Lilly Tech Systems
Committed Use Discounts Intermediate
Committed Use Discounts (CUDs) provide 37-55% savings on GCP compute resources in exchange for a 1 or 3-year commitment. For AI workloads with predictable, steady-state GPU usage, CUDs offer the best guaranteed pricing.
CUD Types for AI
| CUD Type | 1-Year Discount | 3-Year Discount | Best For |
|---|---|---|---|
| Resource-based | 37% | 55% | Specific vCPU and memory amounts |
| Spend-based | 25% | 52% | Flexible across GPU types and regions |
Purchasing a CUD
Bash
# Purchase a 1-year resource-based CUD for A100 GPUs
gcloud compute commitments create ai-gpu-commitment \
--region=us-central1 \
--plan=12-month \
--resources=vcpu=96,memory=624GB \
--resources-accelerator=type=nvidia-tesla-a100,count=8
When to Use CUDs
- Production inference: Endpoints running 24/7 with stable GPU requirements
- Continuous training: Teams that run training jobs consistently throughout the month
- Baseline compute: Commit to your minimum steady-state usage, use Spot for burst
CUD Strategy for AI Teams
Recommended Approach: Analyze 3 months of GPU usage to identify your baseline. Commit to 60-70% of your average usage via CUDs. Use Spot VMs for the remaining burst capacity. This hybrid approach maximizes savings while maintaining flexibility.