[sachalmalick]

Summary

Add an independent CDK entry point for GPU node groups that target externally owned On-Demand Capacity Reservations. Separate regional VPC/EKS foundations from fixed-size compute and application deployments, retain compute dependencies, and use MINIMAL updates to avoid requiring surge capacity.

Use current operator credentials with an S3-only asset setup, pinned EKS NVIDIA releases and a pinned device plugin. Add read-only AWS preflight checks and an operator runbook; deployment identifiers remain in ignored local configuration.

Closes #17263.

Validation

• TypeScript build and 50 CDK tests passed.
• Documentation lint, CODEOWNERS coverage and privacy checks passed.
• Live deployment completed: two GPU nodes in each of four regions, all managed node groups ACTIVE and all eight nodes Ready with one allocatable GPU each.
• Host nvidia-smi and Kubernetes GPU container smoke tests passed on all eight nodes across L4, L40S and A10G hardware.
• EC2 verified all eight instances explicitly occupy the intended reservations and have no public IPs.
• Retention-only foundation updates reviewed for unchanged resource properties; instance IDs remained unchanged after applying the updates.
• Stack deletion protection, dependency retention, fixed counts and fully occupied reservations verified.

No reservation was created, modified, extended or cancelled by this change. Actual model-serving performance is a separate validation step.

Superseded

Replaced by #17418, which combines this complete CDK stack into a single commit against the current global-router-v2. The new PR includes the already merged #17270 and #17271 changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.