What Happened
Two AWS updates were announced on 2026-08-25 relevant to production AI and regulated workloads:
- Capacity Reservation Resource Groups now support Capacity Blocks for ML and interruptible Capacity Reservations. You can add any Capacity Reservation type (not just On‑Demand Capacity Reservations) to a reservation resource group, target that group in launch requests, and let EC2 Fleet and EC2 Auto Scaling groups prioritize across reservation types with automatic fallback to On‑Demand when reservations are exhausted [1]. No additional charge to use the feature; availability excludes AWS GovCloud (US) and China Regions [1].
- AWS Lambda MicroVMs now support AWS PrivateLink. PrivateLink VPC interface endpoints allow private connectivity from your VPC to MicroVM control APIs (create images, launch MicroVMs) and to each MicroVM’s HTTP endpoint so traffic does not traverse the public internet. Endpoints can be created via Console, CLI, CloudFormation or SDKs and are available in Regions that offer Lambda MicroVMs [2].
Why It Matters to Businesses
These changes address two operational needs for AI/ML and regulated applications:
- Predictable capacity for ML at scale. Capacity Blocks for ML plus interruptible reservations in reservation groups let teams combine reserved guaranteed capacity with lower‑cost, interruptible capacity and maintain a single target for launches. That reduces failed launches during scale events and simplifies scheduling for training and batch inference jobs [1].
- Network isolation for regulated workloads. PrivateLink support for Lambda MicroVMs lets organizations avoid public internet exposure for control plane and runtime HTTP traffic — a requirement for many financial, healthcare and government workloads while retaining serverless microVM performance characteristics [2].
Kimbodo Engineering Perspective
Both features reduce operational friction but introduce trade‑offs that teams must manage.
- Capacity mix management. Grouping On‑Demand, interruptible and ML Capacity Blocks simplifies launches, but you must design prioritization and fallback policies (EC2 Fleet/Auto Scaling) to avoid cost and availability surprises. Interruptible reservations reduce costs but can be preempted; ensure workload tolerance or checkpointing.
- Complexity vs predictability. A single reservation group improves predictability for orchestration systems (job schedulers, cluster autoscalers), but it can mask fragmentation of real capacity across AZs or instance types. You still need placement awareness and quota checks in provisioning logic.
- PrivateLink and security posture. PrivateLink improves network isolation but does not remove the need for strong authentication, endpoint policies, careful security group configuration, and logging. PrivateLink introduces per-endpoint costs and additional network resources to manage.
- Operational observability. Both features require updates to provisioning tooling, CI/CD templates, monitoring dashboards (capacity utilization, reservation exhaust events, PrivateLink endpoint health) and runbooks for interruption and failover scenarios.
How We Would Implement It
Step A — ML capacity with Capacity Reservation Resource Groups
- Create a Capacity Reservation Resource Group and add the desired Capacity Reservations: Capacity Blocks for ML, interruptible reservations, and any On‑Demand Reservations you maintain [1].
- Use EC2 Fleet or EC2 Auto Scaling with explicit capacity targeting and prioritization settings so the orchestration layer prefers guaranteed ML blocks first, then interruptible capacity, then On‑Demand as fallback [1].
- Integrate with schedulers (batch systems, Kubernetes node provisioning, or Karpenter/EKS nodegroups) so pod/node placement requests target the reservation group via launch templates or instance profile mappings.
- Deploy CloudFormation/Terraform modules that encapsulate reservation group creation, IAM roles, and launch templates. Add automated checks to verify availability across AZs and to reconcile reservation expirations.
- Instrument monitoring: CloudWatch metrics for reservation utilization, CloudTrail for launch requests, and alerts for reservation exhaustion and preemption events. Implement graceful interruption handlers and checkpointing for training jobs.
Step B — Lambda MicroVMs with PrivateLink
- Create an Interface VPC Endpoint (PrivateLink) for the Lambda MicroVMs service in each VPC that must access MicroVM control APIs or microVM HTTP endpoints. Use CloudFormation/CLI/SDK to provision endpoints and set endpoint policies to restrict callers [2].
- Configure security groups to allow only required subnets and services (e.g., inference clients, model registries) to access the endpoint. Use DNS integration so service names resolve to the endpoint IPs inside the VPC.
- Ensure MicroVM image management and lifecycle calls (create image, launch) are routed over PrivateLink; likewise, serve synthesis/inference HTTP endpoints through the endpoint to avoid public exposure [2].
- Apply API authentication and mTLS for HTTP endpoints running on MicroVMs, and add WAF or API Gateway in front if additional request filtering or rate limiting is needed.
- Monitor VPC Flow Logs, CloudWatch metrics, and CloudTrail for PrivateLink calls and microVM HTTP access. Automate endpoint rotation and least‑privilege IAM for the control plane calls.
Risks, Costs and Security
- Costs: There’s no extra fee to use Capacity Reservation Resource Groups, but you still pay for the underlying reservation types (including charges for reserved or interrupted capacity). PrivateLink incurs per‑endpoint and data processing charges — include those in TCO for high‑throughput inference [1][2].
- Availability & region constraints: Capacity Blocks for ML and interruptible ODCRs are not available in AWS GovCloud (US) and China Regions for this feature; PrivateLink availability mirrors Lambda MicroVMs regional presence — validate region support before design [1][2].
- Preemption and fragmentation: Interruptible reservations can be terminated; design workloads to tolerate interruptions and avoid capacity fragmentation across AZs and instance types that can prevent successful launches.
- Security posture: PrivateLink reduces internet exposure but does not replace auth, authorization, TLS/mTLS, and endpoint policies. Misconfigured VPC endpoints or overly permissive endpoint policies can create internal attack paths. Require IAM least privilege for control APIs and deploy network ACLs, security groups, and logging.
- Operational overhead: New building blocks require updates to provisioning code, runbooks, and monitoring. Automated testing of fallback scenarios (reservation exhaustion, endpoint failure) is essential to avoid silent outages.
References: AWS announcements and docs for Capacity Reservation Resource Groups and Lambda MicroVMs PrivateLink [1][2].
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Application Development practice, or Estimate My AI Application.
