Amazon disclosed that its custom silicon business—comprising Graviton CPUs, Trainium AI accelerators, and Nitro networking chips—crossed a $20 billion annual revenue run rate in Q1 2026, growing at triple-digit rates year-over-year with nearly 40% sequential growth. AWS reached a $150 billion annualized run rate (28% YoY growth, fastest in 15 quarters), with AI-related revenue alone exceeding $15 billion annualized. CEO Andy Jassy estimated that if the chip business were treated as a standalone vendor selling to external customers, the annual run rate would approach $50 billion.
Supply is constrained, not demand. Trainium2 is largely sold out; Trainium3 began shipping in early 2026 and is nearly fully subscribed. Much of Trainium4 (18 months from broad availability) has already been reserved. Amazon holds over $225 billion in Trainium revenue commitments from customers including OpenAI (~2 GW) and Anthropic (~5 GW). Trainium2 delivers ~30% better price-performance than comparable GPUs; Trainium3 adds another 30–40% improvement. Amazon deployed over 2.1 million AI chips in the past 12 months and plans to deploy over 1 million Nvidia GPUs starting in 2026—a clear signal that custom silicon and third-party accelerators are complementary, not competitive.
For architects: Amazon has crossed from internal-use-only to a vertically integrated competitive position. The $20–50B run rate makes Amazon's custom silicon business a top-3 global data center chip operation by its own measure. The full $50B construct assumes merchant sales; actual external revenue is still transactional. Watch whether AWS begins disaggregating chip pricing separately in SEC filings (currently lumped into cloud services). The margin arbitrage—using custom silicon to reduce AWS's own capex by tens of billions annually while expanding market share—is the real lever.