Amazon is shifting its hardware strategy into high gear, moving beyond its traditional cloud ecosystem to challenge Nvidia directly. The tech giant is in discussions to sell its custom-designed Trainium AI accelerator chips and full server racks directly to third-party clients for use in independent, external data centers.

The strategy aggressively highlights a cost-per-performance advantage, with Amazon offering its homegrown silicon at deep discounts compared to Nvidia’s market-dominating hardware.

The Strategic Pivot: Shifting to Direct Sales

For years, Amazon designed its custom silicon (including the Trainium, Graviton, and Nitro lines) exclusively for internal infrastructure, allowing businesses to rent the computing power strictly through Amazon Web Services (AWS).

The decision to start offering full racks of Trainium chips to outside buyers—confirmed by Amazon AI chief Peter DeSantis—marks a major transition from a cloud utility model into a direct semiconductor vendor layout.

• The Financial Run Rate: Amazon’s total custom-silicon division has already quietly scaled into a powerhouse, crossing a $20 billion annual revenue run rate while expanding at a triple-digit pace.
• The $50 Billion Target: Amazon CEO Andy Jassy previously projected that expanding its chip business to encompass both internal AWS deployments and external direct sales could eventually propel the operation into a $50 billion annual revenue engine. That scale would instantly place Amazon on par with traditional semiconductor heavyweights like Intel.

Undercutting Nvidia on Price and Accessibility

The central pitch to external enterprise buyers is a massive relief on their capital expenditure budgets. Amazon has aggressively targeted Nvidia’s client base by pricing its infrastructure significantly lower.

AWS has actively pitched customers on Trainium instances that offer equivalent computing power for complex training and inferencing workloads at roughly three-quarters of the cost of Nvidia’s highly sought-after H100 processors—effectively a 25% discount.

This aggressive pricing framework serves two strategic purposes:

1. Targeting the Commercial Enterprise Tier: While top-tier AI labs often demand Nvidia’s maximum-performance GPUs for frontier models, everyday enterprise buyers looking to train smaller, domain-specific models or handle reasoning and chain-of-thought (CoT) inferencing don’t always need Nvidia’s premium price tier. Amazon’s discounted entry point democratizes access for mid-market corporate data centers.
2. Mitigating “CUDA Lock-In”: Nvidia’s primary competitive moat has long been its proprietary CUDA software framework, which binds developers to its hardware. By selling hardware directly, Amazon offers a cheaper hardware alternative that can incentivize companies to invest the development time required to break away from Nvidia’s software ecosystem.

High International Demand for “Sovereign AI”

Beyond pure price undercutting, Amazon is finding immediate traction due to severe global shortages of standard AI processing hardware and a growing geopolitical push for localized computing resources.

According to DeSantis, a major driver behind the talks is intense international demand—particularly from European corporate entities and governments facing regulatory and political pressure to establish sovereign AI infrastructure. Purchasing Trainium hardware directly allows these organizations to build locally controlled data footprints that remain independent of external cloud networks.

The broader market has responded favorably to the expansion. Despite the looming threat to Nvidia’s near-total market monopoly, industry analysts suggest the overall AI pie is expanding fast enough to support both ecosystems. Highlighting this rapid market growth, even as Amazon expands its Trainium sales, it remains one of Nvidia’s largest institutional buyers, maintaining an active commitment to deploy over one million Nvidia GPUs across its network.