It's not a scam, but the risk is real! Big bear Bury warns the AI Big Five are hiding a $3 trillion risk. “When the music stops, promises outside the box will soon turn into real liabilities”

The Zhitong Finance App learned that Michael Burry, the prototype “Big Short” who became famous due to accurate shorting of the 2008 subprime mortgage crisis, once again stood on the opposite side of market consensus. And this time, he pointed the finger at the heavy ledger behind the artificial intelligence (AI) arms race.

Bury once again warned in his blog that Amazon (AMZNS.US), Meta (META.US), Alphabet (GOOGL.US), Microsoft (MSFT.US), and ORCL.US (ORCL.US) — the five hyperscalers (hyperscalers) — have quietly accumulated more than $3 trillion in external promises, yet the market is generally ignoring this risk.

He estimates that uninitiated lease commitments are close to $1.2 trillion, and procurement promises exceed $1.5 trillion; if special purpose entities, guarantees, and contingent support commitments are included, the total scale exceeds $3 trillion.

“These liabilities, I think, are essentially growing at a very high rate,” Barry wrote. In particular, he emphasized that these obligations are expanding far faster than the company's own growth.

An “off-surface machine” that has been in operation for two years

To understand what $3 trillion means, we must first look at a set of comparative data. Some analysts discovered in July by sorting through the footnotes in the latest financial statements that the total debt disclosed in the balance sheets of these five companies is about 1.35 trillion US dollars, which is still manageable for companies with annual free cash flow of tens of billions of dollars. But the same five companies are burdened with additional obligations of about $1.65 trillion outside the balance sheet — a figure that exceeds the sum of all of their disclosed debts and has grown about eight times over four years, almost accurately corresponding to the trajectory of the AI capital expenditure competition.

The causes of this mechanism are not complicated, and there is nothing new to talk about. According to US GAAP (GAAP) ASC 842, lease obligations are recorded on the balance sheet only when the lease actually “begins”. Until the data center is in operation, those signed, legally binding leases only need to lie in the financial footnote. GPU purchase agreements, leasing contracts for unopened data centers, joint venture arrangements with private equity funds — these will all generate real, contractually bound future cash expenses, but under current accounting standards, they are not included in debt until the facility “begins operation.”

Meta is the most extreme example of this structure. According to the analysis at the time, its total off-balance sheet obligations had reached about $420 billion, nearly three times the disclosed debt. A large portion of this was related to Blue Owl Capital (OWL.US)'s joint venture arrangement to finance the construction of a very large data center in Louisiana — Meta held minority shares and bought back computing power through long-term leases rather than directly owning the facilities. Oracle's hidden debt has expanded about 30 times in four years, which is closely related to its construction promises to meet OpenAI's computing power requirements. By the end of May, Oracle's separately disclosed uninitiated lease commitments reached US$260 billion, almost seven times its confirmed lease liabilities (US$37.89 billion).

The Bank for International Settlements issued a warning about “shadow lending” in AI construction in March of this year. In its February report, Moody's also separately quantified the five companies' uninitiated data center lease commitments totaling $662 billion — equivalent to 113 percent of the group's adjusted debt at the time.

Term mismatch: 20-year lease versus 18-month chip cycle

The core logic of Burry's warning this time is not whether these promises themselves are legal, but rather a deeper mismatch of deadlines.

The construction period for data centers is usually three to five years, and leases are often as long as 13 to 20 years. However, the power density and cooling requirements of AI accelerators may change significantly within 12 to 18 months. In other words, the physical infrastructure designed for 20 years contains chips that need to be iterated every year or more.

Bury quoted Microsoft CEO Satya Nadella's previous statement: “I don't want to be locked in the massive depreciation of a generation of hardware.” The meaning behind this statement is intriguing — what Nadella is worried about is exactly what is happening on Microsoft's own balance sheet.

This mismatch had two levels of accounting consequences. Burry pointed out, first, that the five companies currently have a total of more than 400 billion US dollars in construction assets. As these assets have not been put into use, there is currently no depreciation under GAAP. “Nvidia (NVDA.US) chips may have depreciated economically while waiting in the warehouse, while GAAP depreciation is still zero.”

Second, even if the asset is already in use, the rationality of the depreciation assumption itself is highly controversial. Mainstream cloud vendors around the world generally set a depreciation period of 5 to 6 years for AI servers, but the actual economic life of GPUs — considering that Nvidia's iteration cycle was reduced from 18 to 24 months to about 12 months — is closer to 2 to 3 years in Bury's view. Bury previously estimated that by extending the GPU usage period to 5 to 6 years, hyperscale cloud vendors will reduce depreciation expenses by about 176 billion US dollars between 2026 and 2028.

