Amazon's custom AI chip unit got a rare public spotlight this week when TechCrunch toured the company's Austin lab — and the numbers inside are striking.

Claude Runs on a Million Trainium Chips

Anthropic's Claude is already running on over 1 million Trainium2 chips deployed across AWS. That figure is part of a total of 1.4 million Trainium chips across all three generations now in service. Trainium2 also handles the majority of inference traffic on Amazon's Bedrock service, which underpins thousands of enterprise AI applications.

The pivot from training to inference is deliberate. Inference — generating model outputs in real time — is now the primary compute bottleneck in the industry, and Amazon's team has tuned Trainium explicitly for it.

OpenAI Is Next, at Gigawatt Scale

As part of AWS's $50 billion deal with OpenAI announced last month, Amazon has committed to supplying 2 gigawatts of Trainium computing capacity. AWS becomes the exclusive cloud provider for OpenAI's new agent builder platform, Frontier. Microsoft has since disputed whether the deal violates its own partnership with OpenAI.

Competing on Price and Portability

Amazon says Trainium3 UltraServers cost up to 50% less than comparable cloud instances, with custom Neuron switches enabling low-latency chip-to-chip communication across entire server clusters. The team also addressed the long-standing switching cost problem: models built with PyTorch now run on Trainium with what engineers describe as a one-line code change and a recompile.

AWS also recently announced a partnership with Cerebras Systems, integrating Cerebras inference chips alongside Trainium in a combined stack targeting ultra-low latency. The chip lab traces its roots to Annapurna Labs, an Israeli startup Amazon acquired in 2015 for around $350 million.