Amazon is preparing to transfer roughly $8 billion worth of Nvidia AI servers to a special purpose vehicle and lease the hardware back, the Financial Times reported, citing people familiar with the plan. SiliconANGLE, which summarized the report Friday, said the structure would let Amazon remove the equipment and its associated debt from its balance sheet while continuing to run the chips in its data centers.

What Happened

According to the FT, the vehicle would fund the purchase by issuing debt and by selling a 10% equity stake to outside investors. Amazon has been in talks with potential backers for several weeks, the newspaper reported. Amazon has not publicly confirmed the plan.

The hardware is built on Nvidia's Grace Blackwell accelerators, which pair two Blackwell GPUs with one Grace CPU over Nvidia's NVLink-C2C interconnect. The systems are installed across five Amazon data centers, some owned by Amazon and some leased, according to the report.

RelatedNvidia Plans to Invest in Inference Chip Startup d-Matrix, The Information Reports

SiliconANGLE estimated the size of the fleet using cost figures from cloud provider Modal, which puts each Grace Blackwell unit at $60,000 to $70,000. On that basis, $8 billion would cover roughly 114,000 to 133,000 units. That is SiliconANGLE's calculation, not a figure from Amazon or the FT.

Why It Matters

A sale-and-leaseback lets a company turn hardware it already owns into cash and replace a large upfront asset with ongoing rental payments. Because the vehicle, not Amazon, would carry the borrowing, the arrangement would improve Amazon's debt-to-equity ratio, SiliconANGLE noted, which can support its credit ratings and leave more room to borrow for new capacity.

The hardware in question is no longer Nvidia's newest. Blackwell was succeeded at the top of Nvidia's lineup by Blackwell Ultra, and the company's current flagship is Rubin. Selling older-generation systems into an investor vehicle shifts some of the risk of their declining value to outside capital, while Amazon keeps the compute.

Amazon is also adding chips at a rapid pace. In August, it agreed to buy 2 million more Nvidia GPUs through 2028, spanning Blackwell Ultra, Rubin and the upcoming Rubin Ultra, SiliconANGLE reported. Financing older fleets off balance sheet is one way to free up capacity for those purchases.

Context

Large technology companies have increasingly turned to structured finance to pay for AI buildouts. The Wall Street Journal recently reported that Amazon, Google and eight other technology giants together carry about $3 trillion in off-balance-sheet obligations, much of it tied to AI infrastructure such as data center leases. Meta sold an 80% stake in its Hyperion data center project to outside investors, SiliconANGLE noted.

These deals have drawn scrutiny because they make companies' true commitments harder to read from their financial statements. Lease payments still have to be made, and investors in the vehicles are betting that the hardware will stay in demand long enough to pay back the debt.

What to Watch

The next steps are confirmation from Amazon and details the FT report did not disclose: who the outside equity investors are, how much debt the vehicle would raise and on what terms, and how long the leaseback would run. Amazon's next quarterly report could show whether the transaction closed and how it is accounted for, and other cloud providers with large Blackwell fleets may consider similar structures.