For two years, the artificial intelligence buildout was mostly an equity story, paid for with venture rounds and the cash piles of the largest technology companies. This week’s deal tape suggests the next chapter will be written in the credit markets.
On a single Monday, Bank of America, Blackstone, Citi and Morgan Stanley unveiled a $63 billion special purpose vehicle to buy Broadcom chips and lease them to Anthropic, and SpaceX disclosed a binding commitment for a finance lease with principal payments of about $40 billion on GPUs supplied through an Apollo-managed lessor. Together that is more than $100 billion of debt-style financing for hardware that has not yet been delivered.
Chips financed like airplanes
The structures will look familiar to anyone who has followed aircraft or rail car finance. A vehicle owns the equipment, the user pays rent, and lenders are repaid from that rent. SpaceX’s filing spells it out: quarterly rent equal to one-twentieth of each tranche’s cost plus fixed-rate interest, over about five years. The Anthropic vehicle splits the risk between first-lien bank debt and second-lien money from Blackstone.
The difference is the collateral. A jetliner can fly for decades. An AI accelerator may be economically obsolete in a few years as faster generations ship. That is why the most important line in the Broadcom announcement may be the least flashy one: Broadcom is providing residual value support equal to the first-lien portion of the financing. Senior lenders are, in effect, leaning on the chipmaker’s balance sheet as much as on the chips.
Where the risk goes
That shifts the question for investors. The credit quality of these transactions depends heavily on a handful of strong guarantors and on lessees with very large cash needs. Anthropic’s IPO filing, released the same day, shows rapid revenue growth and a deep cash cushion, but also how much compute it intends to keep buying. Moody’s has already flagged that Broadcom’s risk would rise if it piles up several such backstops in quick succession.
The model is spreading down market, too. Inference cloud startup General Compute, which announced a $100 million Series A this week, has struck a $400 million agreement to buy chips from SambaNova, billed as the first debt financing of inference chips. Smaller operators will not have Broadcom-sized guarantors behind them.
What to watch
Three things will tell us whether this is sound structured finance or a late-cycle stretch. First, pricing: SpaceX has not yet set the fixed rate on its lease, and the spread lenders demand will be the cleanest signal of perceived risk. Second, depreciation: watch how lessees and lessors account for chip lives in coming filings. Third, concentration: if the same banks and private credit funds keep showing up in every vehicle, a stumble at one AI developer could ripple further than the headline sizes suggest.
None of this means the money is misplaced. Demand for computing remains strong, and the borrowers are among the best-capitalized companies in technology. But when an industry starts financing its core assets like airplanes, it should expect to be judged like an airline: on cash flow, leverage and what the collateral is worth when the music slows.