AWS AI Updates: October 31, 2026

1. SageMaker AI Adds MIG-Based GPU Partitioning for Real-Time Endpoints

AWS added GPU partitioning to Amazon SageMaker AI real-time inference, using NVIDIA Multi-Instance GPU (MIG) to split one physical GPU into as many as seven isolated partitions, each with its own compute slices and dedicated memory so one model cannot read another’s memory or take its throughput. Partitioning is enabled with two optional API fields and no new operations or IAM actions: AcceleratorPartitionLayouts on the production variant of an endpoint that hosts inference components, and AcceleratorPartitionType on each inference component, using partition types such as mig-1g.10gb or mig-3g.40gb. Supported families are P4d (A100 40 GB), P5 (H100 80 GB), G7 (RTX PRO 4500 32 GB), and G7e (RTX PRO 6000 96 GB); a P5 H100 can be cut into seven 10 GB partitions, and a G7e RTX PRO 6000 into four 24 GB partitions. Clients keep invoking models by inference component name, but the partition layout is fixed when the endpoint is created (changing it requires a new endpoint), and partitioned endpoints do not publish GPU, CPU, or memory utilization metrics. AWS targets small and mid-size models such as fine-tuned variants, embedding models, rerankers, and classifiers, and the feature is available in us-east-1, us-east-2, us-west-2, ap-southeast-1, ap-northeast-1, eu-central-1, and eu-west-1. Source

2. Bedrock in GovCloud Gains GPT-5.6 Runtime APIs, Cross-Region Inference, Daybreak Red, and the ChatGPT Desktop App

AWS expanded OpenAI model support in Amazon Bedrock for AWS GovCloud (US). GPT-5.6 Terra and Luna are now served on the bedrock-runtime endpoint with the Responses, Converse, and Chat Completions APIs, and a new Cross-Region Inference option routes requests across both GovCloud (US) Regions for throughput and redundancy while keeping data inside US sovereign boundaries, with usage attributable per model in model invocation logs, CloudWatch metrics, Cost Explorer, and the Cost and Usage Report. Daybreak Red, which provides OpenAI’s GPT-5.6 Cyber model for authorized vulnerability research, exploit reproduction, and mitigation development, is now available in GovCloud (US-East) alongside US East (Ohio), where it launched in August; access requires enrollment in OpenAI’s Trusted Access for Cyber program and approval through an AWS account team, and AWS says inference data is not used for training or shared with OpenAI. A companion blog post documents running the ChatGPT desktop app (Work and Codex) against the Bedrock Mantle endpoint in us-gov-west-1 over the Responses API, with API-key or SDK credential-chain sign-in, a managed requirements.toml that allows only the Bedrock endpoint, denies OpenAI and ChatGPT domains, and disables browser use, computer use, plugins, and memories, billed per token with no seat licenses. Separately, Kiro’s web search and web fetch tools are now available in GovCloud (US), with citations in answers and organization-level switches in the Kiro console to disable them or strip them from individual custom agents. Source Source Source Source

3. AWS Batch Quota Management Lets Teams Lend, Borrow, and Preempt GPU Capacity for SageMaker Training Jobs

AWS detailed quota management, a new AWS Batch scheduling capability that divides GPU capacity into quota shares per team, project, or workload for Amazon SageMaker Training jobs. Each share has a capacity limit and one of three strategies (reserve idle capacity, lend it, or lend and borrow), shares under their limits are scheduled first, and when jobs arrive in a share that has lent out capacity AWS Batch reclaims it through cross-share preemption. Optional in-share preemption lets higher-priority jobs displace lower-priority jobs that are scheduled, starting, or running, and priorities can be raised after submission with the UpdateServiceJob API. Setup consists of a scheduling policy, a service environment with instance-based capacity limits, and a SageMaker Training job queue with attached shares; jobs are submitted to a named share with a priority through the aws_batch module of SageMaker Python SDK v3. AWS’s walkthrough configures three shares each capped at one ml.g5.xlarge instance to show lending, borrowing, and preemption, and notes the feature pairs with SageMaker AI training plans for reserved capacity. Source

4. SageMaker Unified Studio Adds Custom Designations for Granular Project Permissions

AWS added custom designations to Amazon SageMaker Unified Studio, named permission sets that control which actions each project member can perform. Members could previously only be assigned Owner or Contributor, which AWS describes as nearly identical; administrators can now assemble permissions from categories including Catalog, Subscriptions, Connections, Notebooks, and Data Agent, for example a Data Steward that approves subscriptions and publishes assets, or an Analyst limited to querying data and running notebooks. Designations are defined at the domain level and can be assigned to members of any project in that domain, and each domain supports the two managed designations plus up to 10 custom ones. The feature is available in all Regions where SageMaker Unified Studio is offered. Source