Amazon may sell nearly $8 billion worth of Nvidia AI systems and then rent the same equipment back. It is a financing plan built around a familiar corporate trick: turn an asset into cash, while keeping it in use and adding a recurring bill.
How the proposed deal would work
A separate owner: Amazon is discussing a special-purpose company, or SPV, that would buy thousands of Grace Blackwell systems already operating in more than a dozen Amazon data centers across five US states.
The SPV would raise additional money through debt financing. Amazon would then pay to use the chips under a lease arrangement. Investors could receive up to a 10% stake in the financial vehicle, while Amazon would not own it.
The talks are at an early stage, and the terms could change. The proposed transaction is not a completed sale, despite the unusual neatness of the sale-and-rent-back idea.
The basic trade: Amazon could receive money for equipment already installed without giving up access to it. In return, it would take on lease payments and the obligations attached to the arrangement. The report describes the structure as a way to shift some spending from conventional capital investment toward a lease-like model.
Why Amazon is looking for financing
The plan comes as Amazon commits heavily to computing capacity. The company has raised its planned capital spending for 2026 to $220 billion, with a significant share tied to AWS, data centers and AI equipment.
Amazon and Nvidia also announced plans in August to deploy another 2 million Nvidia GPUs across AWS’s global infrastructure. The proposed SPV would offer a financing route linked to equipment already in service, rather than only to future construction or purchases.
Amazon has said it is prepared to accept near-term pressure on free cash flow in pursuit of future AWS revenue. The chip-backed financing discussion suggests it is also considering how investors might fund part of the infrastructure buildout. Spending on AI has become large enough that finding new ways to pay for it is part of the product strategy, even if it is not usually the part shown in a demo.
Chips are not buildings
The structure also puts a spotlight on an awkward feature of AI infrastructure: its most expensive components do not necessarily last as long as the places that house them.
Amazon estimates the useful life of chips, servers and networking equipment at about five to six years. Data centers, by comparison, are expected to last more than 30 years. That difference matters to investors financing physical assets. A building can support many generations of hardware, while a chip may need replacement well before the building does.
The timing risk: The financing relies on equipment whose economic life is relatively short. The source report does not detail how the SPV would handle depreciation, replacement or a decline in the chips’ value, so those terms remain part of what investors would need to assess.
What investors might get from it
The SPV could borrow against the equipment, with Amazon’s credit standing helping the vehicle seek an investment-grade rating. That rating, if achieved, could make its debt more attractive to a wider pool of buyers, including insurers and pension funds.
The proposal would connect investor capital directly to AI hardware already running in Amazon’s data centers. It could bring cash into the infrastructure effort without requiring Amazon to stop using the systems, but it would not make the underlying costs disappear. It changes who owns the chips and how Amazon pays for access to them.
For now, the distinction is mostly a financing question, not a change in computing capacity. The chips would still be doing their jobs. The paperwork would simply give them a second one.