AWS Lambda Adds GPU Instances for Event Functions on Lambda Managed Instances
As of November 18, Lambda Managed Instances can run event functions on G-series GPU instances: G4dn, G5, G6, G6f, Gr6, Gr6f and G6e. This brings GPU inference, such as embedding generation, image processing and small-model serving, to Lambda's event-driven model without running your own EC2 fleet.
The limits are sized for ML workloads:
- Container images up to 32 GB.
- 2,048–32,768 MB of memory and 3–48 GB of accelerator memory.
- Up to 48 concurrent requests per execution environment.
- Up to 15,000 vCPUs per capacity provider.
- A 90-minute timeout for asynchronous and event-source-mapping invocations.
You pay the EC2 instance price plus a 9.75% GPU management fee and $0.20 per million requests. Your account also needs a G-series quota increase first.
Read more — AWS News Blog
AWS Nitro Sandboxes: A MicroVM per EKS Pod via runtimeClassName
AWS Nitro Sandboxes run an opted-in Kubernetes pod, ECS task or EC2 workload in its own lightweight VM with a separate guest kernel. The Nitro hypervisor carves that VM out of a running EC2 instance. On EKS, you opt in with one pod field: set runtimeClassName to isolated-sandbox or confidential-sandbox. The feature is available in all commercial regions at no extra charge. Under the hood, a containerd shim starts a Firecracker microVM, and an agent inside the VM talks to the host over virtio and vsock.
In isolated mode, the sandbox trusts its parent instance. Confidential mode moves enforcement inside the VM. It applies a Cedar policy generated from the pod spec, uses Nitro Secure Module measurements, and gates secrets with KMS key policies. AWS's docs warn that confidentiality is incomplete if you don't verify the measurements. At launch there's no GPU or device passthrough. Total sandbox vCPUs are capped at 5x the node's vCPUs, and SMT roughly halves the usable vCPUs on x86 instances. The main use is per-pod kernel isolation for untrusted code, such as AI coding agents, without running a separate node pool.
Read more — InfoQ
AWS Weekly Roundup: Cognito Token Exchange, ElastiCache Serverless Search, OpenSearch Cross-Region DR
AWS's November 16 weekly roundup listed several more developer-facing launches beyond Nitro Sandboxes:
- Amazon Cognito now supports OAuth 2.0 token exchange, so a service can swap a user's token for one scoped to a downstream service.
- ElastiCache Serverless adds search and aggregations on Valkey 9.
- Amazon OpenSearch Service adds managed cross-Region disaster recovery for OpenSearch 3.7 and later, with a recovery time under 20 minutes.
- CloudWatch Logs supports account-level retention settings.
- The next-generation AgentCore runtime is now the default.
Account-level log retention is the quickest of these to adopt. Setting it once stops new log groups from silently keeping data forever. Token exchange removes a common reason to build a custom token broker in front of Cognito.
Read more — AWS News Blog
Gemini 4 Neon GA on Vertex AI and Deferred-Tier Pricing for Agents
Google made Gemini 4 Neon (gemini-4-neon) generally available on November 18. It has a unified token limit of 1,048,576 and implicit caching for prompt prefixes of 4,096 tokens or more. One migration catch: passing thinking_budget, or thinking_level="MINIMAL", returns HTTP 400. Code written for Gemini 3.x that sets those parameters needs updating. On November 19 the model also became GA in BigQuery's AI functions.
On November 16, a Deferred tier became generally available on the Interactions API, ADK (google-adk v2.10.0 and later) and Deep Research. Deferred requests cost 50% less than Standard pay-as-you-go, target completion within 24 hours, and keep state for 7 days. This suits batch-style agent work, such as overnight research or bulk evaluation, where latency doesn't matter.
Read more — Google Cloud Release Notes
Google Cloud Patches Java4Shell Across Managed Java Runtimes
On November 20, Google rolled out fixes for the OpenJDK Java4Shell vulnerability (CVE-2026-114938, bulletin GCP-2026-082). Patched runtimes 17.0.21.1+2, 21.0.13.1+2 and 25.0.5.1+2 cover App Engine standard and flexible, Cloud Run, Cloud Run functions and Buildpacks. Services with automatic updates are covered. Otherwise, rebuild and redeploy, then remove any temporary JDK_JAVA_OPTIONS mitigation flags.
Dataflow patched Apache Beam 2.78.0 worker images. Its bulletin warns that the patched JDK can hit out-of-memory errors under ParallelGC and suggests --experiments=java_gc=UseG1GC. Managed Service for Apache Airflow published fixed versions: composer-2.18.2-airflow-2.11.1 for Gen 2, and 3.3.1-build.10, 3.2.2-build.14 and 2.11.1-build.27 for Gen 3. Until you upgrade, Google's temporary mitigation for Airflow is -XX:+IgnoreUnrecognizedVMOptions -XX:UseAVX=2 in JAVA_TOOL_OPTIONS.
Read more — Google Cloud
Links & Sources
- AWS Lambda now supports GPU instances for Event Functions on Lambda Managed Instances
- AWS Nitro Sandboxes Put Each EKS Pod in its Own MicroVM with One RuntimeClass Field
- AWS Weekly Roundup: AWS Nitro Sandboxes, AWS Emulator, and more (November 16, 2026)
- Google Cloud release notes
- Dataflow security bulletins (GCP-2026-082)