AWS AI Updates: October 28, 2026

1. SageMaker AI Training and Processing Jobs Can Lazy-Load SOCI-Indexed Container Images

AWS added Seekable OCI (SOCI) indexing support to Amazon SageMaker AI training and processing jobs. When a job’s image in Amazon ECR has a SOCI index, SageMaker AI fetches only the files needed to start the container and streams the rest in the background, and it falls back to a standard download if the index cannot be read. In AWS benchmarks across eight images from 1.8 GB to 35.4 GB, image download time dropped from as long as 7 minutes 17 seconds to 3 to 11 seconds, and GPU idle time during image pulls fell by up to 90%. The first training epoch ran 3 to 16 seconds longer while files were fetched on first access, with later epochs unchanged, for a net saving of 1 to 6 minutes per job. AWS publishes pre-indexed Deep Learning Containers with a -soci tag suffix for PyTorch Training, TensorFlow Training, vLLM, and SGLang, and custom images can be indexed with nerdctl image convert --soci. No changes to job definitions are required beyond the image tag. Source

2. SageMaker AI MLflow Apps Move to MLflow 3.16 With a Redesigned Trace Explorer

AWS updated Amazon SageMaker AI MLflow Apps to MLflow 3.16. The release makes a redesigned trace explorer the default for debugging generative AI applications and agents, with adjustable row height, reorderable columns, and custom columns built from any trace tag or metadata field. It also speeds up metric charts for experiments with large metric histories. New MLflow Apps are created on 3.16 by default, and existing apps will be upgraded during an upcoming maintenance window with no action required. It is available in all Regions where MLflow Apps are offered. Source

3. Amazon ECS Adds GPU Sharing With NVIDIA Multi-Process Service

AWS added GPU sharing to Amazon ECS using NVIDIA Multi-Process Service (MPS), so multiple ECS tasks can run concurrently on one NVIDIA GPU. AWS targets workloads that use only part of a GPU, such as small-model inference, embedding generation, multi-model serving, and AV/ADAS simulations. Users set the GPU memory each container needs in the task definition and can optionally cap its share of GPU compute, and ECS places tasks on GPUs with available capacity, with no application code changes. It works on ECS Managed Instances and ECS on EC2; on Managed Instances, ECS also monitors the MPS control daemon and replaces the instance if the daemon becomes unhealthy. It is available in all AWS commercial and AWS GovCloud (US) Regions. Source