Published: October 3, 2026 · Last updated: October 3, 2026
TL;DR: Amazon is reportedly exploring a deal that would shift roughly $8 billion of Nvidia’s Grace Blackwell AI chips onto outside investors through a financing vehicle, while AWS keeps running the same hardware under a lease.
Eight billion dollars worth of the world’s most sought-after AI chips, and Amazon reportedly doesn’t want to own them.
According to the Financial Times, Amazon is exploring a deal that would move roughly $8 billion of Nvidia’s advanced Grace Blackwell chips off its own balance sheet and onto outside investors’ books, through a special-purpose financing vehicle built for exactly this purpose. The structure is simple enough once you see it: investors buy the hardware, Amazon leases it back, and AWS keeps using the same processors in US data centers it already runs. Amazon gets the compute. Someone else carries the asset.
Nothing is signed yet. The report describes the arrangement as under discussion, not a completed transaction, and Amazon hasn’t confirmed specific terms publicly. But the fact that this kind of structure is even on the table says something about how AI infrastructure spending has outgrown the normal capital-expenditure playbook.
Amazon expects to spend roughly $220 billion this year, a huge share of it going toward cloud infrastructure, chips, and the data centers that house them. That number alone explains the appeal of off-balance-sheet financing. Buying chips outright means absorbing their depreciation as AI hardware gets replaced on faster and faster cycles. Leasing them through a financing vehicle lets Amazon report the expense differently, spread the risk across investors who want exposure to AI infrastructure without running a cloud business themselves, and keep its own books looking lighter even as its actual compute footprint keeps growing.
This isn’t a brand new idea in tech finance, but it is new at this scale and with this specific class of asset. GPUs depreciate fast, both financially and functionally, since a chip that’s state of the art this year is a discount rack item once the next generation ships. A financing vehicle that owns the hardware is making a bet that the lease payments over the chip’s useful life will beat whatever it’s worth on the used market once Amazon wants the newer thing. That’s a specific, calculated wager on AI demand staying strong enough that someone, somewhere, will always want to lease last year’s Blackwell chips.
It also puts Amazon in interesting company. Microsoft and Google are running their own versions of this capital math, each spending tens of billions of dollars a quarter on AI infrastructure while trying to keep investors comfortable with the pace. Meta disclosed a 2026 capex range of $130 billion to $145 billion just this year. None of the major cloud providers are slowing down, and all of them are getting more creative about how that spending shows up on a balance sheet instead of just how much of it there is.
The risk sits with whoever ends up owning the chips if AI demand cools even slightly. A lease only works if the lessee keeps paying, and the hardware only holds value if there’s a line of other customers waiting to use it once the original lease ends. Right now that line exists. Whether it still exists in three years, when this generation of Blackwell chips is no longer Nvidia’s best, is the exact bet the investors in this vehicle would be making.
Related: Google Put a TPU in Orbit, and It’s Not Just a Stunt and Your Next Galaxy Phone Costs More Because Robots Need RAM Too.
Bottom Line: Amazon isn’t buying fewer chips. It’s just figuring out how to keep using them without technically owning them, which tells you more about how strained AI infrastructure budgets have gotten than any earnings call would.
Follow Teck Hustlers: Facebook · X · Instagram · LinkedIn · TikTok · Pinterest
