Amazon Trainium Chips Threaten Nvidia Lead

Amazon Trainium chips are moving toward external sales after CEO Andy Jassy signaled in an April 2025 Letter to Shareholders that AWS could sell full server racks of its AI accelerators to third-party data centers, intensifying competition in the data-center AI-accelerator market.

Trainium Commercialization Push

AWS is actively discussing selling its custom Trainium accelerators and full server racks directly to external data-center customers, expanding beyond internal use. In the shareholder letter, Jassy described selling Trainium racks to third parties as feasible given strong demand for Amazon’s in-house chips.

Amazon introduced the first Trainium training accelerator in 2020 to reduce model-training costs compared with third-party GPUs. The company has since developed a second-generation chip, Trainium 2, as it scales deployments.

Jassy said Amazon’s custom-chip portfolio—including Trainium, Inferentia inference accelerators, Graviton CPUs, and Nitro offload silicon—has surpassed an annual revenue run rate of more than $20 billion, growing at triple-digit year-over-year rates. He noted this figure understates the economics because much of the chip value is realized through AWS compute services rather than standalone hardware sales.

He added that if monetized like a standalone hardware vendor, the chip portfolio could reach about a $50 billion annual revenue run rate. Trainium at scale should save Amazon tens of billions of dollars in annual capital expenditures and improve AWS operating margin by several hundred basis points compared with relying on external chips.

AI Chip Market Dynamics

Nvidia still controls roughly 80% of the AI accelerator market for data centers, maintaining a dominant position despite hyperscalers developing their own silicon. Nvidia’s data-center revenue topped $194 billion in its latest fiscal year, supported by a backlog exceeding $1 trillion for next-generation Blackwell and Vera Rubin architectures. This scale gives Nvidia significant pricing and ecosystem leverage with cloud and enterprise buyers.

Alphabet is pursuing a similar strategy, marketing its Tensor Processing Units (TPUs) more broadly so cloud customers can buy its accelerators alongside software and services, mirroring Nvidia’s hardware-plus-ecosystem model. This approach, combined with hyperscalers’ internal silicon, increases competitive pressure in the specialized accelerator market.

A MarketsandMarkets report projects the U.S. AI-chip market will grow from about $61.9 billion in 2025 to $173.1 billion by 2032, a 15.8% compound annual growth rate. Nvidia, Intel, and AMD are identified as leading players. The market’s expansion creates room for multiple suppliers even as shares remain concentrated.

The AI inference segment, which covers running models rather than training them, has roughly doubled recently, with Nvidia’s share rising to about 74% from 66%. Growth in inference workloads motivates hyperscalers to invest in custom accelerators tailored to model-serving economics.

If AWS proceeds with selling Trainium externally, it could reprice procurement for data centers by offering an alternative to GPU incumbents and expanding hardware choices for large AI workloads. Alongside hyperscalers’ internal silicon and cloud providers’ TPU commercialization, this shift intensifies competition as agentic AI drives larger, more data-intensive workloads.