Amazon Wants Investors to Buy Its $8B Nvidia Chip Bill
Amazon Wants Investors to Buy Its $8B Nvidia Chip Bill
The Financial Times says Amazon has a plan for its colossally expensive AI habit: stop owning the chips and start renting them. Per the FT, the cloud giant is quietly sounding out investors on a deal: roughly $8 billion of Nvidia Grace Blackwell chips go into a special-purpose vehicle mostly owned by outsiders — and then Amazon leases the silicon right back. A mountain of capital spending becomes a much tidier line on the balance sheet. If it closes, it will rank among the AI boom's boldest financing stunts. Possibly the boldest. If it collapses, it still shows where the industry's anxiety sits right now. Not whether AI demand exists — who actually pays for it.
One thing has to sit above the rest of this story, because most of today's coverage will bury it: this is a single-sourced report. The FT's story, published Friday around 04:00 UTC and citing unnamed "people familiar with the matter," is the sole origin of every detail below. We traced every version of this story we could find this morning — Reuters, Bloomberg, the Economic Times, Channel News Asia, Emirates 24|7, and a small forest of aggregators — and every one of them routes back to that one FT piece. No outlet has independently confirmed a single line of it. Amazon and Nvidia, for their part, said nothing at all. Keep that in mind, and the story gets a lot more interesting.
Amazon's $8B chip SPV: what the FT actually reported
Reuters moved the FT's reporting at 04:12 UTC (bylined Shivani Tanna in Bengaluru), and the mechanics are straightforward enough. In recent weeks, Amazon has held talks with investors to gauge interest in moving thousands of Grace Blackwell chips — the top-of-the-line Nvidia accelerators it is currently installing in US data centres — into a special-purpose vehicle. Amazon would then lease the chips back from the SPV, which would fund itself by issuing debt to outside investors. To sweeten the pitch, Amazon reportedly plans to offer investors an equity stake of up to 10% in the vehicle, keeping the rest for itself.
The footprint is large: the chips were bought or leased by Amazon and are installed in more than a dozen US data centres across five states, including Nevada and Virginia. The FT's stated rationale is that the move would "strengthen" Amazon's balance sheet and give the company a more "asset-light" posture as AI spending soars. The story broke at 4 a.m. London time — around midnight in New York, when both companies' PR desks were dark — which is why Reuters could only note that Amazon and Nvidia "did not immediately respond ... outside regular business hours." One suspects the phone lines will be warmer now.
Why $8 billion is pocket change — and a thunderclap
Context is everything here. Back on February 5, 2026 — alongside fourth-quarter 2025 results — Amazon told investors it expects about $200 billion of capital expenditure in 2026, overwhelmingly for AWS and AI infrastructure. The stock dropped 10 to 11 percent after hours on the news. Jassy didn't flinch in the company's official statement: "With such strong demand for our existing offerings and seminal opportunities like AI, chips, robotics, low earth orbit satellites, we expect to invest about $200 billion in capital expenditures across Amazon in 2026, and anticipate strong long-term return on invested capital."
On the earnings call that day, Jassy spelled out where the money goes — "If you look at the capital we're spending and intend to spend this year, it's predominantly in AWS. Some of it is for our core workloads, which are non-AI workloads, because they're growing at a faster rate than we anticipated. But most of it is in AI, and we just have a lot of growth and a lot of demand," he told analysts, as reported by Data Center Dynamics.
Against a $200 billion capex plan, $8 billion in chips is four percent. That's the pocket-change reading. The thunderclap reading is different: if Amazon is exploring sale-leasebacks on its GPUs, it means the people running the world's biggest cloud are actively hunting for ways to make the AI buildout look smaller than it is. This summer the Wall Street Journal calculated that nine big tech firms are carrying roughly $3 trillion in AI commitments largely off the balance sheet — unstarted data-centre leases and purchase commitments dwarfing the reported capex. Amazon's SPV, if real, is the next logical step in that direction: not hiding the spending, exactly, but reclassifying it into something prettier.
The balance-sheet trick everyone is trying at once
Amazon would not be inventing this playbook. Just yesterday, Reuters reported on Nvidia's own push to get its chips treated as financeable collateral, with Blackstone, Apollo and KKR involved in a plan offering residual-value guarantees of up to 25%. Morningstar analysts were unimpressed, questioning the private-credit structures and circular-deal risk, and banking sources told Reuters the market isn't ready to treat compute as investment-grade collateral the way it treats aircraft. A Nvidia spokesperson told Reuters on October 1 that "Our financing partners independently assess each opportunity, including customer commitments, expected cash flow and residual value," adding that "Financing structures will vary as this market develops."
