Why Michael Burry Is Shorting NBIS

Disclaimer: This article is for informational and educational purposes only. It does not constitute financial, legal, or investment advice. Views are based on public information as of the publication date and are subject to change.

On August 3, Michael Burry shorted Nebius at $211.77. Nebius is a Dutch company that rents computing power to AI firms. On his Substack, he described the trade as "a bit like shooting fish in a barrel," meaning it was almost too easy. The day the news got out, the stock fell 13%.

Nine days later, Nebius reported second quarter earnings. The stock jumped more than 30%, running from $190 to above $250. The shorts got buried. But Burry just added more to his position.

If you actually follow what this argument is about, you will notice that the earnings report answered almost none of the questions Burry raised.

The Landlord Business

Nebius used to be Yandex, the Russian search giant. After the war brought sanctions, it divested its Russian assets in 2024, moved to Amsterdam, and repackaged itself.

The business model is simple. Secure land, find power, build data centers, fill them with NVIDIA cards, and rent out the compute by the megawatt. It does not design chips and it does not train models. It's a landlord, and its tenants happen to be AI labs.

Second quarter revenue came in at $582 million, up 454% year over year. Capital expenditure over the same period was $5.66 billion, close to ten times revenue. A landlord has to put up the building before collecting rent, and in this case the building is mostly a stack of graphics cards with a shelf life.

That is where the whole argument starts.

The Seller Who Won't Sell

The pricing disclosures were the real news.

Nebius signed four contracts in the quarter, each above a billion dollars, to customers including Reflection and Cohere. Rates ran $20 to $25 million per megawatt annually, with tenants prepaying more than half the construction cost up front.

Then it did something strange. It held capacity back from those long term deals, reserving a slice for contracts of six months or less that clear at $40 to $50 million per megawatt. CEO Arkady Volozh noted the price is “sometimes above” that range.

The company also ran an auction. Chief Revenue Officer Marc Boroditsky explained the thinking: "In a market where we have several buyers for every GPU, we let the market tell us directly." Multiple bidders per card, so stop guessing and take offers. The auction cleared 15% above any price the company had ever gotten.

The most important sentence was this one: “we could sell today our entire 2027 capacity on these terms if we wanted to. But we are not doing this.” They can sell out next year right now. They'd rather wait, because they believe the price will continue climbing.

That is a seller who would rather leave inventory on the shelf than lock in revenue, which is a very different posture from the memory giants rushing to sign LTAs. The market understood the implication immediately, and the stock took off.

What Dr Burry Actually Wrote

Almost every summary of his thesis reduced it to a claim that GPUs only last two or three years.

Let's dispose of that first. Graphics cards can run six or seven years without much trouble, and this isn't seriously contested. Did anyone really believe Burry hadn't figured that out?

Silicon is durable. A data center GPU fleet is a maintained population, not a single device you discard when it fails. Meta measured an annualized failure rate around 9%, meaning you replace cards on a rolling basis and the fleet never falls off a cliff. Bernstein’s channel work concluded that six to seven years of service life is entirely realistic.

Burry is aiming somewhere else entirely.

His actual claim was that NVIDIA’s chips sit on a two to three year product cycle, and that buying assets like that in volume should not be causing companies to lengthen their depreciation schedules. A product cycle measures how long a chip generation keeps its pricing power. Failure rate measures when the hardware dies. They have almost nothing to do with each other.

His words: “Understating depreciation by extending useful life of assets artificially boosts earnings, one of the more common frauds of the modern era.”

The evidence he assembled is a set of before and after comparisons. Meta stretched server depreciation from 3 years in 2020 to 5.5 years by 2025. Oracle went from 5 to 6. Across five hyperscalers he tallies roughly $176 billion of depreciation expense pushed into the future between 2026 and 2028.

The question underneath all of it is uncomfortable and hard to dismiss:

In the fastest hardware turnover cycle any of us have lived through, why is everyone deciding their equipment lasts longer?

Microsoft’s own CEO answered it out loud: “I don’t want to take on a 4-5 year depreciation burden on one generation.” Meanwhile the industry moved server lives from three or four years to six, which quietly removes about $18 billion a year from the depreciation line.

