Michael Burry Just Shorted the AI Cloud — and Today Is the Test

Michael Burry Just Shorted the AI Cloud — and Today Is the Test

Let me tell you something that's been sitting heavy with me since the options data crossed my desk this morning. A company that grew revenue 684 percent in a single quarter is getting shorted by the man who called the 2008 housing crash. Not lightly shorted, not hedged with cheap puts — shorted outright, with his own money, because he thinks the trade is that easy. And the company in the crosshairs reports earnings today.

I've been running hosting infrastructure for over a decade. I've watched this AI buildout go through its GPU shortage phase, its power shortage phase, its community backlash phase, its debt phase. This morning it enters a new phase: the accounting phase. When Michael Burry says "a bit like shooting fish in a barrel" about a stock that's up 130 percent this year, you don't shrug that off. You go read the balance sheet. I did. And what I found is the most interesting earnings setup of this entire cycle.

The Setup — Everything Happened Before the Bell

Here's where we are. Nebius Group — ticker NBIS, the Amsterdam-based "AI cloud" company that rents out AI compute, trains models, and runs inference — reports its second quarter this morning before the US market opens. Consensus is roughly $575 million in revenue and a loss of about 52 cents a share, according to Zacks. That's a 442 percent revenue jump from a year ago. Sounds incredible. It is incredible. And it's not the story.

The story is the guy on the other side. On August 6, in a post on his Substack Cassandra Unchained, Michael Burry disclosed fresh short positions in Nebius at $211.77 a share and Oracle at $144.63. He described the trades as "a bit like shooting fish in a barrel." For Nebius he took a direct equity short instead of options because put implied volatility was above 100 percent — in his words, "exorbitantly priced." He'd already closed a previous Oracle put position on August 5 for what he called a substantial profit, then re-entered the trade as conditions shifted.

Burry's core argument is worth taking seriously because it's not about demand. He's not saying nobody wants AI compute. He's saying the companies selling it are overstating their earnings by depreciating expensive GPUs too slowly — he estimated Oracle's profits could be overstated by as much as 26 to 27 percent by 2028 — and that long-term lease obligations that dwarf current revenue make these companies "very fat, very large, easy to shoot" fish. That's not a market call. That's an accounting allegation. And it lands right on the neocloud business model.

The Bridge — the Number Nobody's Adding Up

Let me give you the single most under-discussed number in the AI infrastructure complex right now. Nebius exited the first quarter with an annualized run-rate of roughly $1.92 billion in revenue. Management guides to $7 billion to $9 billion of ARR by the end of this year. That is not a growth rate — that is a bridge with a $5.1 billion to $7.1 billion gap in the middle and three quarters to cross it.

Do the arithmetic. To hit the $7 billion low end, Nebius must add roughly $1.69 billion of ARR in each of the second, third, and fourth quarters. To hit the $9 billion top end, it needs about $2.36 billion a quarter. Read that against the base: the company has to add close to 90 percent of its entire existing run-rate, every quarter, three quarters running, just to hit the bottom of its own guidance. Meanwhile consensus has second-quarter revenue near $575 million, which annualizes to about $2.3 billion. A perfectly respectable quarter still leaves almost the entire bridge to be built in the second half of the year.

This is the clearest reason the stock round-tripped from $286.69 in June to $190.28 on Monday while the sell-side average target sits above $258 with a consensus Buy rating. The market is not disputing the demand. It is discounting the schedule.

Reading One — the Demand Story Is Real

Now let me give the bulls their due, because they're not wrong about demand. Nebius is not a shell leasing capacity inside somebody else's building. It is building and owning its own sites — CEO Arkady Volozh said on the Q1 call that owned contracted capacity now accounts for more than 75 percent of total power. The company announced its second owned gigawatt-scale US site in Pennsylvania this quarter, supporting 1.2 GW at full buildout, and expects more than 4 GW of contracted power by year-end, up from a prior forecast of more than 3 GW. It broke ground on a gigawatt-scale AI factory in Independence, Missouri in May.

The demand behind all that concrete is contracted, not hoped for. Meta signed a long-term agreement for up to $27 billion in computing capacity over five years. Microsoft is taking tranches of capacity — Nebius says it recently delivered the latest planned tranche and remains on track for the rest. Nvidia invested $2 billion in March as part of a strategic partnership covering AI factory design and inference infrastructure, and it took a 9.3 percent stake. Q1 revenue beat analyst estimates by about 8 percent, adjusted EBITDA was positive at $129.5 million, and management's own words on the call were blunt: "Everything we build, we sell, and we are still in the very early days."

And the demand story is not isolated to Nebius. CoreWeave — the other big neocloud name — reported a record quarter yesterday: $2.58 billion in revenue, up 112 percent year over year, with a revenue backlog that surged to $104 billion. Its stock jumped 14 percent. Two neoclouds, back to back, doubling revenue in the same week. That is not a demand problem. That is the strongest demand signal in the entire tech sector.

