Amazon's custom silicon business just crossed a $20 billion annual run rate. CEO Andy Jassy told investors it'd be closer to $50 billion if they counted what AWS uses internally, putting Amazon among the top three datacenter chip businesses on the planet. The portfolio spans Graviton processors, Trainium AI training chips, and Nitro security chips, all growing north of 100% year over year. What started as a way to cut costs on Amazon's own infrastructure is now a legitimate competitor to Nvidia and AMD.
Trainium2 is sold out. Trainium3, which started shipping in early 2026, is nearly fully subscribed. Trainium4 won't launch for another 18 months and chunks of capacity are already reserved.
OpenAI committed to 2 gigawatts of Trainium capacity through AWS, with the deal ramping in 2027. Anthropic locked down 5 gigawatts of capacity. Meta signed up for tens of millions of Graviton cores specifically for agentic AI workloads. Amazon now holds over $225 billion in total Trainium commitments. Whether all of it translates to guaranteed revenue is a fair question, but the reservations alone tell you where demand is heading.
For anyone tracking AI agents specifically, two things matter here. Amazon launched Bedrock AgentCore, a set of infrastructure tools for building and deploying AI agents. The company says an agent gets deployed through AgentCore every 10 seconds. Separately, OpenAI's GPT-5.4 landed on Bedrock in limited preview, with GPT-5.5 coming soon. A break from Microsoft Azure exclusivity, signaling OpenAI needed compute badly enough to diversify its infrastructure as OpenAI breaks Azure exclusivity. Claude Opus 4.7 also hit Bedrock. The platform processed more tokens in Q1 than in all prior years combined, with customer spending up 170% quarter over quarter.
Jassy made one observation worth sitting with. As AI systems shift from answering questions to taking actions, the compute load shifts toward CPUs. Meta picked Graviton for agentic workloads for exactly this reason. Amazon claims it delivers 40% better price performance than competing x86 chips, and 98% of the top 1,000 EC2 customers now use them. The agentic transition is a hardware story too, and Amazon is quietly building the compute layer underneath it.