Amazon sells its AI chips, then rents them back

Amazon is selling its AI chips and renting them back. The company is in talks with investors to offload roughly $8 billion worth of Nvidia Grace Blackwell chips into a special-purpose vehicle, then lease the same hardware back, the Financial Times reported late on October 1, 2026 (UTC). It is an old financial trick — airlines have done this with planes for decades — applied to the hottest hardware in the world. And it tells you something important: the AI buildout has gotten expensive enough that even Amazon needs to move some of it off its balance sheet.

Under the structure being discussed, thousands of Grace Blackwell chips already installed across more than a dozen U.S. data centers in five states — including Nevada and Virginia — would be transferred into the SPV. The vehicle would raise money from outside investors through debt issuance, and Amazon could sell investors an equity stake of up to 10%. Amazon then leases the chips back from the SPV and keeps using them exactly as before. The chips stay in the same racks, running the same workloads. The only thing that changes is who owns them on paper.

The numbers explain why Amazon is doing this now. The company spent $44.2 billion on capital expenditures in Q1 2026 alone and expects to deploy roughly $200 billion for the full year — nearly double its 2025 outlay. Most of that money goes to AWS data centers and the AI hardware inside them. Management has already warned that heavy near-term capex will weigh on free cash flow until new facilities start generating returns. Selling $8 billion of chips and leasing them back frees up capital for the next wave of purchases without forcing Amazon to slow its buildout.

Not the first to try this

Amazon is not inventing the playbook. Meta used a similar structure last year to fund its giant Hyperion data center, partnering with Blue Owl to move hardware into a separately financed vehicle. The pattern is spreading: as AI capex consumes ever-larger slices of hyperscaler balance sheets, financial engineering is becoming as important to the AI race as the chip engineering itself.

The appeal for investors is straightforward. They get a steady stream of lease payments backed by Amazon — one of the strongest credit profiles in the world — while gaining exposure to AI infrastructure without buying chip stocks directly. For Amazon, the appeal is equally clear: it gets the hardware off its books and recycles capital into the next generation of equipment.

The catch is in the collateral

There is a problem with using AI chips as collateral for long-term debt: they depreciate fast. Nvidia releases a more powerful generation roughly every year, and the secondary market for previous-generation chips can soften quickly. A Grace Blackwell GPU that costs $30,000 today may be worth a fraction of that in three years if the next architecture delivers a meaningful leap in performance per watt. Unlike a Boeing 737 — which retains value over 20 years — or an office building — which can last 50 — an AI accelerator is a depreciating asset with a useful life measured in single digits.

That makes chip-backed debt structurally different from aircraft-backed debt. Lenders are effectively betting that the lease payments will cover the debt service before the collateral loses most of its value. In a rising-rate environment where AI companies are already borrowing at record levels — SpaceX returned to the bond market for $20 billion days after listing, and Anthropic disclosed a nearly $42 billion loss in its IPO filing — the pool of investors willing to take that bet is not infinite.

Why it matters

The $8 billion figure is large, but the signal is larger. When the company operating the world's biggest cloud business starts selling its own hardware and renting it back, it is acknowledging that AI infrastructure costs have outgrown even a $2 trillion balance sheet. This is not a sign of distress — Amazon remains highly profitable and cash-generative — but it is a sign of scale. The AI buildout has entered a phase where no single company's balance sheet is large enough to absorb it all.

The broader implication is that AI infrastructure is becoming a financialized asset class, much like real estate or aircraft leasing. If Amazon's SPV structure works, Microsoft and Google are likely to follow. That would create a new market for AI-chip-backed securities, with pension funds, insurers, and sovereign wealth funds as the ultimate lenders. It also means the AI boom is increasingly funded by debt rather than equity — a shift that amplifies returns in good times and concentrates risk if demand slows.

There is a competitive angle worth watching. Nvidia benefits directly from this trend: if hyperscalers can finance chip purchases through SPVs rather than tying up their own capital, they may be willing to deploy more hardware faster. That is bullish for Nvidia's near-term shipments. But it also creates a latent oversupply risk: if demand for AI compute cools, the chips financed through these vehicles become stranded assets, and the investors holding the debt discover what Grace Blackwell collateral is really worth in a down market.

The timing is also notable. Amazon's move comes days after Broadcom began assembling a $60 billion debt package to finance chip purchases by Anthropic and other customers, and as Anthropic prepares for an IPO targeting a $1.8 to $2 trillion valuation. The AI industry is simultaneously raising equity, issuing debt, and sale-leasebacking hardware — all to fund the same infrastructure buildout. The capital flows are becoming circular, and the system has not been tested in a downturn.

What to watch next

The key indicator is whether Amazon actually closes the SPV deal and at what cost of debt. If investors demand a high yield to hold chip-backed paper, that signals skepticism about collateral durability. If the deal is oversubscribed at attractive rates, it validates the model and opens the door for much larger transactions.

Also worth tracking: whether Microsoft or Google announces a similar structure in the next quarter. If both follow Amazon's lead, AI-chip sale-leasebacks become an industry standard rather than a one-off experiment.

One prediction: by mid-2027, at least two more hyperscalers will have announced AI hardware SPV structures, and the total value of chips held in off-balance-sheet vehicles will exceed $50 billion. The AI buildout is too big for any one balance sheet — the financial system is about to find out whether it is big enough for all of them.