The Seattle-based cloud giant has held talks with investors in recent weeks to gauge interest in the deal, which would allow the company to spin off thousands of Grace Blackwell chips it is deploying in data centres across the US into a special-purpose vehicle.

Amazon will then lease the advanced AI chips back from the SPV, which would tap outside investors through debt issuance. By unloading the expensive semiconductors to investors, the company could adopt a more asset-light approach to its balance sheet.

Amazon declined to comment.

The proposed move comes as tech giants seek creative ways to finance massive spending on data centre infrastructure. Much of the cost associated with the build-out of these facilities is tied to the chips used to train advanced AI models.

Tech companies are pursuing a range of strategies to move debt off their balance sheets to preserve their creditworthiness. They have used residual value guarantees, providing lenders with assurances about the future value of chips or data centres, without borrowing cash to finance projects themselves. The approach has obscured the amount of risk that tech giants are taking on.

Investors expect the entity will receive an investment-grade credit rating based on Amazon’s current double A rating, paving the way for a broader base of investors such as insurance and pension funds to join the deal.

Amazon also plans to offer an equity stake of up to 10 per cent in the vehicle, meaning that it will not own any stakes in the entity. Discussions between Amazon and investors are ongoing and are subject to change, the people added.

Amazon bought or leased the chips involved in the proposed deal and they have been deployed in more than a dozen US data centres across five states, including Nevada and Virginia, the people added.

Nvidia’s Grace Blackwell chips are some of the semiconductor group’s most advanced but will soon be superseded by its latest Vera Rubin offering.

Leading AI labs such as OpenAI and Anthropic — in which Amazon has committed to invest as much as $83bn — use the latest chips to train their models. Previous generations of silicon continue to be used for running applications. Amazon expects each series of semiconductors to last at least five years, according to regulatory filings.

Amazon is expected to spend $220bn in capital expenditure this year, the majority of which will be allocated towards its cloud unit AWS for the purchase of advanced chips and the wider build-out of AI data centres.

The company has tapped the market to help fund this investment. In March it laid out plans to raise about $50bn through corporate bond issuances, increasing the offering from $37bn following strong demand. But it faced weakening interest in long-dated debt when it sold $25bn in bonds in July, with investors demanding higher yields.

Financing backed by graphics processing units has proven popular among companies such as CoreWeave, which have used their ready access to chips to fund borrowing.

In order to lower borrowing costs for its clients, Nvidia in August offered to backstop up to $125bn of such debt through a $500bn financing platform with major Wall Street groups.