원문 보기
The man who called the 2008 crash has a new warning. And it isn't that AI stocks are too expensive. Michael Burry just pointed his readers to a diagram from the credit giant Ares mapping $573 months. Bonds, leases, GPU-backed loans, project debt. And Burry's point is brutally simple. Nearly all of it depends on one thing continuing.
Not AI getting smarter. AI spending never slowing down. Tonight, how the AI boom quietly became a leveraged bet, the deals at the center of the web, and the four numbers that will tell us whether Burry is right. Forget the spaghetti. The Ares map asks one question. Who ultimately has to keep spending money for all these obligations to work?
On one side, 26 disclosed financings, corporate bonds, construction loans, data center leases, GPU-backed borrowing, and guarantees. On the other side, the same names appear again and again. Meta, Oracle, Microsoft, Amazon, OpenAI, Anthropic, Google, Broadcom, Nvidia, with CoreWeave as a key infrastructure node.
The lines mean someone has to pay, someone guarantees the payment, or two giants are commercially joined at the hip. And the whole diagram sits under one sentence. Everything depends on sustained AI capital spending. Ares' own warning. If that spending stops, pressure propagates through every linked obligation, and guarantees get called at exactly the moment the guarantors are weakest.
They compared it to vendor financing in the 2000 telecom bust. And to be clear, none of this is hidden or illegal. It's all disclosed. Burry's argument is narrower and stronger. If you only look at ordinary corporate debt, you're undercounting the real commitment. Leases, purchase obligations, guarantees, joint ventures, the balance sheet misses part of the picture.
The easiest example, Open AI needs staggering amounts of compute. Oracle agreed to provide it. A cloud contract worth more than $300 over 5 years, roughly 4 and 1/2 gigawatts of capacity. Follow the chain. Open AI's demand lets Oracle sign enormous commitments. Oracle builds data centers, buys Nvidia chips, and raises debt to fund it all.
Bond investors and insurers finance that debt. So, the asset on paper is Oracle infrastructure debt, but the economics underneath ask a different question. Will Open AI keep paying for enormous compute for years? And here's the uncomfortable part. Open AI doesn't need to go bankrupt for trouble to start.
Imagine growth merely disappoints. Instead of 10 gigawatts, they need seven. Oracle already holds the land, the power agreements, the construction contracts, the GPUs, the financing. None of that disappears. A slowdown in expected demand is enough. That is textbook capital cycle behavior. Now, the loop Burry finds most interesting.
Coreweave buys enormous quantities of Nvidia GPUs. Nvidia books the revenue. But, Nvidia also invested $2 billion directly into Coreweave, and separately agreed to purchase up to $6.3 billion of Coreweave's unused compute capacity through 2032. Conditional, subject to delivery terms. Disclosed, legal, but Burry's question is the right one.
How much of Nvidia's demand is genuine end-user demand, and how much is demand amplified by financing, commitments, and backstops? Then, there's the loan. Coreweave recently took a $2.6 billion facility backed by its GPUs and infrastructure. The loan runs about 5 years. The customer contracts backing it average about three.
Coreweave's own announcement says lenders are underwriting renewal risk, a bet that someone keeps renting those GPUs after today's contracts expire. And GPUs are not real estate. An Ares chart Burry cited models H100s losing roughly half their residual value over 3 years. New generations arrive, old silicon discounts fast.
If the collateral is worth less than the lender assumed, the math behind every GPU backed loan in the web gets worse at once. Microsoft plays it differently. It doesn't have to own every data center. It signs a long-term compute contract, someone else builds the facility, and that builder uses Microsoft's promise to raise financing.
Nvidia signed roughly $17.4 billion with Microsoft through 2031. And Nvidia told investors in its own filing that it expects to fund the buildout with, quote, debt secured against the contract at terms reflecting the credit quality of the counterparty. Read that again. Borrowing against revenue that does not exist yet, priced on Microsoft's creditworthiness.