Reading Two — the Balance Sheet Is Screaming

Here's the other half of the tape. To build all that owned capacity, Nebius is spending money like a country at war. First-quarter capex came in at approximately $2.5 billion, up from $544 million a year earlier — a fourfold increase in a single year. Management raised full-year capex guidance from $16-20 billion to $20-25 billion, against full-year revenue guidance of just $3.0-3.4 billion. Let me say that plainly: the company plans to spend roughly seven times its expected revenue this year, funded almost entirely with borrowed money.

Long-term debt more than doubled to $8.4 billion. On top of that: $4.3 billion in convertible senior notes from March, and in July a $775 million senior secured debt facility — its first — priced at SOFR plus 2.50 percent, backed by deployed GPU infrastructure and contracted cash flows. The deal was oversubscribed and the company calls it a "repeatable financing framework" for more than $40 billion of additional contracted revenue from investment-grade customers. Translation: the debt machine is designed to keep running.

Now overlay Burry's depreciation argument and the picture sharpens. When a company spends seven times revenue on assets and stretches the useful life of those assets, today's earnings are partly a choice, not a fact. Short interest is roughly 27 percent of the float — one of the heaviest short positions in the entire AI complex. And the options market is pricing a move of nearly 16 percent around this morning's print, with puts costing 13.6 percent more than calls — traders are paying up disproportionately for downside protection. Implied volatility sits at 120 percent, the 89th percentile of the past year. That's not a market expecting a calm readout.

The Secondary Bottleneck — September 11, Not August 12, Is the Real Date

Here's the piece of this story that almost nobody is talking about, and it's the reason the options skew is rational. Nvidia's 9.3 percent stake in Nebius is not a simple block of shares. The Schedule 13G filed July 13 shows 22,256,412 shares total, of which only about 1.19 million are directly owned. The remaining 21.07 million sit underneath a pre-funded warrant acquired on March 11. And Nvidia cannot sell any of it before September 11 — the lock-up expiry, 30 days after this earnings print.

So the market is being asked to absorb two events inside a month: a print this morning, and the release of a potential supply overhang on September 11. The bull reading is that Nvidia has every strategic reason not to sell — Nebius is a customer, a partner, and a showcase for Nvidia silicon, and dumping the position would signal doubt about the exact demand Nvidia is selling into. The bear reading is that a 21-million-share warrant is a liquidity event waiting for a window, and the mere possibility caps the stock into September no matter what this morning brings. Both are true simultaneously. That's why the options are expensive.

What This Means for Independent Hosting Providers

First — read the ARR bridge as your demand canary. If Nebius prints a number that puts a credible first span across that bridge — say, revenue well above the $575 million consensus with raised guidance — the neocloud demand thesis stays intact and GPU pricing stays firm. If it prints the number and walks back the bridge, start modeling what a softer compute market does to your own pricing in the fourth quarter.

Second — watch September 11, not just today. If Nvidia's warrant shares hit the market, expect GPU and hardware secondary-market pressure that ripples into every colo operator and hosting provider who bought at the top of the cycle. A 21-million-share overhang is a supply signal for hardware, not just a stock chart.

Third — treat "adjusted" earnings from capital-heavy AI clouds as a marketing number until the depreciation schedule is on the table. Burry's argument — that stretched useful lives inflate current profits — is the single best lens for evaluating any provider that spends seven times revenue on equipment. If you're comparing a neocloud's pricing against your own margins, compare against their cash cost of serving, not their adjusted EBITDA.

Fourth — this is your moment to be the capital-light, contracted, real-margin alternative. The AI clouds are trading growth for debt and hoping the bridge gets crossed. Your advantage is that you don't have to. Lock in multi-year contracts with customers who need reliable capacity, keep your utilization honest, and let the giants fight over whose balance sheet bends first.

The Bottom Line

Here's what I actually believe after reading every number I could get my hands on this morning. The demand for AI compute is real — CoreWeave and Nebius both doubling revenue in the same week proves that. But demand being real and the business being priced right are two different sentences, and the market has spent the last six weeks learning that lesson the hard way. The stock went from $286 to $190 while analysts kept their Buy ratings and their $258 targets. The schedule, not the story, is what's being discounted.

Michael Burry has been early before, and he's been wrong before. But he called the last housing crash by reading the balance sheets when everyone else was reading the press releases. Today, the AI cloud's biggest pure-play growth story answers for its balance sheet in public. However it lands, the lesson for every independent operator is the same one I've been preaching for months: the buildout is real, the financing is fragile, and the guys who own their assets outright — not the guys renting the future on margin — are the ones who survive the bridge. Watch the print. Watch September 11. And keep your own house boring.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: Nebius Q1 2026 results (May 13), Nebius senior secured debt release (July 17), FinanceFeeds, Foreign Policy Journal, Computeforecast, Zacks consensus, CNBC/Investing.com CoreWeave coverage (August 11).