Amazon structures an $8B chip-leaseback agreement with NVIDIA, financing GPU procurement through financial leasing rather than direct capital purchase.
Amazon is in talks to move roughly $8 billion worth of Nvidia Grace Blackwell chips off its own balance sheet and into a newly created financing vehicle, then lease the same hardware straight back for use in its U.S. data centers, according to a report from the Financial Times citing people familiar with the matter. Neither Amazon nor Nvidia has publicly confirmed the deal, and the proposal is still described as under discussion rather than signed. But the structure itself marks a notable shift in how the biggest cloud providers are paying for the AI buildout, and it puts a dollar figure on a financing trend that has been building quietly all year.
If the deal closes, thousands of Grace Blackwell chips, already installed or being installed in U.S. data centers, would be placed inside a special-purpose vehicle (SPV). That SPV would raise money through debt issuance from outside investors, use the proceeds to effectively buy the chips, and then rent the same silicon back to Amazon Web Services. The hardware never leaves the building. What changes is who technically owns it and who carries the debt used to pay for it.
Strip away the financial engineering and the core claim is simple: Amazon wants to turn roughly $8 billion of Nvidia Grace Blackwell chips into a leaseback arrangement instead of carrying them as a straight capital purchase. The Financial Times, credited with first reporting the story, cited unnamed sources familiar with the discussions—a meaningful caveat, since nothing here has been confirmed on the record by Amazon or Nvidia, and several details circulating online remain unverified. What is solid, per the reporting, is the shape of the deal: an SPV holding Grace Blackwell chips, debt raised against that SPV, and Amazon leasing the hardware back for AWS workloads. The chips are already slated for, or already running in, U.S. data centers.
The mechanics of a chip leaseback SPV are not exotic in finance circles, even if they are new territory for AI infrastructure at this scale. The SPV reportedly plans to raise its funding through debt issuance rather than a straight equity sale. That detail matters because debt investors care about one thing above all: whether they get paid back. An SPV backed by a lease commitment from Amazon looks very different to a bond buyer than, say, a startup's GPU cluster. Amazon's credit profile effectively becomes the backbone of the structure, even though the debt technically sits outside Amazon's own balance sheet.
On the other side of the ledger, Amazon pays rent to the SPV to keep using the chips. That rent becomes the cash flow that services the SPV's debt. It is functionally similar to a company selling its headquarters building to a landlord and then signing a long-term lease to keep working out of the same offices. The building doesn't move. The company just converts a big upfront asset into a recurring expense.
Nvidia's Grace Blackwell platform is the chipmaker's flagship AI training and inference hardware for the current generation, pairing Nvidia's Grace CPU with Blackwell GPUs in a tightly coupled rack-scale design. It is expensive, it is in enormous demand, and it is exactly the kind of asset that depreciates fast on paper even when it keeps running useful workloads for years. That combination—high upfront cost plus fast on-paper depreciation—is precisely what makes a chip attractive for an off-balance-sheet financing structure.
There is also a succession question hanging over the hardware. Nvidia has already signaled its next-generation Vera Rubin platform as the follow-up act. Every hyperscaler buying Grace Blackwell today is implicitly betting that the chips stay useful long enough to justify the spend before Vera Rubin capacity becomes widely available. Structuring the purchase as a lease, rather than owned capital equipment sitting on the books for years, gives Amazon more room to swap in newer silicon later without stranding an asset.
What makes this story resonate beyond Amazon is the pattern it fits into. AI compute has become so capital-intensive that hyperscalers are increasingly borrowing tools from private equity and structured finance rather than just writing checks from cash flow. An SPV lets a company keep buying hardware at the pace AI demand requires while keeping the related debt off its own primary balance sheet, which can matter for credit ratings, reported leverage ratios, and how Wall Street analysts score the company's core financials each quarter.
This is not a new idea in corporate finance generally. Airlines have financed jet fleets this way for decades. What is new is applying it to GPUs that depreciate on a completely different timeline than aircraft, and doing it at a scale where a single reported transaction touches $8 billion in hardware.
Amazon's reported SPV talks land in the middle of a busy year for chip-adjacent financing and infrastructure deals across the AI hardware stack. The year 2026 has seen Synopsys and Amazon move forward on a custom chip design investment valued at $1 billion or more for AWS silicon. Nvidia announced a $150 billion share buyback program despite heavy AI capex demands. Blackstone acquired a data center campus from Nvidia for $99.6 million to support AI operations. CoreWeave deployed 11,264 Nvidia custom CPU cores, signaling that cloud providers are leaning on Nvidia's non-GPU silicon as well. And Nvidia's B200 cloud rental rates have climbed to $8.01 per hour, up 79 percent, reflecting the rising cost of renting the compute these deals finance.
Taken together, the pattern is consistent. Every major player touching Nvidia hardware this year, whether buying it, leasing it, housing it, or renting it out by the hour, is dealing with the same underlying pressure: AI compute costs more than any single balance sheet wants to absorb in one go.