Uploaded September 2026 | Updated November 2026, 1 hour ago
Scaling agentic reinforcement learning requires thousands of scheduling decisions per second, which creates unique infrastructure challenges.
At Ray Summit 2026, Poonam Lamba and Nisha Johnson from Google present a reference architecture using Ray, Kubernetes, and Google TPUs: how Ray enables sub-second scheduling and fault tolerance, how GKE time-slicing and smart inference scheduling maximize accelerator utilization, and how PyTorch and JAX stacks run on TPUs, closing with RL performance benchmarks and Google's product roadmap.
You'll leave with a reference architecture for RL at scale.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
Scaling agentic reinforcement learning requires thousands of scheduling decisions per second, which creates unique infrastructure challenges.
At Ray Summit 2026, Poonam Lamba and Nisha Johnson from Google present a reference architecture using Ray, Kubernetes, and Google TPUs: how Ray enables sub-second scheduling and fault tolerance, how GKE time-slicing and smart inference scheduling maximize accelerator utilization, and how PyTorch and JAX stacks run on TPUs, closing with RL performance benchmarks and Google's product roadmap.
You'll leave with a reference architecture for RL at scale.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com










