3nm Trainium3 Chips Pivot AWS to Direct Merchant Silicon

Amazon Web Services has transitioned from a cloud-only hardware operator to a direct merchant silicon vendor. During a June 17 investor briefing, Amazon CEO Andy Jassy confirmed the company will begin selling its custom AI chips as standalone hardware to external data centers. The strategic shift breaks the previous exclusivity of Trainium and Inferentia processors to AWS infrastructure, turning the cloud provider into a direct competitor against Nvidia, AMD, and Intel.

Jassy framed the policy change as a $50 billion revenue opportunity. For enterprise customers and independent colocation facilities, the move unlocks access to Amazon’s proprietary silicon without requiring workloads to live inside an AWS data center.

Trainium3 and Nitromorph Interconnect

The commercial hardware catalog leads with the AWS Trainium3 processor. Manufactured on a 3nm process, the chip utilizes a modular tile architecture designed for foundation model training. This physical design allows external operators to scale deployments from individual server racks up to localized UltraClusters.

AWS is pairing the silicon sales with licensing for its Nitromorph high-speed interconnect technology. By offering Nitromorph externally, third-party data centers can link Trainium3 clusters with the same bandwidth and latency profiles previously restricted to AWS Availability Zones.

For production AI inference workloads, AWS is also selling the Inferentia3 processor. This chip focuses on low-latency, high-throughput execution of pre-trained models.

Hardware Pricing and Availability

AWS targets the Nvidia H200 and B200 deployment market with aggressive cost claims. While specific per-unit MSRPs are not yet public, AWS states the total cost of ownership for training on its hardware will run 35-40% lower than comparable Nvidia architectures.

Rollout happens in two phases:

• Q3 2026: Validation Kits ship to select colocation providers like Equinix and Digital Realty.
• January 2027: General availability for enterprise hardware purchasing.

Decoupling the Software Stack

Selling physical chips requires a software ecosystem that does not assume an AWS cloud environment. To support this, AWS open-sourced the Neutron-SDK. This compiler and runtime suite bridges the gap between common machine learning frameworks and the proprietary hardware.

Developers looking to scale PyTorch training or run JAX workloads can use the Neutron-SDK to execute code directly on standalone Trainium hardware. The SDK ensures that teams building custom models avoid vendor lock-in at the cloud level, isolating their dependency strictly to the silicon and compiler layers.

If you operate private infrastructure to comply with data sovereignty requirements, this hardware availability changes how you reduce LLM API costs. Large-scale AI labs, including Anthropic, have already signaled plans to deploy Trainium3 within localized, private clusters to keep proprietary training data strictly on-premises while leveraging modern 3nm chip efficiencies.