Amazon Explores Sale-Leaseback Structure for $8 Billion in Nvidia Chips
This article was created with the help of artificial intelligence.
Key Takeaways
- Amazon is exploring a financing structure to remove approximately $8 billion worth of Nvidia chips from its own balance sheet.
- The Grace Blackwell GPUs are already installed in more than a dozen data centers across at least five US states or are currently being installed.
- Amazon told investors to expect capital expenditures of $220 billion for 2026 – $20 billion more than initially announced.
- Free cash flow for the past twelve months fell to negative $7.6 billion, driven by a 66.1 billion year-on-year increase in capital expenditures.
- Grace Blackwell is already one generation behind Vera Rubin, Nvidia's newer architecture, with token costs for Vera Rubin estimated at one 35th of those for Grace Blackwell Ultra.
- The discussions remain in an early stage and could still change; whether and when such a vehicle will actually be launched remains open.
Amazon is exploring a financing structure to remove approximately $8 billion worth of Nvidia chips from its own balance sheet. According to information from October 3, 2026, the company recently surveyed investors about a so-called sale-leaseback model: thousands of Grace Blackwell GPUs, which are already installed in more than a dozen data centers across at least five US states – including Nevada and Virginia – or are currently being installed, would be transferred to a special-purpose entity and leased back from there. Amazon and Nvidia declined to comment.
Special-Purpose Entity with Debt Financing and External Participation
The structure envisages Amazon contributing the chips to the special-purpose entity, which would refinance itself through bond issuance. Investors could receive up to 10 percent of the shares in the special-purpose entity; Amazon itself would hold no ownership stakes. The hardware would remain installed in US data centers, Amazon pays lease payments and retains access to the computing capacity.
The initiative marks one of the clearest attempts by a hyperscaler to date to treat AI chips as a standalone asset class capable of supporting debt and leasing structures – rather than carrying them permanently as tangible assets on the balance sheet. Through the sale-leaseback, Amazon could shift the obligation into lease disclosures and protect its creditworthiness.
Capital Expenditures Rise to $220 Billion
The considerations reflect the enormous capital requirements for AI infrastructure. Amazon told investors to expect capital expenditures of $220 billion for 2026 – $20 billion more than initially announced. The bulk goes to data centers and chips for the AWS cloud division. Chief Executive Andy Jassy cited rising memory chip prices and sustained strong demand for AI capacity in connection with the second-quarter earnings in July as reasons for the increased guidance. Jassy acknowledged that Amazon likely would not be able to meet all customer demand through 2026 and into 2027.
The financial strain shows in cash flow figures: free cash flow for the past twelve months fell to negative $7.6 billion, driven by a 66.1 billion year-on-year increase in capital expenditures – primarily for AI investments. Long-term liabilities climbed by 53.5 billion within a year to 119.1 billion US dollars. In the second quarter of 2026 alone, Amazon spent $54.2 billion on capital projects, an increase of roughly 68.6 percent compared to the prior year quarter.
Bond Issuances Meet Mixed Reception
To finance the expansion, Amazon repeatedly tapped the bond market. In March 2026, the company announced a corporate bond offering of approximately $50 billion – more than the originally planned $37 billion after demand proved strong. A second issuance of $25 billion in July encountered greater resistance; investors demanded higher yields for longer maturities. In total, Amazon raised $75 billion through bonds in 2026.
Jassy explained in analyst calls that Amazon had already accessed debt capital markets and had multiple financing options available.
Technological Obsolescence as a Risk for Investors
A central risk for potential investors in the special-purpose entity is the rapid technological obsolescence of AI chips. Grace Blackwell is already one generation behind Vera Rubin, Nvidia's newer architecture. Nvidia estimates token costs for Vera Rubin at one 35th of those for Grace Blackwell Ultra. The chip manufacturer estimates revenue per gigawatt at approximately $25 billion for Grace Blackwell and roughly $40 billion for Vera Rubin – each new generation reduces the value of the collateral.
According to Jassy, servers reach breakeven "a bit under three years" and last at least five to six years. In a sale-leaseback model, Amazon would retain the early, profitable years of chip usage, while creditors would rely on later years when a chip is two or more generations obsolete and must be refinanced or resold.
Discussions Remain in Exploration Stage for Now
According to October information, Amazon spent several weeks informing investors about the planned structure. Discussions remain in an early stage and could still change. Whether and when such a vehicle will actually be launched remains open.