Friday, November 20, 2026 Edition No. 244
Search
Daily News Fully automated · 5 min read

AWS AI Updates: November 19, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Amazon OpenSearch Memory Reaches General Availability as Serverless Persistent Memory for AI Agents

Amazon OpenSearch Service now offers OpenSearch Memory as generally available, a serverless capability that gives AI agents persistent memory across sessions without replaying full conversation history to the model. It stores agent interactions verbatim as events that become searchable within seconds, then derives structured memories in the background, handling extraction, embedding, indexing and consolidation, where new information can add to, update or skip existing memories. Retrieval supports keyword, semantic and hybrid search, access is controlled per memory store through AWS IAM with actor and session scoping, and expiration periods can be set for events and memories. It is available in US East (N. Virginia and Ohio), US West (Oregon), Europe (Frankfurt) and Asia Pacific (Tokyo). Source

2. AWS Glue ETL Gets an Apache Spark Optimization Agent for Cost and Performance Recommendations

AWS Glue introduced an Apache Spark optimization agent that analyzes completed job runs and returns cost or performance recommendations, typically within minutes. A user picks a cost or performance goal, and a deterministic recommendation engine checks job configuration, logs and Spark telemetry for problems such as insufficient parallelism, skewed join keys, memory spill and idle executors. It also checks platform-level practices such as moving to a newer Glue version or using Graviton processors. Each recommendation includes the configuration change, the supporting telemetry and the reasoning, and is tagged by layer: infrastructure, Spark configuration, code or data layout. It is available in the Glue console or through the Amazon SageMaker Unified Studio MCP Server, requires Glue 4.0 or later and has no additional charge, in all Glue Regions except AWS GovCloud (US) and China. Source

3. SageMaker AI Adds Cross-Region Inference Routing That Shifts Traffic When Local Capacity Runs Short

Amazon SageMaker AI now supports cross-Region inference routing, which sends requests to endpoints in other Regions and AWS accounts when one Region runs short of capacity. A user creates an inference router and registers existing real-time endpoints, asynchronous endpoints and inference components, with no changes to models or containers. The router prefers endpoints in its own Region, shifts new requests to healthy endpoints elsewhere as errors rise, and returns traffic gradually once the local endpoint recovers. Endpoint owners control which routers in other accounts can use an endpoint through an endpoint policy, and can set a cross-Region traffic cap to reserve capacity for local callers. Real-time inference is supported in 32 commercial Regions and asynchronous inference in 25, with no additional charge beyond standard SageMaker AI and CloudWatch pricing. Source

4. Amazon ElastiCache Adds a Distributed KV Cache Tier for Self-Hosted LLM Inference

Amazon ElastiCache now supports distributed key-value caching for self-hosted LLM inference on ElastiCache for Valkey. It adds a shared cache tier beyond each GPU’s local cache, so any server can reuse stored KV state across sessions, routing changes and server failures, with retention measured in days. GPUs fetch cached state over RDMA using Elastic Fabric Adapter. AWS cites up to 40% lower inference cost and up to 80% lower time to first token. It requires cache.i8ge.48xlarge nodes in the same Availability Zone as the GPU fleet and the LMCache connector for vLLM, SGLang or TensorRT-LLM, and is available in 13 Regions. AWS also points to KVTrace, a free open-source tool that estimates the savings on a team’s own traffic. Source

5. SageMaker HyperPod EKS Clusters Support Higher Pod Density per Node

Amazon SageMaker HyperPod now supports higher pod density on EKS-orchestrated clusters, so each node can run more pods and make better use of GPU, CPU and memory. The Amazon VPC CNI adds network capacity automatically as pods are scheduled, and Karpenter autoscaling honors the higher pod limits on newly provisioned nodes. AWS lists use cases including more model replicas per GPU instance on smaller types such as ml.g5 and ml.g6, one pod per Multi-Instance GPU slice on instances such as ml.p5en.48xlarge, and room for system DaemonSets. Enabling it requires updating the cluster’s IAM permissions, cluster software and the VPC CNI add-on, and it is available in all Regions where HyperPod is offered. Source

6. AWS Lambda Managed Instances Add GPU Support With vLLM, SGLang and PyTorch Base Images

AWS Lambda now supports GPU-accelerated Event Functions on Lambda Managed Instances, which run on EC2 G-series instances while Lambda manages GPU drivers, frameworks and scaling. Preconfigured GPU base images include vLLM, SGLang and PyTorch, container images are supported with up to 32 GB of NVMe ephemeral storage, and Seekable OCI lazy loading and direct streaming from Amazon S3 reduce startup time. Scaling responds to GPU utilization, CPU utilization and concurrency. AWS lists AI inference first among target workloads, alongside media processing, 3D rendering and scientific computing. Pricing is the EC2 On-Demand price for the instance plus a 9.75% Lambda management fee and $0.20 per million requests. Source