The parallel is hard to miss. Both Amazon and Nvidia are asking the same question of Wall Street: will you lend against GPUs? And just this week we reported on Broadcom's $42 billion arrangement with Anthropic, in which the chipmaker effectively financed its customer's lease of its own silicon. Add Nvidia's $668 million bet on Taiwan's GMI Cloud GPU infrastructure, and a pattern snaps into focus: the AI supply chain is increasingly financing itself. Vendors lend, customers lease, and the distinction between a buyer and a banker gets blurrier by the quarter.
The accounting catch nobody should skip
A note of care before anyone calls this "off-balance-sheet" and moves on: under modern lease accounting (ASC 842), a sale-leaseback doesn't magically make obligations vanish. Amazon would still book a right-of-use asset and a lease liability; the economic commitment — paying for those chips for years — remains. What changes is the optics and the cash-flow timing: reported property and equipment grows more slowly, certain return-on-asset ratios look stronger, and the upfront cash hit is somebody else's problem. Credit analysts, the good ones, will look straight through it. The FT's framing — "strengthen its balance sheet" — is Amazon's framing, and it deserves to be read as what it is: the company's own sales pitch to investors, laundered through anonymous sources.
There is also the small matter of what, exactly, is being moved. The FT itself hedges that the chips were "bought or leased" by Amazon — meaning even the premise of the transaction is unresolved in the reporting. And the identity of the investor counterparties is unknown; the FT says only that talks to gauge interest happened "in recent weeks."
What this says about the AI economy in October 2026
Step back and the story is less about Amazon than about the moment. Six months ago the debate was whether the AI buildout had enough demand. Now the debate is whether it has enough financing. The Bank of England's Andrew Bailey made the same point to the BBC this week: markets are pricing AI as if every player wins. "Not everybody is a winner," Bailey said. And his Financial Policy Committee is now warning, on the record, about debt-financed AI investment and what it calls "circular arrangements." Set against that, Amazon taking an $8 billion GPU portfolio to bond investors reads less like genius and more like a stress test. Two open questions: is there a real market for AI-infrastructure debt? Or are the vendors themselves holding up the edifice? KKR's credit desk put a number on that anxiety this week — roughly $600 billion in AI-related debt, or 6% of the US investment-grade market.
Our position, stated plainly: clever if it closes — but not cleansing. Leasing GPUs from an SPV doesn't change how many chips Amazon needs, how fast they depreciate, or whether the AI revenues ever arrive to pay for them. What it does is spread the risk to investors willing to price it — and the honest question is whether those investors understand GPU residual values any better than the Morningstar analysts who just tore apart Nvidia's version of the same idea. Amazon built its empire on other people's warehouses and other people's money; applying the same logic to its own AI chips is on-brand. But a balance sheet that looks asset-light while the obligations quietly compound in the footnotes is not a strengthened balance sheet. It's a prettier one.
The talks may yet collapse — most exploratory financing talks do, and this one has been reported by exactly one newspaper. But watch what happens next: if Amazon finds buyers for GPU-backed paper, every hyperscaler with a capex problem will be at the same window within a quarter. And that, more than any $8 billion number, is the story.
Sources
Financial Times (via Reuters, October 2, 2026 — moved 04:12 UTC, by Shivani Tanna, Bengaluru; Bloomberg's follow; syndication via Economic Times, Channel News Asia, Emirates 24|7); Reuters, "Nvidia's bet that its chips can finance the AI boom gets a Wall Street reality check" (October 1, 2026); Data Center Dynamics on Amazon's $200B 2026 capex (February 2026); GeekWire and CNBC TV18 on Amazon Q4 2025 earnings; Wall Street Journal analysis of big-tech off-balance-sheet AI commitments (August 17, 2026, via write-up); BBC interview with Bank of England governor Andrew Bailey (October 1, 2026).
Read next: Amazon finally weighs in on AI safety — and won't pick a side.
Daniel Rao is the founder and editor of TechPulse USA, covering AI, chips and the geopolitics of the intelligence age — from data-centre financing to export controls.
Comments
Post a Comment