Amazon is reportedly in talks to do something that would have sounded strange a year ago: sell roughly $8 billion worth of its most advanced Nvidia chips to outside investors, then rent them straight back. The Financial Times first described the discussions, Reuters recapped them, and both Amazon and Nvidia declined to comment. Nothing is signed, and the terms could still change. But the shape of the deal is a clear signal of where the AI buildout is heading, because it treats a pile of silicon the way airlines treat jets and landlords treat office towers.
The mechanics are a classic sale-and-leaseback. Amazon would move thousands of Nvidia Grace Blackwell chips, already installed across more than a dozen data centers in five US states, into a separate entity known as a special purpose vehicle. That vehicle would borrow money from investors, buy the chips, and lease them back to Amazon for long-term use. Amazon would reportedly keep no equity in the vehicle while offering outsiders up to a 10 percent stake. The point is to keep using the hardware without carrying its full cost on the balance sheet.
Why bother? Because the numbers have become enormous. Amazon has guided to around $220 billion in capital spending this year, up from an earlier $200 billion, with much of it flowing into AWS data centers. Chief executive Andy Jassy warned in July that even at that pace the company would not have enough capacity to meet demand in 2026 or 2027, and called the volume already booked for 2028 extraordinary. Financial engineering is how a company keeps spending at that scale without letting its books buckle.
What makes the Amazon structure revealing is who the bonds are pitched to. Investors are reportedly being told to lean on Amazon's double-A credit rating rather than the resale value of the chips, which is precisely what lets conservative money such as pension funds and insurers take part. In other words, the deal is sold as a bet on Amazon's promise to keep paying, not on the hardware itself. That distinction matters, because the hardware is the weak link.
Nvidia is working the same problem from the other end. The company is in early talks, alongside the broker Howden Re, to help insure lenders who finance smaller cloud firms against the risk that those borrowers default and the GPUs pledged as collateral do not fetch enough on resale to cover the debt. Nvidia has reportedly handed at least one insurer its own data on how fast its chips depreciate. Back in August it lined up six financial heavyweights, among them Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, to build a platform aimed at pulling in more than $500 billion of outside capital, and floated guaranteeing up to $125 billion of hardware residual value itself.
The tension sits in that word, residual. Aircraft and buildings hold value for decades, which is what makes securitising them sane. AI accelerators do not. Grace Blackwell is cutting edge today, but Nvidia's next-generation Vera Rubin is already coming, and Amazon has quietly trimmed the assumed useful life of some gear from six years to five. Every party here is quietly guessing how quickly a chip becomes a doorstop, and then building hundred-billion-dollar financial products on top of the guess. If the guesses are right, this is clever capital efficiency. If they are wrong, a lot of investment-grade paper is resting on hardware that ages like fruit.