Graviton hits $20B run rate

Amazon’s custom chips just turned from an AWS feature into a business big enough to matter on its own. On May 1, Amazon said its chips business had passed a $20 billion annual revenue run rate and was still growing at triple-digit rates year over year. That number includes Graviton CPUs, Trainium AI accelerators, and Nitro networking hardware, not Graviton alone. But Graviton is a huge part of why this matters — because it shows AWS is no longer just renting Nvidia-powered AI clusters. It is increasingly steering customers onto Amazon-designed silicon for the boring but expensive parts of modern computing. (ir.aboutamazon.com) ### What exactly hit $20 billion? The key correction is simple: Amazon did not say “Graviton alone” hit a $20 billion run rate. Andy Jassy said the broader chips business — Graviton, Trainium, and Nitro together — topped that level in Q1 2026. He also said the business grew nearly 40% q(ir.aboutamazon.com)rate would be closer to $50 billion. That tells you the internal transfer-pricing issue is masking how much silicon AWS is really consuming. (ir.aboutamazon.com) ### Why is Graviton the interesting piece? Trainium gets the AI headlines, but Graviton is the chip that spreads everywhere. It runs general compute — web apps, databases, containers, analytics jobs — and AWS pitches it on one simple promise: better price-performance than comparable x86(ir.aboutamazon.com)preprocessing, serving logic, and agent loops. A lot of that lands on CPUs, and cheaper CPUs move cloud bills in a big way. (aws.amazon.com) ### Why are AI customers suddenly buying so much CPU? Because agentic AI is surprisingly CPU-hungry. Meta said on April 24 that it would bring tens of millions of AWS Graviton cores into its compute portfolio, explicitly tying that decision to agentic AI’s growing CPU demands. Amazon made a similar point in its own explainer this week — Trainium handles model training and i(aws.amazon.com)nding application and infrastructure work. Basically, the more AI systems behave like software workers calling tools and services, the more conventional compute you need around the GPU core. (about.fb.com) ### Is this just Meta? No — and that is the other reason the number jumped so fast. Amazon also expanded its Anthropic partnership in late April, saying Anthropic plans to spend more than $100 billion over the next decade on AWS technologies, including current and future Trainium generations and tens of millions of Gr(about.fb.com)uration. That helps AWS plan capex with more confidence, even if it also raises concentration risk. (aboutamazon.com) ### Why does the February comparison matter? Because it shows how fast this curve is steepening. In Amazon’s fourth-quarter 2025 results, the company said Trainium and Graviton together had just crossed a $10 billion annual run rate. Three months later, the broader chips business was above $20 billion. The categorie(aboutamazon.com) the same bucket. But directionally, it still shows acceleration, not just steady growth. (ir.aboutamazon.com) ### What does this change for cloud competition? It raises the bar. AWS now has a fuller stack: Nvidia when customers need it, Trainium for Amazon’s own AI economics, and Graviton for the huge layer of CPU work surrounding AI and standard cloud apps. That means Amazon can compete on more than raw (ir.aboutamazon.com) that makes the rest of AWS’s AI buildout easier to afford. (aws.amazon.com) ### Bottom line? The clean takeaway is not “Graviton alone hit $20 billion.” It is that Amazon’s custom silicon business has become massive, fast-growing, and strategically central — and Graviton is the piece pulling ordinary cloud compute and agentic AI infrastructure into the same flywheel. (ir.aboutamazon.com).aspx))