Off the Books
The four biggest tech companies plan to spend about $725 billion on AI this year, and even they can't pay for it out of pocket. Amazon is moving chips off its balance sheet, Alphabet is selling stock and Meta is renting its own data centers back from investors. Who can afford the war says more about their moats than the spending does.
The decision
On October 2, the Financial Times reported that Amazon is in talks to move about $8 billion of Nvidia Grace Blackwell chips into a special-purpose vehicle. Outside investors would fund the vehicle, it would buy the chips, and it would rent them back to Amazon. The chips never leave AWS data centers. Only the owner, and the debt, change. Neither Amazon nor Nvidia has confirmed the talks.
$8 billion is small for a company guiding to $220 billion of capex this year. The structure is what matters. Amazon, one of the richest companies on earth, is looking for someone else to carry its chips.
It isn’t alone. On June 1, Alphabet announced an $80 billion stock sale and priced it at about $85 billion, including $10 billion from Berkshire Hathaway. It bought back zero shares in the first half of 2026, after buying back $28 billion in the same period of 2025. Meta financed its Hyperion data center in Louisiana through a separate vehicle with Blue Owl in late 2025, about $27 billion of debt kept off its own books, and it has since sold more data center debt.
The 4 biggest spenders guided in July to about $725 billion of capex in 2026, up 77% from 2025. Analysts expect more than $1 trillion in 2027. Goldman Sachs projects hyperscaler debt issuance of $420 billion next year.
The richest companies in history are now raising money like startups. That is the decision.
The strategic why
Each company chose a different way to pay, and each way says something about its moat.
Alphabet paid in the open. Selling stock is the most expensive way to raise money, because it dilutes every existing shareholder. It’s also the most honest. There’s no off-books vehicle and no lease to unwind later. Alphabet’s capex ran ahead of its operating cash flow in Q2, even with revenue up 24%. Its message to the market was plain: we would rather own every chip and every building outright, and we’ll pay for that with our shares. Google Cloud sells compute to outside customers, so a share of that spending comes back as revenue.
Amazon is moving the bill. A sale-leaseback keeps the chips running while moving the debt to investors who want exposure to AI hardware. It protects Amazon’s credit metrics, and that protects its ability to borrow cheaply for the next round. It’s also aimed at the one problem chips have that buildings don’t: they lose value fast. Nvidia’s next platform is already coming. Letting someone else own the depreciation is a smart trade. And AWS rents those chips to customers by the hour. The SPV collects rent from Amazon, and Amazon collects rent from everyone else.
Meta is renting its own castles. Meta has the same appetite as the others, with $115 billion to $135 billion of capex guided for 2026, but one big difference. It has no cloud business. Every GPU Meta buys serves Meta: its ads, its feeds, Muse. Amazon, Microsoft and Google can spread the cost of a data center across thousands of outside customers. Meta carries its alone, and when it finances off books, it pays investors a return on top. AI-linked debt is pricing at about 115 basis points over Treasuries, against 78 for the broader investment-grade market. That gap is the price of being a heavy borrower in a crowded market.
Here’s the pattern. Financing doesn’t create a moat. It shows you which moats are real. A company that sells compute can turn its bill into a product. A company that only uses compute has to turn its bill into future profits, and it pays interest while it waits.
The question isn’t who spends the most. It’s who gets paid back while they spend.
What it defends against
The credit ceiling. Even the largest companies have a limit on how much they can borrow before rating agencies react. Off-books vehicles and stock sales keep each company under that ceiling, so it can keep spending through 2027 when everyone expects the bill to pass $1 trillion.
Falling behind on compute. Every company here believes the one with the most capacity wins the next model, the next agent and the next customer. Running out of money first is the only outcome worse than overspending. Each financing structure is a way to make sure that isn’t them.
The depreciation trap. A GPU bought today is worth much less in 3 years. Leasebacks and vehicles move that risk onto investors who have chosen to take it, at least on paper.
The risk
Off-books doesn’t mean off the hook. Amazon will still owe the lease payments, Meta will still owe Hyperion rent and the rating agencies know where to look. One analyst estimate puts big tech’s AI-related obligations held outside balance sheets in the trillions. If AI revenue arrives more slowly than planned, these structures don’t shrink. They come due on schedule.
There’s also crowding. The same pool of private credit funds Meta’s vehicles, Oracle’s data centers, xAI’s debt, CoreWeave’s loans and Nvidia’s financing partners. When everyone borrows from the same lenders at the same time, the price rises for all of them. That rising price is the 115 basis points.
Finally, there’s the Alphabet lesson. Selling $85 billion of stock is a reminder that the most profitable companies in the world are now spending faster than they earn. If that’s true for Alphabet, it’s true for everyone. Investors have been patient so far.
That’s why Meta’s scale and cost score drops from 8 to 7. It isn’t spending less than its rivals. It’s paying more for every unit of compute, carrying the full cost alone and financing part of it at a premium. That moves Meta from 3.26 to 3.20. It’s still Deep.
Everyone is hiding part of the bill. Only some of them can send it to a customer.