Amazon custom silicon hits $20B ARR; Trainium3 nearly sold out, Trainium4 largely reserved

Amazon's custom silicon portfolio—spanning Graviton CPUs, Trainium AI accelerators, and Nitro DPUs—has surpassed $20 billion in annual revenue run rate as of Q1 2026, growing at triple-digit rates year-over-year. CEO Andy Jassy disclosed in his April 2026 shareholder letter that if the chip business were valued as a standalone entity selling to both AWS customers and third parties, its annual revenue run rate would be roughly $50 billion. Amazon deployed over 2.1 million AI chips in the past 12 months, positioning custom silicon as a top-three datacenter chip business globally.

Trainium capacity is nearly fully allocated. Trainium2 has largely sold out, with 30% better price-performance than comparable GPUs. Trainium3, which entered production in early 2026 and offers 30-40% better price-performance than Trainium2, is nearly fully subscribed. Trainium4, still approximately 18 months from broad availability, has already been significantly reserved. Notably, Trainium4 will feature NVLink Fusion interoperability with NVIDIA GPUs, allowing customers to combine Trainium accelerators with NVIDIA silicon in the same system. Major commitments include ~2 GW from OpenAI and up to 5 GW from Anthropic.

Graviton CPUs have achieved broad enterprise adoption: 98% of the top 1,000 EC2 customers now use Graviton, with more than 100,000 total customers. Graviton delivers up to 40% better price-performance than x86 processors. Meta recently signed a multibillion-dollar deal for Graviton5 chips, and AWS is exploring selling Trainium chips directly to external customers outside of AWS—a potential shift from cloud-only service delivery to merchant hardware sales. AWS Bedrock inference services now run most workloads on Trainium, reflecting the maturation of custom silicon economics in production AI inference.