Amazon's custom chip operation has become a major business. The company's Trainium AI training chips are now generating revenue at a $20 billion annual pace, as disclosed by CEO Andy Jassy.
If Amazon's custom silicon division were standalone, its total revenue across Trainium, Graviton, and Nitro chips would approach $50 billion. That would place it among the world's largest chipmakers.
The growth is not incremental. Amazon reported triple-digit year-over-year revenue increases for its custom chip business.
Amazon has secured more than $225 billion in committed Trainium revenue through multi-year deals with AI companies Anthropic and OpenAI.
AWS Bedrock, Amazon's managed AI service built on Trainium infrastructure, now counts over 125,000 customers. Roughly 80% of Fortune 100 companies use the platform.
Amazon began shipping Trainium3 in early 2026, delivering a 30-40% improvement in price-performance over the previous generation.
Reports indicate Amazon is exploring direct sales of Trainium chips to third-party data centers. This would change Amazon's distribution model from a cloud-services-only approach to something closer to how Nvidia operates.
The explosive demand for custom AI silicon confirms that compute scarcity is structural. A world where Trainium chips are available outside AWS would shift the compute supply curve, with pricing implications across the AI infrastructure stack.
Amazon's $225 billion commitment pipeline suggests alternatives to Nvidia are arriving faster than the market expected. This could compress margins across the entire AI chip sector while expanding the total addressable market.