Amazon is in talks with investors to move thousands of Nvidia Grace Blackwell chips into a special-purpose vehicle and then lease them back, a structure that could take roughly $8 billion of expensive AI hardware off the company's balance sheet.

The figure, reported by Moneycontrol, anchors what would be one of the largest hardware sale-and-leaseback arrangements ever attempted in the AI infrastructure market. The deal is not closed. The talks could still change in scope or fall apart, and Amazon has not confirmed the identity of the investors involved or the final size of the package.

The mechanics matter for anyone tracking AI capex. Under the proposed structure, outside investors would own the Grace Blackwell chips outright through a special-purpose vehicle. Amazon would lease the same hardware back and continue deploying it in its data centers. The chips themselves never move. Only the ownership does.

That distinction is the whole point. AI accelerators have become the single largest capital line item for hyperscalers, and Grace Blackwell — Nvidia's flagship data-center platform pairing Grace CPUs with Blackwell-generation GPUs — sits at the top of the cost curve. Moving billions of dollars of that hardware off the balance sheet would relieve pressure on reported capital expenditure while preserving full operational access to the compute.

The commercial logic for the investors is straightforward in outline, though the report does not detail the terms under discussion. Buyers of the special-purpose vehicle would hold a pool of income-producing assets with a contracted lessee of Amazon's credit quality. Amazon, in turn, converts a massive up-front capital outlay into an operating expense spread over the lease term.

For Nvidia, the arrangement changes nothing on the demand side — the chips are already sold and deployed — but it signals how its largest customers are now financing the hardware. If the structure works, it creates a template that other hyperscalers burdened by AI capex could copy. That would broaden the buyer base for future accelerator generations by making them financeable assets rather than pure balance-sheet drains.

The scale is unusual but not unprecedented in kind. Sale-and-leaseback structures are standard tools in real estate, aircraft, and telecom infrastructure. What is new here is applying them at an $8 billion scale to semiconductors with short useful lives and rapid generational turnover — a risk profile lenders and investors are still learning to price. The report does not say how the parties would handle depreciation of the chips or what happens at the end of the lease term.

For Amazon specifically, the timing aligns with the industry-wide surge in AI infrastructure spending. Hyperscaler capital expenditures have climbed steeply as companies race to build out training and inference capacity, and financial engineering that trims reported capex while keeping the compute in place offers visible appeal to investors watching margins.

The talks also underline how scarce and expensive top-tier AI silicon remains. Grace Blackwell systems carry premium pricing and long lead times, which is precisely what makes them attractive as collateral-like assets for outside investors seeking exposure to AI compute demand.

Nothing is finalized. Amazon and the investors are still negotiating, and the roughly $8 billion figure describes the potential value of hardware involved rather than a signed commitment. If the deal closes, it will mark the most significant test yet of whether institutional capital will fund hyperscaler AI hardware at scale — and whether lease-backed accelerators become a permanent fixture of how the industry pays for compute.