| Instance | AWS | CoreWeave | Savings |
|---|---|---|---|
| 1x A100 80GB | ~$3.60/hr (p4d) | ~$2.21/hr | ~39% |
| 8x A100 80GB | ~$32/hr (p4d.24xl) | ~$17.70/hr | ~45% |
| 1x H100 80GB | ~$6.50/hr (p5) | ~$4.76/hr | ~27% |
# If running on bare EC2/GCE, containerize first
docker build -t inference-server:v1 .
docker push ghcr.io/myorg/inference-server:v1
Key changes from AWS EKS / GKE:
gpu.nvidia.com/class instead of nvidia.com/gpu.product
shared-ssd-ord1)Run both old and new infrastructure simultaneously, gradually shift traffic.
Decommission old GPU instances after validation period.
| Issue | Solution |
|---|---|
| Different CUDA drivers | Match container CUDA to CoreWeave node drivers |
| Storage migration | Use rclone or rsync to move data to CoreWeave PVC |
| DNS changes | Update ingress/load balancer DNS |
| IAM differences | CoreWeave uses kubeconfig, not IAM roles |
This completes the CoreWeave skill pack. Start with coreweave-install-auth for new deployments.