AWS AI Updates: April 18, 2026

1. SageMaker HyperPod Instance Groups Go Multi-Type and Multi-Subnet

AWS. HyperPod users can now specify multiple instance types and subnets in a single instance group, with automatic fallback to lower-priority types when capacity for the preferred option is unavailable. The change eliminates the operational workaround of creating separate groups per combination and pairs with Karpenter autoscaling to detect and provision the best available GPUs for both training and inference workloads. Available everywhere HyperPod runs with EKS orchestration. Source

2. SageMaker JumpStart Adds Task-Aware Optimized Deployments for 30+ Foundation Models

AWS. JumpStart now ships pre-configured deployments tuned for cost, throughput, latency, or balanced performance, covering Llama 3.1/3.2, Phi-3, Mistral, Qwen 2/3, Gemma, and Falcon3 among others. Users pick an optimization target and the service deploys to either SageMaker Managed Inference endpoints or HyperPod clusters, exposing P50 latency and other metrics before traffic is routed. It shortcuts the hand-tuning that most teams otherwise do themselves when standing up a frontier open-weight model. Source