Same pattern at IREN, about 9.7 billion for Microsoft for GPU Cloud, a $5.8 billion equipment bill with Microsoft prepaying roughly 20% of each tranche. The bank underwriting the Nvidia loan isn't betting on Nvidia, it's betting on Microsoft's promise. Now, scale that across Microsoft, Meta, Google, Amazon, Oracle, and dozens of builders.
Hyperscaler contracts become the foundation of an enormous credit ecosystem. That is what the Ares diagram exposes. Meta is the most interesting test of Burry's thesis because it might prove him wrong on the stock even if he's right on the cycle. Meta's Hyperion project, an 80/20 joint venture with Blue Owl funds.
The venture's vehicle issued about $27.3 billion of senior secured debt rated A+ by S&P. And that bond now trades around 94 cents on the dollar. Meta's own 10-Q discloses $28 billion in residual value guarantees with maximum exposure near 46 billion. Again, disclosed, not hidden. But it shows how much commitment sits outside the ordinary debt investors usually track.
Here's the thing though. Meta has Facebook, Instagram, WhatsApp, an advertising machine printing cash. Burry scenario doesn't need Meta to go bankrupt. The real question is return. Does AI driven advertising and commerce generate enough incremental profit to justify the CapEx? If yes, the spending is brilliant.
If not, it's the cycle's most expensive lesson. At Meta, watch incremental AI revenue against incremental AI CapEx. That ratio decides everything. And this is where Burry sees a system, not a stock call. If AI revenue growth disappoints relative to spending, watch the loop. Hyperscalers trim CapEx growth.
Utilization softens. GPU rental and resale values fall. GPU backed collateral is worth less. Lenders tighten. New projects can't get financed. Fewer GPUs get ordered. Nvidia and Broadcom growth falls. Multiples compress. Hyperscalers cut deeper. Notice what isn't required. AI failing. AI can keep growing enormously.
Capital spending merely has to overshoot eventual demand. That is a normal slowdown turning into a capital cycle bust and supposedly diversified credit exposures becoming correlated all at once. Burry keeps coming back to telecom in 2000 and the lesson is precise. The internet wasn't a fraud. It became one of the most important technologies in history.
The mistake was too much capital chasing it at the same time. Fiber got built everywhere. Equipment makers financed their own customers. Investors assumed exponential traffic justified any capacity. Then capacity outran monetization. Prices collapsed. Financing vanished. The internet kept growing. Technology right.
Investment cycle wrong. That is exactly the analogy Burry is drawing with AI. Now the honest counter argument, because Burry still needs one thing for this to activate. Monetization is real and it's fast. Anthropic's annualized revenue went from about $9 billion at the end of 2025 to 30 billion by April 2026 by the company's own numbers.
The hyperscalers generate enormous cash from businesses that have nothing to do with AI. This is nothing like financing pre-revenue dot-coms in 1999. So the bear case has a single trigger condition. AI revenue has to disappoint relative to AI capital spending. Not fail, just grow slower than the infrastructure Burry assumed.
Without that, the web holds, the debt gets serviced, and Burry's diagram stays a curiosity. Which is why the trigger matters more than the valuation right now. Four numbers to watch: AI revenue growth, hyperscaler capex guidance, GPU rental and resale values, and AI credit spreads. That last one is already moving.
AI-linked bonds now yield about 115 basis points over Treasuries versus 78 for the broad market. Investors charging more for AI debt. With roughly $420 billion of hyperscaler issuance expected next year. Not a crisis, but exactly the variable Burry would watch. If all four turn together, the diagram stops being an interesting theory and starts becoming a tradeable capital cycle reversal.
Burry isn't betting that AI fails. He's betting the financing cycle overshot, and cycles always collect. If this breakdown helped, subscribe. New videos three times a week. AI stock ideas, earnings breakdowns, and market trends. No hype, just the numbers that matter. Educational content only, not financial advice. See you in the next one. Mhm.
Comments 0