Amazon Looks to Offload $8B of Nvidia AI Chips From Its Balance Sheet - AOL.com

Tech Business

Amazon Seeks to Move $8B of Nvidia AI Chips Off Its Books

Amazon is looking to shed roughly $8 billion of Nvidia AI accelerators from its balance sheet, a move that would ease depreciation strain amid record AI capex and a growing Trainium push.

By
Grace Kim
Filed
Channel
Tech Business
Read
3 min read

Amazon is looking to offload roughly $8 billion worth of Nvidia AI chips from its balance sheet, according to a report carried by AOL.com. The figure, if confirmed in regulatory filings, would represent one of the largest single-vendor GPU positions ever held by a cloud provider — and a striking signal about how hyperscalers now manage their AI capital expenditure.

The report gives no detail on which Nvidia product families sit inside the $8 billion position. Nvidia's current data-center lineup spans the H100 and H200 accelerators built on the custom 4NP node at TSMC, alongside the Blackwell generation — the B200 and GB200 systems — that began volume shipments to cloud customers in late 2024 and early 2025. Amazon is simultaneously deploying its own Trainium and Inferentia silicon, which it designs in-house and manufactures at TSMC, a parallel accelerator program that has grown into a meaningful counterweight to its Nvidia purchases.

Why a hyperscaler would sell GPUs

The accounting logic matters as much as the hardware. Accelerators held as owned assets depreciate over their useful life — typically three to five years for AI silicon given rapid generational turnover. A position of $8 billion carries a heavy depreciation load precisely as Nvidia's cadence compresses the economic life of each generation. Moving chips off the balance sheet, whether through sale, lease structures, or transfers to financing partners, would shift that depreciation burden and free capital for new capacity.

The move also fits a broader pattern among hyperscalers. Microsoft, Google, and Amazon have all pursued sale-leaseback and vendor-financing structures for AI infrastructure, and Amazon has a long history of securitizing equipment, including server fleets, through financing vehicles. An $8 billion divestment would rank among the largest such transactions in the sector's buildout to date.

For Amazon specifically, the timing aligns with a period of record infrastructure spending. The company has repeatedly raised its capital-expenditure guidance to fund AWS AI capacity, and CEO Andy Jassy has said the majority of that capex flows to AI infrastructure for AWS and its retail operations. Shedding part of the Nvidia position would let Amazon keep expanding capacity while easing the balance-sheet strain that investors increasingly scrutinize.

Supply-chain context

The reported divestment comes as demand for top-tier accelerators continues to outrun supply. Nvidia's Blackwell chips remain allocation-constrained, with cloud providers competing for finite advanced-packaging capacity at TSMC — particularly CoWoS output that binds GPU dies to high-bandwidth memory stacks from SK hynix, Micron, and Samsung. In a tight market, an owner of $8 billion in Nvidia silicon holds a liquid asset; secondhand and re-allocated H100 systems have commanded strong prices on gray-market channels since 2023.

That liquidity cuts both ways for Nvidia. A large hyperscaler releasing accelerator inventory could add visible supply into a resale channel that Nvidia does not control, potentially pressuring pricing for prior-generation parts. Nvidia has not commented on the report, and neither has Amazon.

Amazon's own silicon strategy shapes the calculus as well. The company's Annapurna Labs unit has shipped multiple Trainium generations, and AWS has committed to large internal deployments of Trainium2 instances, with a next generation on its public roadmap. Every rack converted to in-house silicon reduces future Nvidia purchases — and makes existing GPU inventory a candidate for monetization rather than long-term retention.

What to watch

The report leaves key questions open: the transaction structure, whether the chips go to a financial buyer or an end user, and which accelerator generations the $8 billion covers. Amazon's next 10-Q filing, along with any disclosure of financing subsidiaries or lease obligations tied to compute assets, will offer the first hard evidence. If the company completes the divestment, expect other hyperscalers holding multi-billion-dollar GPU positions to study the template closely as depreciation pressure builds across the industry's AI fleets.

Source: Google News: AI chips

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering industry trends and analytics at Chip Dispatch.

157 articles

Related articles

« Previous articleNext article »