Amazon is seeking to offload $8 billion worth of Nvidia chips to investors, the Financial Times reported, in a move that would put a hard number on how hyperscalers are now managing the capital intensity of AI accelerator fleets.

The reported figure — $8 billion — is the concrete anchor of the story. According to the FT, Amazon is looking to transfer Nvidia chip holdings to investors rather than carry the full exposure on its own balance sheet. Reuters relayed the report; Amazon has not, according to the available reporting, confirmed the terms, the counterparty structure, or the identity of any participating investors.

What makes the number notable is its scale relative to the AI infrastructure buildout now under way across the top US cloud providers. Nvidia's data center GPUs — the Hopper and Blackwell families that power most large-model training and inference — are the single most expensive line item in hyperscaler capex, and demand for them has consistently outstripped supply since late 2022. An $8 billion position in those chips represents a meaningful slice of accelerator inventory for any single operator.

The mechanism matters as much as the money. Offloading chips to investors is, in effect, a financing structure: the buyer takes ownership of the hardware, and the hyperscaler typically retains operational use of it through lease or capacity arrangements. That converts a capital expenditure into an operating cost, frees balance-sheet capacity for further buildout, and shifts depreciation and obsolescence risk — a real concern for GPUs whose economic life is shortened by each new product generation — onto the investors.

For Amazon, the timing aligns with its stated trajectory. The company has committed to roughly $100 billion in capital spending for 2025, with the overwhelming majority directed at AI infrastructure for Amazon Web Services. Any structure that lets AWS keep GPUs running in its data centers while moving some of the hardware cost off its books would directly extend how far that capex budget stretches.

The FT report does not, in the available summary, specify which Nvidia products the $8 billion covers, whether the chips are deployed or yet to be delivered, or at what valuation relative to list price the transfer would occur. Those details will determine whether the transaction reads as straightforward vendor financing or as a harder test of what AI accelerators are worth as assets in their own right.

There is a broader signal here for the semiconductor supply chain. If financial investors become direct owners of Nvidia silicon — even indirectly, through structures arranged by cloud operators — GPU demand begins to rest partly on capital markets' appetite for hardware yields, not only on cloud customers' willingness to pay for compute. That adds a new class of buyer into a supply chain already defined by allocation constraints at TSMC's leading-edge CoWoS packaging nodes, where advanced GPU output remains the gating factor.

It also sets a precedent other hyperscalers can follow. Microsoft, Google, and Meta are each carrying comparable or larger accelerator programs, and each faces the same accounting question: whether the returns on AI compute justify holding tens of billions of dollars in depreciating chips outright. Amazon moving first, if the reported deal proceeds, would give the industry a price reference for GPU-backed financing.

All figures and intentions in this story come from the Financial Times' reporting as relayed by Reuters; neither Amazon nor Nvidia has publicly confirmed the arrangement. Watch for confirmation in Amazon's next filings and earnings calls, where any financing structure of this size would be difficult to leave undisclosed.