale of $3 trillion, consider this: an analysis of financial statement footnotes from July showed that the five companies' total disclosed debt on their balance sheets was approximately $1.35 trillion, manageable for firms generating hundreds of billions in annual free cash flow. However, the same companies carry about $1.65 trillion in additional off-balance-sheet obligations, exceeding their total disclosed debt and growing roughly eightfold over four years, closely tracking the trajectory of the AI capital expenditure race.

The mechanics behind this are neither complex nor novel. Under US GAAP ASC 842, lease obligations are only recognized on the balance sheet when the lease has actually "commenced." Until data centers become operational, legally binding leases that have been signed remain in financial statement footnotes. GPU purchase agreements, leases for data centers not yet opened, and joint ventures with private credit funds all generate real, contractually obligated future cash outflows, but under current accounting rules, they are not recorded as debt until the facility begins operations.

Meta Platforms Inc (NASDAQ: META) represents the most extreme case of this structure. Its off-balance-sheet obligations have reached approximately $420 billion, nearly three times its disclosed debt. A significant portion relates to a joint venture with Blue Owl Capital (NYSE: OWL) financing the construction of a massive data center in Louisiana, where Meta holds a minority stake and leases back computing capacity on a long-term basis rather than owning the facility outright.

Oracle Corp (NYSE: ORCL) has seen its hidden debt balloon roughly thirtyfold in four years, tied to its construction commitments for OpenAI's computing needs. As of the end of May, Oracle's separately disclosed unstarted lease commitments stood at $260 billion, nearly seven times its recognized lease liabilities of $37.89 billion. The Bank for International Settlements warned in March about "shadow lending" in AI construction, and Moody's separately quantified $662 billion in unstarted data center lease commitments for these five companies in February, equivalent to 113% of their combined adjusted on-balance-sheet debt.

The core of Burry's warning is not about the legality of these commitments, but a deeper maturity mismatch. Data center construction typically takes three to five years, with lease terms often spanning 13 to 20 years. However, AI accelerator power density and cooling requirements can change significantly within 12 to 18 months. In other words, physical infrastructure designed for 20 years now houses chips that iterate every year or so.

Burry cited Microsoft Corp (NASDAQ: MSFT) CEO Satya Nadella's earlier comment: "I don't want to be locked into massive depreciation of one generation of hardware." This statement is telling, as it reflects concerns about what is happening on Microsoft's own balance sheet. This mismatch creates two accounting consequences. First, the five companies currently hold over $400 billion in construction-in-progress assets that, not yet in service, generate no depreciation under GAAP. "Nvidia chips may have economically depreciated while waiting in warehouses, yet GAAP depreciation remains zero," Burry noted. Second, even when assets become operational, the rationality of depreciation assumptions is highly debatable. Major cloud providers typically set AI server depreciation periods at 5 to 6 years, but Burry believes the actual economic life of GPUs, given Nvidia's iteration cycle compressing from 18-24 months to about 12 months, is closer to 2 to 3 years.

Burry previously estimated that by stretching GPU lifespans to 5-6 years, hyperscalers would understate depreciation by approximately $176 billion between 2026 and 2028. Notably, Amazon (NASDAQ: AMZN) in 2025 shortened the useful life of some servers from 6 years back to 5 years, following a study finding that technology iteration in AI and machine learning exceeded prior expectations. This marked the first instance among the five giants where a company walked back its depreciation assumptions. Amazon's adjustment partially validates Burry's view that depreciation periods are being overly extended.

Meta's approach moves in the opposite direction. The company increased server depreciation periods from 4-5 years to 5.5 years, adding approximately $2.9 billion in book profit that year, representing 4% of its pre-tax income. Over the past five years, the other four giants have also repeatedly "optimized" profit figures through similar methods.

Among the five companies, Burry specifically called out Oracle. His reasoning is direct: analysts' average price target for Oracle is significantly above the current market price, which he sees as a sign of excessive optimism. The data supports this observation. According to S&P Global's survey of 43 analysts, Oracle has a consensus "buy" rating with an average price target of approximately $238. Some models even project targets as high as $420. Burry himself began shorting Oracle from around $145 and previously held put options on the stock.

Oracle's off-balance-sheet exposure is indeed the most concentrated among the five. Its $260 billion in unstarted lease commitments is nearly seven times its recognized lease liabilities, an extreme level among large tech companies. S&P Global Ratings has factored this into its assessment. Meanwhile, Oracle's total debt stands at approximately $167 billion, with its capital structure weakening amid large-scale equity and debt financing. These commitments primarily target data centers with lease terms typically ranging from 15 to 19 years. Oracle's situation warrants separate scrutiny because its off-balance-sheet commitments are highly tied to OpenAI's computing demands. If OpenAI's commercialization falls short of expectations, Oracle faces not just idle capacity issues but a chain reaction of contract termination costs, asset impairments, and rising financing costs.

Burry's criticism extends beyond corporate accounting to Wall Street's analytical frameworks. His core argument is that many analysts fail to adequately incorporate maintenance capital expenditures and economic depreciation of AI infrastructure into their valuation models. This is not an easily refutable point. Hyperscaler capital expenditures are expanding at an astonishing pace. According to institutional estimates, the five companies' combined 2026 capital expenditures will reach $769.2 billion, nearly double 2025 levels, potentially exceeding $1 trillion by 2027. Microsoft alone is expected to spend $190 billion in 2026, up approximately 130% year-over-year. The combined 2026 capital expenditures for Amazon, Alphabet, Microsoft, and Meta are projected between $700 billion and $760 billion.

The financing methods for these expenditures are also changing. The five companies have raised over $200 billion through bond markets this year, more than double 2025 levels. Morgan Stanley expects global AI-related debt issuance to approach $570 billion in 2026. Amazon's free cash flow turned negative in Q2 2026, and Alphabet posted its first quarterly negative free cash flow since going public decades ago. Morgan Stanley forecasts that Amazon, Oracle, and Meta will see free cash flow approach zero or turn negative in 2026.

For now, however, the market remains selectively blind to these signals. The Roundhill Magnificent Seven ETF (NYSEARCA: MAGS), which tracks large tech stocks, is heading for a third consecutive monthly gain. Strong cloud business growth last quarter continues to support investor optimism.

Not everyone agrees with Burry's assessment. Tom Lee of Fundstrat publicly refuted Burry's warning comparing the current situation to the Enron scandal, arguing that $3 trillion in off-balance-sheet commitments "reflects almost no actual risk to investors." His argument is that such large-scale paper transactions are normal operations in the financial sector, backed by high-tech profits. The essence of this disagreement lies in whether off-balance-sheet commitments should be viewed as "future arrangements not yet realized" or "real liabilities already locked in." Optimists argue that fulfilling these commitments depends on continued AI demand growth, and current strong cloud business performance validates this premise. Pessimists point out that if demand growth slows, even from "hyper-speed" to merely "fast," these commitments will rapidly transform from footnotes into real numbers on income statements through asset impairments, contract termination costs, and idle capacity.

Burry's language in this latest warning is relatively measured. He explicitly stated he is not alleging fraud but is issuing a serious caution about the downside risks of the current AI construction wave if demand weakens. "When the music stops, these off-balance-sheet commitments will quickly become real liabilities."