AWS EKS now supports massive 100K-node Clusters

The new EKS announcements from the AWS Summit in New York should excite the DevOps professionals looking to deploy massive cluster nodes. 

Amazon’s Elastic Kubernetes Service just rewrote the record books. Clusters created with the new “ultra‑scale” flag can now span 100 000 worker nodes, enough to marshal 1.6 million AWS Trainium accelerators or 800 000 NVIDIA GPUs in a single control plane. The limit—announced this morning at AWS Summit New York—catapults EKS far beyond its previous 10 000‑node ceiling and squarely into the realm of trillion‑parameter model training and other frontier‑AI experiments that demand every accelerator live under the same Kubernetes namespace.

Amazon’s container team says the breakthrough came from repartitioning etcd writes, hardening the API server and introducing a dynamic sharding scheme that boosts control‑plane throughput five‑fold while remaining 100 % upstream‑conformant. Customers spinning up an ultra‑scale cluster in us‑east‑1 or us‑west‑2 today can mix P5/P6 GPU instances with Trn2 chips, wire them into SageMaker HyperPod or Bedrock AgentCore, and schedule jobs with vanilla kubectl; the upgrade also trickles down to services that ride on EKS—meaning SageMaker, Batch and EMR inherit the headroom automatically.

AWS positions the move as a hedge against sprawling multi‑cluster complexity by keeping all pods in one logical fleet, data scientists avoid cross‑cluster networking tax and can saturate hundreds of thousands of accelerators during gradient‑sync phases without bespoke plumbing. Analysts reckon the new scale tier gives AWS a decisive lead over Google GKE’s 50 000‑node limit and Azure AKS’s 30 000, reinforcing the cloud giant’s pitch that “if your model can dream it, we can cluster it.” Enterprises chasing GPT‑class workloads, and startups building open‑weight LLMs, now have the option to stay inside Kubernetes instead of shifting to speciality schedulers—turning EKS into the biggest single playground for AI compute on the planet.