Amazon Wants Investors to Own $8 Billion of Nvidia Chips It Still Uses

The Financial Times reported on October 2 that Amazon has held talks with investors about moving roughly $8 billion of Nvidia Grace Blackwell chips into a separate company, which would raise debt from outside investors, buy the chips, and lease them back to Amazon. Reuters carried the report. No deal has been announced, and Amazon has not confirmed that the talks happened.

The structure is a sale-and-leaseback, the same move airlines have used on aircraft for decades. The point is not to stop using the asset. The point is to stop carrying all of it on your own balance sheet.

How the vehicle would work

Under the reported plan, the chips would sit inside a special purpose vehicle, a standalone company created for a single deal. That vehicle would tap outside investors through debt issuance, and Amazon would lease the hardware back and keep running it. Investors would reportedly be offered an equity stake of up to 10 percent, with Amazon holding no shares itself.

The chips are already installed, according to the reporting, spread across more than a dozen US data centers in five states including Nevada and Virginia. They are expected to stay in service for roughly five more years before Nvidia's successor generation, Vera Rubin, replaces them.

That detail is the whole risk in one line. The vehicle's investors are buying five years of useful life on hardware whose value depends on how fast the next generation arrives. If depreciation runs faster than the lease terms assume, the loss lands on them.

Why Amazon would bother

Amazon said earlier this year that it expects to invest around $200 billion in capital expenditure across the company in 2026, driven by AI, AWS, chips and robotics. That spending has already shown up in its cash flow. Free cash flow turned negative on a trailing twelve month basis, an unusual place for a company of its size, and the cause was a large jump in purchases of property and equipment.

An off-balance-sheet vehicle does not make the chips disappear from the world or from Amazon's operations. It moves the ownership and part of the financing to someone else, which can keep the company's own credit metrics cleaner and free up cash for the next order. Amazon ended June with roughly $123 billion in cash and marketable securities, so this is not a company that needs the money to survive. It is managing how the spending looks and how much of it sits on its own books.

The pattern this fits into

Amazon would not be inventing this. Nvidia agreed with six financial firms on financing platforms worth around $500 billion, where its own chips serve as collateral. Google used a comparable structure to finance Anthropic's compute, with a purpose-built vehicle buying the chips and leasing them out. And a $42 billion credit commitment from Broadcom to Anthropic was announced shortly before the Amazon report, part of a larger package.

The direction is consistent. AI infrastructure is moving from being paid for mostly out of the big technology companies' own pockets toward being financed by outside debt, the way railways and telecom networks were a century ago. One industry estimate cited in coverage puts the potential debt raised for AI infrastructure at up to $1 trillion by 2030.

Whether that comparison flatters the current moment is worth sitting with. Railways lasted generations and their assets stayed useful. A cluster of graphics chips has a useful life measured in a handful of years, and its revenue depends on customers continuing to pay for inference and training. The leaseback spreads the cost. It does not spread the demand risk away from the operator, because Amazon still needs the capacity to earn its keep.

The depreciation problem at the centre

The reason this structure is worth watching is that it moves a depreciation question from Amazon's income statement to someone else's. Amazon already shortened the expected useful life of some servers and networking gear from six years to five, citing faster technology development, and that change added $1.4 billion to 2025 depreciation expense. When a company is buying hardware by the tens of billions, the useful-life assumption is not a footnote. It is a large number multiplied across a huge fleet.

Investors in the vehicle are, in effect, taking a view on that assumption. If the chips earn their keep for five years, the lease payments cover the debt and everyone is fine. If the next generation of hardware makes them uncompetitive in three years, the vehicle is holding assets worth less than it owes, and the loss lands on the debt and equity holders rather than on Amazon's books. That is the trade being offered: better yield in exchange for residual-value risk.

Why the timing looks the way it does

The report surfaced at a moment when the AI buildout is shifting from an engineering story to a financing story. Nvidia has grown into a company whose quarterly revenue dwarfs the annual budgets of most industries, and its own financing arrangements with financial firms run into the hundreds of billions. Google has financed a customer's compute through a similar vehicle. Broadcom committed a large credit line to the same customer shortly before the Amazon news.

When several large buyers reach for the same instrument within weeks of each other, it usually means they have hit the same wall. The ability to buy chips was rarely the constraint. What has become hard is the willingness to keep reporting the associated debt against their own credit, in a market that now watches those balance sheets closely.

What to watch

The clearest early signal is whether other hyperscalers set up similar vehicles that name specific hardware, with disclosed sizes and timelines. Concrete multi-year leases on Grace Blackwell or later platforms would show that off-balance-sheet AI financing is becoming a repeatable part of the order pipeline rather than a one-off arrangement.

The second signal is ratings. The reported vehicle would issue its own bonds, and the entire premise depends on those bonds being rated close to Amazon's own credit quality. That rating is what would draw insurers and pension funds, who avoid direct bets on young AI infrastructure. If the placement works, it becomes a template. If the rating comes in lower than hoped, the financing gets more expensive and the appeal fades.

The honest caveat

This is reporting about exploratory talks. No investors, lenders, lease terms or accounting treatment have been disclosed, and setting up a separate company does not by itself move the chips off Amazon's books under accounting rules. The structure could change, shrink, or never happen.

Still, the direction of travel is visible. The largest buyers of AI hardware are no longer only asking how much compute they can buy. They are asking how to buy it without showing all of it on their balance sheets, and the answer increasingly involves handing the asset to someone else while keeping the use.