From Silicon to Services: Optimizing Every Layer of AI Infrastructure for Token Economics at Scale

16 Sep 2026
Main Stage
Main Stage

The cost of AI is growing exponentially — but so is the opportunity. The winners won't be those with only the fastest chip. They'll be those who deliver the best token economics — and that requires purpose-built infrastructure, end-to-end, from AI chips to services. In this keynote, we'll show how AWS built exactly that: Trainium3 AI Chips, Next Gen UltraServers, Neuron SDK, Elastic Fabric Adapter, SageMaker AI, and Amazon Bedrock — each layer optimized unlocks new capabilities in frontier models, and it makes models more efficient and cost effective to scale. We'll demonstrate how this co-developed system delivers real-world token economics across diverse model architectures, heterogeneous compute, and various deployment options. Purpose-built AI infrastructure is ready. This is why Anthropic and OpenAI choose AWS for their most demanding AI workloads. Come see what's possible.

Speakers
Peter DeSantis
Peter DeSantis, SVP, Foundational AI Models, Custom Silicon & Quantum Computing - Amazon