Amazon's custom AI chip division landed its two biggest customers this decade within the same news cycle. AWS announced a $50 billion investment deal with OpenAI, making AWS the exclusive provider for OpenAI's new AI agent builder, Frontier. The commitment includes 2 gigawatts of Trainium computing capacity. Separately, Anthropic's Claude already runs on over 1 million Trainium2 chips, with Trainium2 handling the majority of inference traffic on Amazon's Bedrock service.

This is not a chip company that emerged last quarter.

The Long Build

Amazon acquired Israeli chip designer Annapurna Labs in January 2015 for approximately $350 million. That team became the foundation for what is now a multibillion-dollar business, per Amazon CEO Andy Jassy's December 2025 statement. Across all three generations, 1.4 million Trainium chips have been deployed.

Project Rainier, one of the largest AI compute clusters in the world, went live in late 2025 with 500,000 Trainium2 chips dedicated to Anthropic. That cluster alone represents a significant fraction of total Trainium2 deployment.

Trainium3: The Hardware Details

Trainium3 shipped in December 2025. It is a 3-nanometer chip produced by TSMC. Previous Trainium generations used air cooling. Trainium3 switched to liquid cooling. New Neuron switches allow every Trainium3 chip to communicate with every other chip in a mesh configuration, which Amazon says reduces latency.

Amazon claims Trn3 UltraServers cost up to 50% less to run compared to classic cloud servers for comparable performance. "Up to" is doing real work in that sentence.

The Migration Story

AWS says migrating from Nvidia to Trainium now requires "basically a one-line change" in PyTorch. If accurate, that removes the main practical barrier to adoption. Trainium previously required custom tooling that discouraged switching.

This month, AWS also announced a partnership with Cerebras Systems, integrating Cerebras inference chips on servers running Trainium for low-latency AI performance. Cerebras is known for wafer-scale chips optimized for fast inference. The combination suggests AWS is building a tiered inference stack rather than betting on one architecture for every workload.

Apple's Quiet Endorsement

In 2024, Apple publicly described using Amazon's Graviton and Inferentia chips, with a nod to Trainium. Apple does not typically discuss its cloud infrastructure. The mention was brief, but the company's standards for supplier endorsement are not.

What's Next

The Trainium chip team is currently designing Trainium4. No timeline or specs were disclosed.

The customer list now includes Anthropic at scale, OpenAI under a $50 billion exclusive agreement, and Apple in some capacity. AWS started this with a $350 million acquisition eleven years ago. The chip team is still designing the next generation. The compounding is straightforward.

Source: Techcrunch