Why Amazon Wants to Move $8 Billion in Nvidia Chips Off Its Balance Sheet

According to Cailian Press and multiple financial media outlets on October 2, Amazon is in talks with investors about an asset financing arrangement: transferring about $8 billion worth of Nvidia Grace Blackwell chips deployed in data centers across the United States to a special purpose vehicle (SPV), then leasing them back from that vehicle for continued use. The plan is still under discussion and has not been finalized.

The chips are still the same chips, and the data centers are still the same data centers. What has changed is the books.

A deal that looks like a sale but is actually borrowing

The structure is not complicated. Amazon transfers ownership of the chips to the SPV, the SPV raises funds from outside investors by issuing bonds, and Amazon pays rent to regain the right to use them. Amazon also plans to give up up to 10% of the SPV's equity, meaning it will no longer hold a stake in the vehicle.

The deal involves thousands of Grace Blackwell chips already installed in more than a dozen data centers across five states, including Nevada and Virginia. The chips were purchased or leased by Amazon itself.

For Amazon, high-priced semiconductor assets on its books become rental expenses, moving its operating model one step toward asset-light. For investors, they get debt backed by Amazon's credit and generating steady rental income. People familiar with the matter say that, thanks to Amazon's current AA rating, the vehicle is expected to obtain an investment-grade rating, allowing more conservative money such as insurance funds and pensions to participate.

This is no small amount. Amazon is expected to spend $220 billion in capital expenditure this year, most of it on AWS, to procure high-end chips and expand AI data centers. It is already using the bond market to fund this spending: in March it announced about $50 billion in corporate bonds, upsized from an original $37 billion due to strong demand; by the time it issued another $25 billion in bonds in July, demand for long-term debt had weakened and investors began demanding higher yields.

Moving assets off the balance sheet is not unique to Amazon. When a company's capital expenditure is mainly spent on rapidly depreciating chips, maintaining its credit rating and financing costs becomes a problem it must address.

The controversy is not about the money, but whether GPUs are still good collateral

What Wall Street is really debating over and over about this deal is whether GPUs can serve as long-term collateral.

Nvidia's argument is that its GPUs and full AI factory systems are versatile, can be transferred among different customers and operators, and can have their economic lives extended through software upgrades. The company says that for some projects it can provide residual value support on a case-by-case basis of up to 25% of project size. In August, it announced plans to establish an independent compute financing platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, aiming to mobilize more than $500 billion in third-party capital over time. Nvidia added that this figure is the total capital pooled across multiple platform initiatives, not its own revenue, nor a single investment commitment to any one customer.

Financial institutions are less optimistic. GPU technology iterates too quickly, and long-term resale value lacks sufficient historical data to support it. When pricing GPU-backed loans, banks typically assume a depreciation cycle of about three to four years; Nvidia believes high-end chips can generate revenue for as long as a decade. The gap between these two sets of assumptions will ultimately show up in collateral terms, interest rates, and debt-service protections.

Earlier, CoreWeave obtained $8.5 billion in investment-grade GPU-backed financing, but the rating on that loan relied heavily on Meta's contractual payment commitments rather than the residual value of the chips themselves. This shows that what the market is currently willing to accept are assets “with a stable customer paying,” while it remains cautious about “how much the chips will still be worth in a few years.”

This time the chips' situation is more delicate. Grace Blackwell is currently Nvidia's most advanced product line, but it will soon be replaced by the newer Vera Rubin. The latest generation of chips is usually used to train frontier models, while the previous generation continues to run business applications. Regulatory filings show that Amazon expects each generation of semiconductor products to last at least five years.

How much risk does off-balance-sheet financing hide?

Academics and regulators have long been wary of such arrangements because they move risk off the balance sheet without truly eliminating it.

Through residual value guarantees, companies promise lenders the future value of chips or data centers, thereby avoiding direct borrowing. The upside is cleaner books; the downside is that it becomes harder for outsiders to judge how much risk a company actually bears. Amazon plans to give up all equity in the SPV, so in theory it is no longer the owner, but as lessee it must keep paying to use the chips. If the chips depreciate ahead of schedule, the lease obligation does not shrink accordingly.

For investors, there are two keys to judging this deal: the ability of rental income to cover the debt, and whether the residual value assumptions hold up. The former depends on whether AWS can continue to sell this compute capacity in the future; the latter depends on whether the GPU market will slow down as Nvidia expects.

What this means for the industry

If the deal goes through, it will be a public experiment in cloud providers converting expensive AI chips into financeable assets. Cloudflare, Microsoft, and Google are all looking for ways to finance compute in different ways, with different models, but the common thread is that capex pressure has grown so large that outside capital is needed to share the burden.

For small and medium-sized businesses looking to buy compute, no direct impact is apparent in the short term. Compute prices depend on supply and demand, not Amazon's accounting treatment. What such arrangements will really change is the funding structure of AI infrastructure. When the collateral value of chips becomes something repeatedly priced by public markets, the entire industry's financing costs will more closely follow capital market sentiment, rather than just technology roadmaps.

Nvidia wants GPUs to be collateralizable long term like real estate, while Wall Street is currently only willing to value them with the rapid depreciation of consumer electronics. Which judgment is right will not have an answer within a few years. Amazon's $8 billion deal is precisely an attempt to put this question on the table.