What the FT says Amazon is doing
Amazon has been sounding out investors on moving roughly $8B of Nvidia Grace Blackwell chips into a special-purpose vehicle, according to the Financial Times, via Reuters on October 2. The SPV would own the hardware and raise money by issuing debt; Amazon would offer investors an equity stake of up to 10% and retain no ownership of the entity itself. Amazon then leases the chips back and keeps running them in the same dozen-plus US data centers across five states, including Nevada and Virginia. It is a sale-and-leaseback, applied to silicon. Amazon and Nvidia did not comment.
The money behind it
The motive is scale. Amazon plans roughly $220B in capital expenditure this year, nearly double last year's, with most of it pointed at AWS data centers and chips. CEO Andy Jassy said after Q2 results that even at $220B, demand outstrips supply, and he expects that imbalance to hold through 2028. In the same week, Amazon raised EC2 Capacity Block rental prices about 15% starting next week, covering Nvidia chips from the older A100 to the newer B300, per MarketWatch. So Amazon is simultaneously charging more for GPU time and shopping for someone else to own the GPUs.
What to make of it
The comparison being made is to aircraft: expensive hardware, owned by financiers, leased to operators. The comparison works until you look at depreciation. A commercial jet flies for 25 years. A flagship AI chip is near-obsolete in three. Whoever owns this vehicle absorbs residual-value risk on hardware that loses value faster than almost any asset Wall Street has financed at this scale. Amazon is not eliminating that risk. It is pricing it and handing it to investors willing to take it.
There is a second story underneath. Memory chip prices are surging, and Jassy already raised the 2026 spending target by $20B in July. At some point the AI buildout stops being a capex story and becomes a financing story: debt, vehicles, lease rates. Amazon is the first hyperscaler to put a number on that transition. The number is $8B.
The skeptical read: none of this creates a single additional GPU. The chips exist, they are racked, they are running. All this structure does is decide whose balance sheet carries the cost of the AI boom. That matters for investors. For everyone else, the more telling number is the 15% rental price hike. The cost of compute keeps going up, and now there is a whole new layer of finance collecting a spread on it.
Sources
- [1] Reuters / Financial TimesRead source
- [2] MarketWatch (via Morningstar)Read source
- [3] TechStartups analysisRead source