One notable detail is that Amazon reduced the service life of some servers from 6 to 5 years in 2025 because a study found that the speed of technology iteration in the field of AI and machine learning exceeded previous expectations. This is the first case of the Big Five “going backwards” in terms of depreciation assumptions. To a certain extent, Amazon's adjustment confirms Bury's judgment that the depreciation cycle has been excessively lengthened.

Meta's approach is in the opposite direction. The company raised the server depreciation period from 4 to 5 years to 5.5 years. This alone increased book profit of about $2.9 billion that year, accounting for 4% of its pre-tax profit. Over the past five years, the other four giants have also used similar techniques to “optimize” profit data more than once.

Oracle: The one named

Of the five companies, Bury named Oracle in particular. His reason is straightforward: the analysts' average price target for Oracle is significantly higher than the current market price, which in his opinion is a sign of excessive optimism.

The data does support this observation. According to S&P Global's survey of 43 analysts, Oracle's consensus rating is “buy,” with an average target price of around $238. In some agencies' models, the target price is even as high as $420. Meanwhile, Burry himself began shorting Oracle at a price of around $145, and had previously held a put option on the stock.

Oracle's external exposure is indeed the most concentrated of the five. Its $260 billion unstarted lease commitment is equivalent to nearly seven times the confirmed lease liabilities, which is an extreme level for large tech companies. S&P Global Ratings has taken this factor into account in the evaluation. At the same time, Oracle's total debt is about 167 billion US dollars, and the capital structure is weakening, along with large-scale equity and debt financing. These commitments mainly refer to data centers, and the rental period is generally between 15 and 19 years.

The reason why Oracle's situation is worth reviewing separately is that its off-sheet promises are highly tied to OpenAI's computing power requirements. If OpenAI's commercialization process falls short of expectations, Oracle will face not only the problem of idle production capacity, but also the chain effects of contract termination costs, asset impairment, and rising financing costs.

What is Wall Street pricing?

Barry's criticism points not only to the accounting treatment of businesses, but also to Wall Street's analytical framework. His core argument is that many analysts do not fully include maintenance capital expenses and economic depreciation of AI infrastructure in their valuation models.

This is not an argument that can be easily refuted. The capital expenses of hyperscale cloud vendors are growing at an alarming rate. According to agency estimates, the total capital expenditure of the five companies in 2026 will reach 769.2 billion US dollars, almost double that of 2025, and may exceed 1 trillion US dollars in 2027. Microsoft alone is expected to have capital expenditure of up to 190 billion US dollars in 2026, an increase of about 130% over the previous year. The combined capital expenditure of Amazon, Alphabet, Microsoft, and Meta in 2026 is estimated to be between $700 billion and $760 billion.

The way these expenses are funded is also changing. Five companies raised more than $200 billion through the bond market this year, more than double that of 2025. Morgan Stanley predicts that the global AI-related debt issuance scale may be close to $570 billion in 2026. Amazon's free cash flow turned negative in the second quarter of 2026, while Alphabet experienced negative quarterly free cash flow for the first time in decades since listing. Morgan Stanley predicts that Amazon, Oracle, and Meta's free cash flow will approach zero or even turn negative in 2026.

But for now, the market is still selectively ignoring these signals. The Roundhill Magnificent Seven ETF (MAGS.US), which tracks major tech stocks, is moving in the direction of rising for the third straight month. Strong cloud business growth in the last quarter continues to support investors' optimism.

The focus of disagreement

Not everyone agreed with Bury's judgment. Fundstrat's Tom Lee publicly refuted Bury's warning that the current situation was likened to an Enron incident, arguing that the $3 trillion off-balance promise “hardly reflects actual risk for investors.” His argument is that such large-scale paper transactions operate normally in the financial sector, and are supported by high-tech profits.

The essence of this disagreement is: should an external promise be viewed as a “forward arrangement that has not yet been implemented” or “a real debt that has already been locked in”? Optimists believe that the fulfillment of these promises depends on continued growth in AI demand, and the current strong growth in the cloud business is verifying this premise. Pessimists, on the other hand, point out that once demand growth slows — even if only from “super speed” to “high speed” — these promises will quickly change from footnotes to real figures on the profit and loss statement in the form of asset depreciation, contract termination costs, and idle production capacity.

Burry's own language was relatively restrained in this round of warnings. He made it clear that he was not accusing him of fraud, but he gave a serious reminder about the downside risks of the current wave of AI construction when demand weakens. “When the music stops, these off-label promises quickly become real liabilities.”