The Chips Stay, the Debt Leaves: Amazon's $8 Billion SPV Would Let It Compute on Hardware It No Longer Owns

The chips will not move an inch. They will keep humming in the same racks, in the same halls, drawing the same power in more than a dozen US data centers across at least five states, including Nevada and Virginia. What changes — if a deal now taking shape comes together — is who owns them. Amazon is in exploratory talks to transfer roughly $8 billion worth of installed Nvidia Grace Blackwell accelerators into a special-purpose vehicle financed by outside investors, and then lease the hardware back from that vehicle, according to a Financial Times report published Friday, with Bloomberg and Reuters subsequently confirming the talks.

Nothing about the transaction is final; terms are unsettled, and neither Amazon nor Nvidia has commented publicly. But the structure itself is the story. It is the clearest signal yet that the AI infrastructure race has entered its financial-engineering phase — a moment when the industry’s constraint is no longer how fast fabs can print chips or how quickly data centers can be wired, but how creatively balance sheets can be arranged around them.

What the deal would actually do

The proposed structure is a sale-and-leaseback executed through a special-purpose vehicle. Amazon would transfer ownership of thousands of already-deployed Grace Blackwell chips — Nvidia’s current flagship lineup, the workhorses behind frontier model training — to a newly established SPV. The vehicle would raise its capital primarily by issuing bonds to external investors, with Amazon offering up to a 10% equity stake in the venture as a sweetener. That detail matters: Amazon itself would hold no equity in the vehicle at all, retaining only usage rights through the leaseback.

For outside investors, the pitch is a new kind of AI exposure. Rather than buying Amazon or Nvidia stock and inheriting everything those businesses do, lenders get direct, collateralized exposure to the hardware itself — with the potential 10% equity stake providing upside beyond interest payments. Their returns depend on one thing above all: Amazon continuing to lease the chips.

The target investor base is telling. Participants in the market expect the SPV’s bonds to secure an investment-grade rating on the strength of Amazon’s own credit profile (S&P: AA, Moody’s: A1, Fitch: AA-), which would open the door to insurance companies and pension funds — conservative pools of capital that rarely touch venture-style AI bets but hold enormous quantities of long-duration money seeking yield. Not coincidentally, that is precisely the market Japanese institutions have been moving into aggressively: Nippon Life Insurance has finalized plans to scale US AI data-center-centered infrastructure financing to ¥2 trillion (about $12.7 billion) by fiscal 2035, and MUFG has committed to an AI infrastructure investment fund exceeding ¥4 trillion (roughly $25.4 billion).

Why now: the $220 billion question

The backdrop is a capex number that would have seemed absurd two years ago. Amazon’s capital expenditure this year is projected to reach roughly $220 billion, with the majority flowing to AWS chip procurement and AI data center expansion. The company committed to purchasing more than 1 million Nvidia GPUs by March 2026, then ordered another 2 million in August, pushing its 2026 orders past 3 million GPUs. On the customer side, it has committed up to $33 billion to Anthropic and $50 billion to OpenAI — commitments that only make sense if the underlying compute exists to serve them.

Amazon has leaned on the corporate bond market to fund this: its March and July 2026 issuances together totaled approximately $75 billion. The March deal, launched into strong demand, was upsized from a planned $37 billion to about $50 billion. By July’s $25 billion issuance, however, investor appetite for longer-dated paper had visibly cooled, with buyers demanding higher yields — a sign that the market’s tolerance for open-ended tech debt expansion is fraying at the margins. An $8 billion SPV is small against that $75 billion; as one-tenth of it, the transaction is less a funding necessity than a proof of concept for a template that can scale.

And Amazon’s free cash flow tells the same story from the other side. The AI capex surge has already crushed AWS-era cash generation, with quarterly free cash flow swinging deeply negative as capital expenditures accelerate past operating inflows. Every chip moved off the balance sheet — while its cost reappears as an operating lease expense — mechanically flatters the metrics that equity and credit markets watch most closely.

A template with a lineage

Amazon did not invent this structure; it is inheriting it. CoreWeave pioneered GPU-collateralized financing, raising $2.3 billion against Nvidia chips as far back as 2023 and closing an $8.5 billion investment-grade GPU-backed facility in March 2026 — the first of its kind. Nvidia itself has since industrialized the model: in August it unveiled a $500 billion compute financing platform built with six institutions including Apollo and BlackRock, proposing to guarantee up to $125 billion of the debt.

The hyperscaler variant arrived in October 2025, when Meta raised $27 billion through a joint venture with asset manager Blue Owl Capital for its Hyperion data center campus in Louisiana — Blue Owl funds hold 80%, Meta keeps 20% and runs operations, and the JV’s assets and liabilities sit outside Meta’s consolidated financial statements. Multiple hyperscale cloud providers are now reportedly exploring similar arrangements. Amazon’s proposal extends the model from buildings to the chips themselves, and from construction finance to a purer form of compute securitization.

The transparency problem

Not everyone regards this as elegant. The Bank for International Settlements, in its March quarterly report, labeled such arrangements “shadow borrowing” — obligations economically equivalent to debt but largely kept off corporate balance sheets. The BIS noted that hyperscalers issued more than $100 billion of corporate bonds in 2025 while simultaneously financing through dedicated vehicles partnered with private credit, funded by bank credit lines. Refinancing pressure at the vehicle level, procyclical swings in private credit appetite, or the triggering of guarantee clauses could each become a channel for transmitting risk exactly when it is least convenient.

The specific fragility of chip-collateralized structures is residual value. Data-center buildings depreciate over decades; AI accelerators are overtaken by successors — Grace Blackwell itself is about to be displaced by Nvidia’s Vera Rubin series — and regulatory filings show Amazon expects each chip generation to remain in service for at least five years. Lenders underwriting five-year economics on hardware with an 18-month reputation cycle are making an implicit bet that inference demand keeps absorbing old silicon. So far it has. The moment it stops, the collateral behind the bonds stops being special.

There is also a competitive reading. If the SPV template works, it decouples AI capacity growth from balance-sheet capacity. The hyperscaler with the best structured-finance desk — not the best technology — could sustain the fastest buildout, at least until the cost of off-balance-sheet capital converges with the on-balance-sheet kind.

What to watch

Three markers will tell us whether this is a one-off or a turning point. First, whether the talks harden into a signed deal and at what spread over Amazon’s own curve — pricing that gap is the market’s honest estimate of how much the structure is worth. Second, whether other hyperscalers follow within quarters, as they did after Meta’s Blue Owl deal. Third, whether regulators or accounting standard-setters respond to the BIS’s framing and move to consolidate “shadow borrowing” back onto the statements it left.

For an industry that has spent three years insisting compute is the new oil, the implication is straightforward: oil fields got securitized long ago. The chips staying in the rack while the debt leaves the balance sheet is not a contradiction — it is what maturity looks like for an asset class. The question is whether the investors now being invited to own the machines understand what they are renting out, and what happens to a five-year lease when the silicon underneath it is two generations old.