Amazon held talks to spin off roughly $8 billion worth of Nvidia AI chips deployed in data centers, the Financial Times reported, a figure that would represent one of the largest single restructurings of AI accelerator assets now sitting on a hyperscaler's balance sheet.

According to the FT report, the discussions centered on separating Nvidia chip holdings — housed inside Amazon's data center infrastructure — into a distinct structure. The $8 billion figure refers to the value of the Nvidia AI silicon involved, not to Amazon's broader server fleet or its custom accelerator programs.

The reported talks come as the industry's largest cloud providers face a shared problem: the capital intensity of AI compute. Nvidia's data center GPUs remain the dominant currency of large-model training, and each accelerator carries a price tag in the tens of thousands of dollars before a server is racked, powered, or cooled. For a company of Amazon's scale, an $8 billion position in a single vendor's chips is material even against a capital budget that runs to tens of billions of dollars annually.

A spin-off structure of the kind described in the report would, in principle, move that hardware cost off Amazon's books while keeping the compute capacity available to its cloud business. The FT report did not state whether the talks reached agreement, what entity would take ownership of the chips, or on what financial terms a transaction would be structured. Amazon and Nvidia had not publicly confirmed the discussions at the time of the report.

The report lands amid intensifying scrutiny of how hyperscalers finance AI infrastructure. Cloud operators across the board have turned to off-balance-sheet vehicles, joint ventures, and vendor-financed deals to stretch capital budgets as GPU demand outruns supply. Nvidia's own results have made the scale of this spending visible: the company's data center revenue has climbed to levels that dwarf the rest of the semiconductor industry, driven overwhelmingly by a small group of U.S. cloud providers.

Amazon occupies an unusual position in that group. AWS designs its own AI accelerators — the Trainium and Inferentia product families — and offers them alongside Nvidia's GPUs to customers. The reported talks, as described by the FT, concern only the Nvidia silicon within Amazon's data centers, not its in-house chip fleet.

A separation of Nvidia chip assets would also carry customer implications. AWS customers who rent Nvidia GPU instances would, under a spun-off structure, effectively be consuming compute owned by a different entity. Whether that changes pricing, availability guarantees, or procurement terms is among the questions the report leaves open.

Nvidia's supply position adds context to any such negotiation. Its flagship data center GPUs remain allocation-constrained, and access to large volumes of current-generation silicon has become a strategic lever for hyperscalers negotiating with both Nvidia and their own enterprise customers. Hardware that is already deployed, powered, and generating revenue is among the most defensible assets in that bargaining environment — a factor that would weigh on how any spin-off is valued.

The FT report did not specify a timeline for the talks or whether they remain active. If a transaction of the reported $8 billion scale proceeds, it would set a precedent for how hyperscalers account for the tens of billions of dollars in AI accelerators they are slated to deploy over the coming years, and could reshape how investors read capital-expenditure guidance across the sector.