Amazon Plans to Sell $8 Billion of Nvidia Chips and Rent Them Back

Amazon is not selling a data center. It is selling the chips inside one.

The Financial Times reported on October 2, citing unnamed sources, that the company is in talks to move roughly $8 billion of Nvidia Grace Blackwell chips. They would go into a special-purpose vehicle.

The chips will not move. Amazon would lease them back and keep using them.

What the Report Says

The chips are reportedly spread across more than a dozen data centers in five U.S. states, including Nevada and Virginia. They number in the thousands.

The vehicle would raise most of its money through debt. Outside investors could take equity stakes of up to 10%.

If the vehicle earns an investment-grade rating, insurers and pension funds could also buy in. Amazon has held recent talks with investors. Nothing is signed and the figures may change. Amazon and Nvidia did not comment.

Reuters, relaying the FT, said the aim is to strengthen the balance sheet through "a more asset-light approach."

It Looks Like a Sale

In substance, it is a way to borrow.

Amazon receives $8 billion in cash today and commits to years of rent. Economically, that closely resembles pledging the chips as collateral for a loan.

The difference is in the accounting. Direct borrowing shows up as debt. A lease, if structured the right way, moves the cost out of capital spending and into operating expense.

The Math: Own It, or Sell and Lease

The figures below are hypothetical, since lease terms were not disclosed. The goal is to put the cash flows of the two paths side by side.

Amazon pays that price so that capital spending falls now and cash flow looks stronger in the early years.

But that only works if the rent does not cover the chips' full value.

Leaving the Balance Sheet Means Handing Over Risk

Under U.S. accounting rules (ASC 842), a sale-leaseback does not always count as a sale.

Here is the tension. The easiest way to win over investors is to pay rent covering the chips' full life. Then investors rely on Amazon's credit, not on the chips.

But the more fully the rent covers the value, the more the deal resembles debt, and the weaker the off-balance-sheet goal becomes.

For the chips to truly leave the books, someone has to own the risk of what a five-year-old Nvidia chip will be worth.

Why Chips Make Hard Collateral

Traditional collateral such as aircraft or buildings holds value for decades. Data center chips do not: Nvidia launches a new architecture every year, and older generations lose value fast on the secondary market.

Meta's partnership with Blue Owl for its Hyperion data center in Louisiana covered buildings and power infrastructure. The structure Amazon is discussing goes a step further and finances the computing hardware itself.

A forecast cited in the same coverage puts cumulative hyperscaler spending near $9 trillion through 2031, with debt issuance possibly reaching $500 billion next year.

Each appeared on this site separately: the Blackstone-Google loan, Nvidia's search for insurers and Broadcom's Anthropic financing. In those three, the money came from the chipmaker or the tenant. At Amazon, the owner of the chips is trying to lighten its own balance sheet.

Scale and Reading

For Amazon, $8 billion is roughly one-twentieth of more than $200 billion in annual capital spending. On its own, it will not change the balance sheet.

The pattern is what matters. If it works, the same structure could repeat with each new wave of chips, and Nvidia hardware would move into the portfolios of insurers and pension funds.

Amazon's last three months give context; the report alone did not carry the stock.

For the broader path of big tech, the Nasdaq 100 fund:

Sources: This story rests on a Financial Times report of October 2, 2026, based on unnamed sources; details were checked against accounts by Bloomberg and Reuters and a Taipei Times report. Amazon's capital spending appears in sources both as "more than $200 billion" and "$220 billion"; this story uses the lower bound. The 6.5% funding cost and five-year life in the math section are hypothetical. The accounting discussion reflects the general principle of the U.S. standard ASC 842; how the deal would be classified cannot be known until its terms are disclosed.