Amazon's chips exceed $20B run rate

Amazon’s chip story stopped being a side project this week. On April 29, Amazon said its custom silicon business had crossed a $20 billion annual revenue run rate, and Andy Jassy argued it would look more like $50 billion if the company counted AWS as a customer the way standalone chip vendors do. That matters because these chips are not just parts inside Amazon’s cloud. They are turning into one of the main ways Amazon controls AI costs, capacity, and product speed. (ir.aboutamazon.com) ### What exactly did Amazon announce? In its first-quarter 2026 results, Amazon said the chips business — Graviton CPUs, Trainium AI accelerators, and Nitro networking and security silicon — exceeded a $20 billion annual revenue run rate and is still growing at triple-digit rates year ove(ir.aboutamazon.com)ing up inside the cloud business already. (ir.aboutamazon.com) ### Why does the $50 billion number matter? The $20 billion figure is the internal accounting view. Jassy’s bigger point was that Amazon consumes a huge amount of its own silicon through AWS, so the business looks smaller only because Amazon is selling to itself. If you treat AWS like an (ir.aboutamazon.com)be about $50 billion. That is a framing move, but it is also a real clue about scale. (aboutamazon.com) ### Which chips are doing the work? Graviton handles general-purpose and CPU-heavy cloud workloads. Trainium is the AI piece — model training and, increasingly, inference. Nitro offloads networking, security, and virtualization functions so AWS servers can do more useful compute. Put together, that gives A(aboutamazon.com)supply at hyperscale. (ir.aboutamazon.com) ### Why is Trainium suddenly such a big deal? Because demand is already outrunning supply. Amazon says Trainium commitments now exceed $225 billion. Trainium2 has about 30% better price-performance than comparable GPUs and has largely sold out. Trainium3 started shipping at the start of 2(ir.aboutamazon.com) broad availability — is already heavily reserved. That is not roadmap theater. That is backlog. (aboutamazon.com) ### Who is actually buying in? Anthropic and OpenAI are the headline names. Amazon says both have made large, multi-year, multi-gigawatt Trainium commitments. Jassy also pointed to Uber using Graviton and Trainium, and Meta committing to tens of millions of Graviton cores for agentic AI workloads. So this (aboutamazon.com). (aboutamazon.com) ### Is this really about beating Nvidia? Not exactly — at least not all at once. Amazon is still deploying Nvidia GPUs at huge scale. But the catch is that hyperscalers do not want their AI economics dictated by one supplier forever. Trainium gives Amazon bargaining power, margin protection, and a way to t(aboutamazon.com) workload it can price, package, and optimize on its own terms. (convergedigest.com) ### Why should product teams care? Because infrastructure assumptions are now product assumptions. If Amazon’s AI services increasingly run on Trainium and Graviton, then cost, latency, availability, and launch timing all depend on chip roadmaps and reserved capacity. That changes planning. Promising an AI-heavy feature without knowing what silicon it will run on starts to look reckless. (aboutamazon.com) ### Bottom line? Amazon is not just renting out cloud anymore. It is building the chips underneath the cloud, then using that stack to pull more of the AI value chain inside the company. Crossing a $20 billion run rate makes that impossible to dismiss as an experiment. (ir.aboutamazon.com)