Happy Hour: AWS Trainium Frontier results on a fixed compute budget, by the teams who built them

​Most architectures we train were shaped by the hardware we happened to have, and each was tuned for one goal at a time. What would you design differently if those bottlenecks went away, and reasoning quality counted as much as training efficiency? Over nine weeks, teams of researchers and developers competed, used AI agents to co-design model architectures and custom NKI kernels, then trained small language models on a fixed compute budget on AWS Trainium, a purpose-built AI chip for training and inference at scale, where familiar assumptions about efficiency no longer hold. Join us during NeurIPS 2026 and AI Week for drinks as the top teams present their designs and results, in person or on video, and share their code and findings. Stay for conversation about reasoning-ready small models, efficient training, agentic model and kernel co-design, and how far you can go when the constraints change.