Burry Shorted Oracle and Nebius: $176 Billion of Costs Deferred

What made Michael Burry's August 6 note in "Cassandra Unchained" worth reading was not the bet but the reason. He disclosed shorts in Oracle at $144.63 and Nebius at $211.77.

His claim: these AI infrastructure builders report larger profits than they have, because of one assumption buried in the footnotes.

That assumption is useful life: the company itself estimates how many years a server keeps earning. Stretch the estimate and this year's expense shrinks, this year's profit grows.

Two Positions Facing Opposite Directions

The pair looks contradictory. Oracle has lost 43% of its value over the past year and is down more than 25% in 2026. Nebius is up 204% over the same months and roughly tenfold over three years.

Depreciation: Accounting's Quietest Choice

A company buying $10 billion of servers does not expense it in one year; it spreads the cost over the years it expects to use them. The cash leaves on day one, the expense in slices.

In early 2023 Alphabet extended server life from four years to six. By its own account the change cut 2023 depreciation by $3.9 billion and lifted earnings by roughly $3 billion.

In the fourth quarter of 2022 Amazon went from four years to five, cutting that year's expense by $3.6 billion. A year later it went from five to six, expecting to add $3.1 billion to 2024 operating income. Since 2020 Meta has moved from three years to five and a half.

The other half matters as much. In the first quarter of 2025 Amazon shortened the life of some servers back from six years to five, citing the accelerating pace of technology in AI.

Burry's thesis is that this second move spreads: real economic life is two to three years, and the industry is deferring roughly $176 billion of expense between 2026 and 2028.

Nebius: $1.9 Billion of Revenue, $25 Billion of Spending

At Nebius, first-quarter revenue reached $399 million, up 684%; annualized recurring revenue hit $1.9 billion and adjusted EBITDA turned positive at $129.5 million.

Long-term debt more than doubled to $8.4 billion; the company raised over $6 billion, $2 billion of it from Nvidia.

Spending up to $25 billion in a year on $1.9 billion of annualized revenue only makes sense if the hardware earns for years. There are contracts: a five-year Meta agreement from March worth at least $12 billion and up to $27 billion, with the Microsoft deal roughly $46 billion in total.

But contracted revenue arrives across years while the spending happens now, and debt bridges the gap — an illustration of why cash flow differs from profit.

Oracle: Half the Order Book Sits With One Customer

Oracle is fragile elsewhere: on July 9 S&P cut its rating to BBB-, one notch above speculative grade, citing the pace of AI spending. The agency had modeled about $60 billion of fiscal 2027 capex, revised after guidance of $90–95 billion, and expects a free cash flow deficit of roughly $42 billion.

The backlog was long the centerpiece of the bull case, and became a headline risk once half of it traced to one customer. Oracle has completed a $5 billion convertible preferred issue and announced a $20 billion equity raise.

The Invisible Half of the Debt

On a Nikkei study cited by Fortune, off-balance-sheet obligations at the five largest American technology companies have grown eightfold in four years to $1.65 trillion. Moody's puts them at $1.2 trillion, of which more than $820 billion is data center construction, against $1.35 trillion of visible debt.

Per Penn Mutual, investment-grade bonds from the big cloud companies traded 20–25 basis points wide of the index in early 2025 and about 35 by July 2026. They had gapped to nearly 50 earlier in the year.

The Other Side

The same note disclosed a Palantir short held through a 30% rally; earlier ones targeted Micron, Caterpillar and the semiconductor index.

What Is Left

On August 7 the market was looking elsewhere: the American economy shed 23,000 jobs in July against expectations of an 80,000 gain. As rate-hike expectations faded the Nasdaq rose 1.3%; Nebius fell to $187.44 that day and Oracle gained 2.4%.

The point is not who turns out to be right, but that reported profit can hinge this heavily on a footnote assumption almost nobody asks about.

This article draws on public sources as of the publication date: the positions Burry disclosed in his own newsletter, S&P Global's rating action on Oracle, the useful-life changes companies disclosed in their own financial footnotes, compilations by Nikkei and Moody's, and Penn Mutual's bond spread data. Projections about future corporate earnings are assertions, not facts. Not investment advice.

Correction (September 13, 2026): prose tightened for rhythm; figures and sources unchanged.