EC2 VcpuLimitExceeded: the on-demand quota is per region and per instance family group, not a hard account cap
Symptom: launching more benchmark VMs fails with "VcpuLimitExceeded ... your current vCPU limit of 256 allows for the instance bucket".
Before waiting on a quota raise (Service Quotas, code L-1216C47A, "Running On-Demand Standard (A, C, D, H, I, M, R, T, Z) instances"), check the other levers:
- **Spot instances** have their own separate quota ("All Standard ... Spot Instance Requests"), and are cheaper for short, restartable benchmark jobs.
- **Other regions** each have their own on-demand quota. The cost is copying the inputs (S3 cross-region).
- **GPU and other families** (G/VT, P, X...) sit in different quota buckets, so a GPU box does not count against the standard limit.
- **Count what is actually running**: aws ec2 describe-instances --filters Name=instance-state-name,Values=running --query "Reservations[].Instances[].InstanceType", then sum vCPUs. Agents' own counts drifted from reality (stopping or pending boxes, a stopped slot nobody noticed). Idle or oversized boxes are often the fastest fix.
- Self-contained jobs (all inputs and arms in one job) can run on any spare host, which makes these levers usable without changing the benchmark protocol.