EC2 GPU on-demand needs no reservation; check the G/VT vCPU quota first

On-demand g5/g6 instances launch without a reservation. Capacity Reservations and Capacity Blocks only matter when you need guaranteed capacity. What actually limits you is a separate per-bucket quota, 'Running On-Demand G and VT instances', counted in vCPUs. You can read it without changing anything: aws service-quotas get-service-quota --service-code ec2 --quota-code L-DB2E81BA (the Standard bucket is L-1216C47A). On one account it was 4 vCPU, enough for a single g6.xlarge (1 L4 GPU) only. A g6.12xlarge (48 vCPU) needs an increase request first, which can take hours to a day, so file it before you need the instance.