| name | gcp-finops |
| description | GCP cost engineering and financial operations. Analyze billing, optimize resources, manage budgets, reduce spending. Use when discussing GCP costs, budget optimization, pricing decisions, cost analysis, or spend management. |
| allowed-tools | Read, Grep, Glob, WebFetch |
| kb-sources | ["wiki/software-engineering/gcp-finops"] |
| updated | "2026-04-19T00:00:00.000Z" |
GCP FinOps
Cost engineering and financial operations patterns for Google Cloud Platform.
Note: Pricing estimates are approximate and based on us-central1 region. Use the GCP Pricing Calculator for current rates.
Cost Optimization Framework
The FinOps Cycle
┌─────────────┐
│ INFORM │ ← Visibility into spending
│ (analyze) │
└──────┬──────┘
│
┌──────▼──────┐
│ OPTIMIZE │ ← Right-size, eliminate waste
│ (reduce) │
└──────┬──────┘
│
┌──────▼──────┐
│ OPERATE │ ← Maintain efficiency
│ (control) │
└─────────────┘
Quick Cost Reference
Compute (Cloud Run)
| Config | Cost/hour | Monthly (24/7) | Use Case |
|---|
| 0.5 vCPU, 256MB | $0.0144 | $10.50 | Minimal API |
| 1 vCPU, 512MB | $0.0324 | $23.60 | Standard API |
| 2 vCPU, 1GB | $0.0648 | $47.20 | Heavy processing |
Key insight: Scale-to-zero saves 80-95% for variable workloads.
GPU Instances (Cloud Run with L4)
| Config | Cost/GPU-sec | GPU Cost/hour | GPU Monthly (24/7) | Use Case |
|---|
| 4 vCPU, 16GB + L4 (min) | $0.0001867 | ~$0.67 | ~$490 | Light GPU inference |
| 8 vCPU, 32GB + L4 (rec) | $0.0001867 | ~$0.67 | ~$490 | Production GPU inference |
Note: Table shows GPU cost only. vCPU/memory charges are additional (see Compute table above). Total cost for 8 vCPU/32GB + L4 is ~$0.67 GPU + ~$0.19 vCPU/mem = ~$0.86/hr.
Key constraints:
- Minimum instance spec: 4 vCPU / 16 GiB RAM (hard GCP requirement)
- Recommended: 8 vCPU / 32 GiB RAM for production workloads
- Cold start penalty: 20-40s (GPU initialization)
- Min-instances break-even: ~2,700 jobs/month justifies keeping 1 instance warm
- GPU billing is per-second ($0.0001867/GPU-sec), vCPU/memory billed separately
Caution: Agent-generated GPU cost estimates are frequently wrong (e.g., confusing $/hr with $/min — a 60x error). Cross-check against Cloud Run pricing before committing to a budget — unverified estimates have shipped with 60x errors that made the FinOps recommendation actively harmful.
Database (Cloud SQL)
| Tier | Monthly | RAM | Best For |
|---|
| db-f1-micro | $16 | 0.6GB | Dev only |
| db-g1-small | $28 | 1.7GB | Staging |
| db-standard-1 | $68 | 3.75GB | Small prod |
| db-standard-2 | $136 | 7.5GB | Medium prod |
Key insight: +35% for HA (REGIONAL).
Cache (Memorystore Redis)
Pricing model: $0.049/GB/hour (BASIC tier, us-east4)
| Size | BASIC | STANDARD_HA |
|---|
| 1GB | ~$35/mo | ~$70/mo |
| 2GB | ~$72/mo | ~$144/mo |
| 5GB | ~$180/mo | ~$360/mo |
Key insights:
- 2x cost for HA (STANDARD_HA), but includes automatic failover
- Pricing is hourly: $0.049/GB/hour x 730 hours/month = ~$35.77/GB/month
- Custom sizes available: 1GB minimum, 300GB maximum (BASIC tier)
- Region variations: us-east4 typically cheapest, check calculator for others
Cost calculation formula:
Monthly cost = Memory_GB x $0.049/hour x 730 hours/month x HA_multiplier
Where:
HA_multiplier = 1.0 (BASIC) or 2.0 (STANDARD_HA)
Example: 2GB STANDARD_HA
= 2 x $0.049 x 730 x 2.0
= $143.08/month
Right-Sizing Decision Matrix
| Current State | Recommendation | Savings |
|---|
| CPU <10% utilization | Downsize tier | 30-50% |
| Memory <30% used | Reduce RAM allocation | 20-40% |
| Scale-to-zero disabled | Enable for dev/staging | 80-95% |
| HA in staging | Disable HA | 26-35% |
| SSD in staging | Switch to HDD | 47% |
Budget Alert Thresholds
| Threshold | Action | Audience |
|---|
| 50% | Informational | Team |
| 75% | Review spending | Tech Lead |
| 90% | Reduce non-essential | Team Lead |
| 100% | Emergency review | Management |
See reference.md for detailed pricing research methods and optimization patterns.
