Amazon Explored $8 Billion Nvidia Chip Financing via Special Purpose Vehicle

Amazon explored a financing structure in early October 2026 to transfer approximately $8 billion worth of Nvidia chips into a special purpose vehicle and then lease the hardware back. The discussions were at an early stage at that time and could have changed.

$8 Billion in Chips as Lease Assets

According to the proposed structure, Amazon would contribute already installed or in-deployment Grace Blackwell GPUs from Nvidia into a Special Purpose Vehicle (SPV). This would be financed through debt, while Amazon would lease the semiconductors back and continue using them in its data centers. External investors could receive equity stakes of up to 10 percent in the special purpose vehicle, while Amazon itself would hold no shares.

According to sources, the chips are already deployed in more than a dozen data centers across at least five US states, including Nevada and Virginia. The transaction would remove the assets from Amazon's balance sheet and shift the lease obligation to the notes rather than reporting it as debt.

Amazon spent the weeks before reports emerged briefing potential investors on the planned structure. The company declined to comment, and Nvidia did not respond to inquiries.

$220 Billion Capital Expenditures Strain Cash Flow

The move comes against the backdrop of enormous financing burdens for expanding AI infrastructure. Amazon expects capital expenditures of $220 billion for 2026, an increase of $20 billion from its previous guidance. The bulk flows into data centers for its cloud division AWS and the chips required for them.

Free cash flow over the past twelve months fell to negative $7.6 billion, driven by a $66.1 billion increase in property, plant and equipment investments compared to the prior year. In the second quarter alone, Amazon spent $54.2 billion on capital projects, 68 percent more than a year earlier.

Long-term debt rose to $119.1 billion from $65.6 billion a year earlier, an increase of 82 percent. To finance the expansion, Amazon repeatedly turned to the bond market. In March, the company announced a corporate bond offering of approximately $50 billion, an increase from the originally planned $37 billion due to strong demand. Another offering of $25 billion in July encountered greater resistance, with investors demanding higher yields for longer maturities.

Jassy: Cannot Meet Demand in 2026 and 2027

CEO Andy Jassy said during quarterly earnings in July that rising memory chip prices and sustained demand for AI capacity are driving the higher spending targets. He acknowledged that Amazon had already used debt capital markets and had multiple financing options available. At the same time, he warned that the company would likely be unable to fully meet customer demand through 2026 and into 2027.

The SPV structure would enable Amazon to maintain creditworthiness while shifting the lease obligation to the notes of the balance sheet rather than appearing as debt on the liabilities side.

Chip Obsolescence as Risk for Investors

A central issue with the structure lies in the rapid aging of hardware. Grace Blackwell is already one generation behind Nvidia's current Vera Rubin architecture. Nvidia puts the token cost of Vera Rubin at 35 times the advantage over Grace Blackwell Ultra and estimates revenue per gigawatt at approximately $25 billion for Grace Blackwell and around $40 billion for Vera Rubin.

According to Jassy, servers amortize in somewhat less than three years and last at least five to six years. Under the SPV structure, Amazon would retain the early, profitable years, while lenders would be faced with later years in which a chip two generations old would need to be refinanced or resold.

In the sources, the agreement is characterized as smart for Amazon but potentially risky for debt providers. Unlike traditional sale-leaseback structures such as with aircraft, which remain in service for decades, semiconductors lose value much more rapidly.

Hyperscalers Treating Chips as Standalone Asset Class

The transaction would be one of the most direct attempts by a hyperscaler to treat AI chips as a standalone asset class that can support debt and lease structures, rather than permanently carrying them as operating assets on the balance sheet. Should the structure be realized, it could set a pattern for other cloud providers facing similar capital constraints.