The $3.1T question: who holds the risk when the music slows.
The most important number in AI right now isn't a benchmark. It's $3.1 trillion.
That's Morgan Stanley's estimate of off-balance-sheet AI commitments across nine tech firms: future data-center leases and chip purchase agreements sitting in GAAP footnotes, not on the balance sheet. About five times those companies' trailing-twelve-month capex. About three times their on-book debt and leases combined.
The AI trade stopped being a technology story a while ago. It's a financing story now: who funds the data centers, where the risk actually sits, and what breaks first.
The machine only has one direction.
The four big hyperscalers are on track for roughly $700–730 billion in combined 2026 capex, up 70–80% from last year, roughly triple 2024. Guidance keeps ratcheting up, not down. Goldman Sachs has the group approaching $1.14 trillion in 2027.
But the market is starting to ask about the return. In February, Microsoft fell 11% in a single day, its worst since March 2020, on slowing Azure growth paired with faster data-center spending. Good isn't good enough anymore; only scarce is.
The real innovation is the financing.
Because you can't spend a trillion dollars a year without getting creative about where it shows up. The structures are legal, disclosed, and invisible to every leverage metric the market watches:
- Anthropic closed a $35 billion private-credit package with Apollo and Blackstone to buy TPUs through an SPV, leased back to itself. Then, on October 1, Broadcom agreed to backstop up to $42 billion of Anthropic infrastructure leases, acting as both chip supplier and lender. Anthropic's own filing flags the conflict of interest.
- Oracle has ~$38 billion of similar structures tied to its $300 billion OpenAI partnership. xAI raised $20 billion through a comparable chip-collateralized vehicle.
- Alphabet's contractual obligations went from $332 billion to $811 billion in a single quarter.
The pattern: the asset and the debt live in the vehicle; the tech company contributes a lease and, often, a guarantee. Legal, disclosed, in the footnotes, invisible to every leverage metric the market watches.
The test case is CoreWeave.
Strip away the structures and look at one company in the open: $35–39 billion of guided 2026 capex against $12.4–13.2 billion of revenue. Three dollars of capex for every dollar billed. Trailing free cash flow: negative $13.65 billion.
The per-contract economics genuinely work. A 100MW cluster costs ~$3.8 billion, and a five-year take-or-pay contract at $1.2 billion a year repays it twice over even if the GPUs are worth zero at the end. The problem is one level up: new five-year debt facilities that outlive the average three-year customer contract. Lenders are underwriting renewal risk that has never been tested, with the top two customers accounting for 65% of revenue.
The equity market has noticed: the stock is ~55% below its highs, Rothschild initiated with a Sell on leverage, Kerrisdale is short. The credit market, meanwhile, just handed CoreWeave an $8.5 billion investment-grade GPU-backed facility at ~5.9%. Somebody's wrong about the renewal.
The wall isn't chips. It's power.
Texas paused new data-center interconnections pending an audit of its 474-gigawatt queue, more than five times the state's all-time peak demand. The grid operator's verdict: most of it is speculative. Wood Mackenzie estimates roughly two-thirds of requested US data-center electricity will never connect. PJM's interconnection wait averages eight years.
And then there's the neighbors. More than $170 billion of data-center capacity has been blocked, withdrawn, or stalled by community opposition since January 2024. New York now publishes a framework pricing community consent at $1 million per megawatt. A Pennsylvania developer is offering residents $10,000 per household to revive a rejected project. Community pushback isn't PR anymore. It's a capex line item.
The skeptics have math.
Michael Burry's September letter put ~$3 trillion in AI infrastructure obligations on the big cloud companies and argued they depreciate AI chips over four to six years while the true replacement cycle is two to three, understating depreciation by ~$176 billion through 2028, with write-downs landing "perhaps in 2028 or 2029." His anchor stat: S&P 500 net investment hit 2.07% of GDP last quarter, the highest in roughly forty years except the March 2000 peak.
The rebuttal ran seven pages. That's how seriously the argument is being taken.
Burry's math hinges entirely on the chip-life assumption: two to three years versus the four to six the cloud companies book. The other side points to six-year-old chips still running commercial workloads. Both can't be right about the earnings. But he's not alone: Chanos sees GPU-backed debt defaults coming, Kerrisdale is short the neocloud model, and lenders are voting with their term sheets. They still won't underwrite silicon as long-duration collateral without guarantees nobody will give.
The honest version of the fiber parallel.
1998–2001 laid over $500 billion of fiber on WorldCom's myth that internet traffic was "doubling every 100 days." The real number was closer to doubling once a year. By 2002, 85–95% of the fiber was dark, bandwidth prices had fallen 90%, and more than $2 trillion of telecom value had evaporated. The glut was mocked as "100 years of capacity." It was consumed in fifteen.
What rhymes: capital flooding toward extrapolated demand, capacity overshooting, vendor financing. Lucent and Nortel financed their customers' purchases then; chip suppliers co-finance buyers now, with Broadcom acting as both supplier and lender to Anthropic.
What breaks the parallel, and it's the whole ballgame: the spenders then were junk-rated new entrants funding speculative demand with high-yield debt. The spenders now are the most cash-generative incumbents on earth, funding largely from operating cash flow against real demand. The failure mode isn't bankruptcy. It's write-downs and multiple compression.
So where does it break?
Not a prediction. A watchlist. The 2027 contract renewals, when the first 12–18-month neocloud deals reprice and renewal risk gets tested for the first time. The depreciation fight, which decides whether $176 billion of earnings were real. Whether community tolls spread beyond New York. Whether lenders ever accept silicon as long-duration collateral. Right now, they won't without guarantees the chipmakers haven't agreed to give.
The buildout is real. The demand is real. The question was never whether AI needs the capacity. It's who holds the risk when the music slows. Right now, the answer is: not the balance sheet. That's not a fraud story. It's a story about where to look.
Top comments (0)