The Rent Is Falling Off a Cliff

Forget when the card stops working. Ask what the card still earns.

That number is public, and it isn’t pretty.

The H100 was the most fought over piece of hardware of 2023. At peak scarcity it rented for roughly $8 an hour. By June 2026, on demand rates sat between $2.01 and $2.53, essentially back to where the industry started before any of this began.

The A100 has fared worse. Cards trade around $18,900. Median rental is $1.90 an hour, the discount providers will hand you one for $0.89, and Azure’s interruptible instances have sunk to $0.68.

Thirty months. Roughly 70% of the rent gone. Call it a third per year, compounding.

This is Burry’s case, and it’s a case about revenue rather than reliability. The card still boots, still takes jobs, still has customers. It simply earns a fraction of what it once did. Extending the depreciation schedule is precisely the accounting choice that keeps that decay from ever showing up in reported earnings.

One inconvenient fact for this view: H100 annual contract pricing went from $1.70 an hour in October 2025 to $2.35 by March 2026, up 40%. Spot is cratering while contracts are firming.

The two aren’t in conflict. What’s changed is that compute has moved from something you rent when you need it to something you reserve in advance. Nebius’s record auction is the same phenomenon viewed from the other end. The customer paying a premium to guarantee capacity next year and the customer hunting bargains on the spot market were never competing for the same machine.

The Answer Is 51%

Since the cards turn over every few years, the business has to recover the full cost of replacing them out of operating cash flow inside that window. Fail that test and you don’t own a company. You own a machine that converts capital into nothing.

So if we assume no growth (including from external financing) and strip capital spending down to replacement only, and ask what margin Nebius needs to sustain itself unaided.

The answer moves with useful life, and the asset base isn’t uniform. GPUs and servers are 75% to 80% of it and cycle every six years. Networking gear runs eight. The shell, the substation, the cooling plant all last twenty five years, and new cards drop into old buildings. Blend those together and you get:

Six year cards means Nebius needs a 51% EBITDA margin to stay level. Not to create value, to breakeven.

The EBITDA margin that drove the stock up 34% was 40.6%.

And there's the joke at the center of this whole episode. The market threw a party over a margin that beat estimates by eleven points and still sits below the level required to sustain the business. Worse, EBITDA excludes depreciation by definition. The D stands for Depreciation.

Investors reached for the one metric that removes the disputed item and used it to settle the dispute.

Can margins get there? Consensus says 62% by 2028, and the most bullish analyst on the name says 69%. Both clear the bar comfortably, and if either proves right this is a genuinely good business. But if Burry's four year estimate holds, the bar jumps to 71%, and no realistic path reaches it.

So Who’s Right?

The useful life debate is finished, and Dr Burry won the half that counts. He never argued the cards would break. He argued they’d get cheap, and the rental market has been proving him right for two years.

He’s also less dangerous than the bulls seem to fear. Even with four year lives and rents falling 30% annually, this business still generates cash. It just fails to earn its cost of capital. That’s a mediocre investment, not a wipeout.

A wipeout takes two things happening together: residual values collapsing at the same moment the financing window closes. My read is that’s the tail he’s actually betting on.

Worth remembering too that the trade was crowded before he ever showed up. Short interest hit 27% of the float going into the print, which explains most of how one decent quarter produced a 34% single day move.

If you keep one thing from all of this, remember this figure: 51%.

Whether Nebius closes the gap to that line or drifts away from it over the next few quarters will tell you more than revenue growth, contracted backlog, or gigawatts under agreement ever could. Management has already put a date on it.

From the CFO: “expect capacity coming online from our own data center to begin improving margins in the second half of next year.”

A final note on the accounting. Under US GAAP, useful life is management’s own estimate. There is no industry standard to violate, and depreciation is explicitly defined as allocation rather than valuation. Dr Burry is never going to prove fraud in a courtroom on these facts. He also doesn’t need to. It just means the arithmetic is yours to do, because nobody is going to correct that number on your behalf.

Subscribe now

Refer a friend