Amazon Moves $8 Billion of NVIDIA Chips Off Its Books and Into a New AI Asset Class

Amazon is moving about $8 billion of NVIDIA chips into a newly created outside entity that now holds the hardware on its books. Amazon would keep using the chips under a leaseback, with up to 10% of the entity's equity offered to outside investors, according to the Financial Times, citing people familiar with the matter.

The chips are thousands of NVIDIA Grace Blackwell processors installed across more than a dozen U.S. data centers, per Reuters. Bloomberg independently confirmed the FT report on October 2. The structure is a special-purpose vehicle, a financing entity created to hold a specific pool of assets and funded largely by debt from outside investors. Amazon would service the debt through lease payments while retaining operational use of the chips.

A hyperscaler is treating its AI accelerators the way airlines treat jetliners and telecoms treat cell towers: parking the capital-intensive hardware in a separately financed entity so the operator's own balance sheet stays lighter and a wider circle of investors can take on the asset's depreciation risk. The chips stay plugged in. The depreciation lives somewhere else.

This is the largest single instance so far of a pattern that has been building for roughly a year. London-based Nuway Capital issued the first GPU-backed bond in 2025. NVIDIA announced in August a separate initiative with Blackstone, Apollo, and KKR to create asset-backed financing structures for its chips, a plan that the FT and others have pegged at roughly $500 billion in scale, per Wccftech. Investopedia noted the Amazon move is the first hyperscaler-scale test of whether the model works at the top of the demand curve.

Sale-leaseback is an old tool. Airlines have parked aircraft in separate trusts for decades. Telecoms have offloaded cell towers into third-party operators for the same reason. The novelty in 2026 is the asset. AI accelerators are depreciating in a market that is still defining the curve, with each generation more capable than the last and a roadmap of successors pushing residual values in directions no comparable asset class has had to price before.

That uncertainty is exactly what makes some investors nervous. The FT's reporting flagged that Wall Street is growing wary of NVIDIA's broader financing plan, a sentiment echoed in a Reuters dispatch cited by aggregators. The concern is residual value: if today's Grace Blackwell chips depreciate faster than the SPV's debt schedule assumes, the outside equity holders absorb the loss. NVIDIA's hardware roadmap has historically held value better than consensus feared, with successive generations often commanding premium pricing even as predecessors aged. The bet on a multi-year useful life is now being made at trillion-dollar scale, and the people funding it are not the same people running the data centers.

A few practical pieces are still missing. Amazon, AWS, and NVIDIA have not commented on the talks. The SPV's debt size, coupon, maturity, and rating are not public. The identity of the outside investors, and whether Microsoft, Google, or Meta are pursuing comparable structures, has not been disclosed. Dealroom is tracking the deal as still in the "exploring" stage, and the Reuters wire as republished by the Island Packet describes the talks as preliminary.

What the structure does, if it works, is expand the pool that can fund AI's next build cycle. Hyperscalers no longer have to absorb every dollar of GPU capex on their own balance sheets. Pension funds and credit funds can take slices of the depreciation exposure in return for yield. That makes more chips deployable, faster, than the four largest cloud companies' combined free cash flow would allow on its own. AWS customers see no operational change on day one: the same racks run the same workloads. Competitors face a different problem, because matching that capacity with on-balance-sheet purchases would mean funding depreciation out of operating cash flow.

The first concrete milestone is whether Amazon confirms the structure and discloses its debt terms. The second is the first quarterly disclosure from any GPU SPV showing actual chip utilization and lease coverage, the number that will tell the market whether the financing matches the physics.