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

An investor betting against a stock is not, by itself, news. Thousands of wagers on where a price is headed open every day, and almost none of them matter.

What made Michael Burry's August 6 note in his newsletter "Cassandra Unchained" worth reading was not the bet but the reason behind it. Burry disclosed short positions in two companies: Oracle at $144.63 and Nebius at $211.77. His claim was not that the shares are expensive. His claim was that the profits these AI infrastructure builders report look larger than they are because of a single assumption buried in the footnotes.

That assumption is the useful life used for depreciation. A company estimates for itself how many years a server, or a graphics processor, will keep earning money. Stretch the estimate and this year's expense shrinks — and this year's profit grows.

Two of those figures are prices. Two are assertions. The gap between them is the subject of this piece.

Two Positions Facing Opposite Directions

The pair Burry picked 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 twelve months and roughly tenfold over three years.

Betting against a stock that has already collapsed and one that has gone vertical only makes sense if something links them. For Burry, the link is that both businesses do the same thing: buy billions of dollars of hardware, rent it out, and assume it will keep earning for a long time.

In the note, he described the trades as being "a bit like shooting fish in a barrel." He also explained why he shorted Nebius shares outright rather than buying puts: implied volatility on the puts had pushed above 100%, making the hedge too expensive. On Oracle, he closed an earlier put position at a substantial profit on August 5 and re-established the bet in the stock the next day.

Depreciation: Accounting's Quietest Choice

When a company buys $10 billion of servers, it does not expense that money in one year. It estimates how many years the hardware will be used and spreads the cost across them. The cash leaves on day one; the expense arrives in slices.

That is why useful life is a decision that directly sets reported profit. And because it is a decision, it can be revised.

Seeing the three schedules side by side explains why the argument has become so heated.

Not a Hypothetical — a Practice With a Record

What Burry is pointing at has already happened repeatedly, and the companies disclosed the effect themselves.

Alphabet extended server life from four years to six in early 2023; by its own account the change cut 2023 depreciation by $3.9 billion and lifted earnings by roughly $3 billion. Amazon went from four years to five in the fourth quarter of 2022, reducing that year's expense by $3.6 billion, then from five to six a year later, expecting to add $3.1 billion to 2024 operating income. Meta has moved from three years to five and a half since 2020 in a series of steps.

The other half of the record matters just as much. In the first quarter of 2025, Amazon shortened the life of a subset of its servers back from six years to five, citing the accelerating pace of technology — specifically in artificial intelligence. One of the companies that stretched the assumption conceded that it may have stretched too far.

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

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

Nebius shows how growth and risk come from the same source. First-quarter revenue reached $399 million, up 684% year over year. Annualized recurring revenue hit $1.9 billion. Adjusted EBITDA turned positive at $129.5 million.

On the other side of the balance sheet, long-term debt more than doubled to $8.4 billion. The company raised over $6 billion in fresh capital, $2 billion of it from Nvidia.

A company carrying $1.9 billion of annualized revenue spending up to $25 billion on hardware and facilities in a single year only makes sense if that hardware earns for many years. It does have contracts: a five-year agreement with Meta signed in March, worth at least $12 billion and up to $27 billion, and combined with the Microsoft deal, roughly $46 billion announced in total. But contracted revenue arrives across years while the spending happens now. Debt bridges the gap — one of the clearest illustrations of why cash flow differs from profit.

Oracle: Half the Order Book Sits With One Customer

Oracle is fragile in a different place. On July 9, S&P cut its credit rating to BBB-, one notch above speculative grade. At the top of the reasoning was the pace of AI infrastructure spending: the agency had modeled about $60 billion of fiscal 2027 capital expenditure and had to revise after the company guided to $90–95 billion. S&P expects a free cash flow deficit of roughly $42 billion in that year.

The size of that backlog was the centerpiece of the bull case for a long time. The same number became a headline risk for a rating agency once half of it traced to a single customer. Oracle has responded by completing a $5 billion convertible preferred issue and announcing plans for a $20 billion equity raise.

The Invisible Half of the Debt

Beyond individual companies, there is a wider floor under this argument. According to a Nikkei study cited by Fortune, off-balance-sheet obligations at the five largest U.S. technology companies have grown eightfold in four years to $1.65 trillion. Moody's puts off-balance-sheet liabilities at $1.2 trillion, of which more than $820 billion is data center construction commitments. Debt actually visible on the balance sheets is $1.35 trillion.

The bond market has started to price it. Per Penn Mutual's data, investment-grade bonds from the big cloud companies traded 20–25 basis points wide of the index in early 2025 and about 35 basis points wide by July 2026 — after gapping out to nearly 50 basis points earlier in the year.

The chart is not evidence for a claim. It shows the price level at which this argument is being conducted.

Timeline

The Other Side

There is no settled answer here, and the counterarguments are not weak ones.

Burry's own record belongs in the table too. In the same note he disclosed that he is holding his Palantir short through a 30% rally in the stock. He has previously run similar positions against Micron, Caterpillar and the semiconductor index.

What Is Left

On August 7 the market was looking at something else entirely: the U.S. 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; Oracle gained 2.4%.

The chart below shows the general condition of the ground on which this argument is taking place.

So after the first weekend, Burry is ahead on Nebius and behind on Oracle. A thesis like this is not settled in weeks; the next checkpoint is Nebius's report on August 12.

For a reader, the point is not who turns out to be right. It is that a company's reported profit can hinge this heavily on an assumption sitting in a footnote that 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.