Reserved Instance

Reserved Instances and Savings Plans

Cost Optimization Options
| Option | Commitment | Savings | Flexibility |
|--------|------------|---------|-------------|
| On-Demand | None | 0% | Full |
| Spot | None | 60-90% | Low (interruptions) |
| Reserved | 1-3 years | 30-72% | Low |
| Savings Plans | 1-3 years | 30-72% | Medium |

Reserved Instances
Commit to specific instance type in specific region:

p3.2xlarge in us-east-1
- On-demand: $3.06/hr = $26,825/year
- 1-year RI: $19,929/year (26% savings)
- 3-year RI: $12,964/year (52% savings)

Savings Plans
More flexible commitment to compute spend:

Compute Savings Plans
Works across:
- All instance types
- All regions
- EC2, Fargate, Lambda

EC2 Instance Savings Plans
Works across:
- All sizes within instance family
- All AZs in region

# Example commitment
# Commit to $10/hr spend
# Covers any mix of instances up to that amount

ML Workload Strategy
| Workload | Strategy |
|----------|----------|
| Always-on inference | Reserved/Savings Plan |
| Variable inference | On-demand + Spot |
| Training | Spot with checkpoints |
| Development | Spot |

Calculating Requirements

# Estimate steady-state compute
baseline_gpus = 8  # Always running
peak_gpus = 24     # During training

# Cover baseline with Savings Plan
# Cover peak with Spot + On-demand

# Baseline cost with g4dn.xlarge
baseline_hourly = 8 * 0.526  # $4.21/hr
baseline_yearly = baseline_hourly * 24 * 365  # $36,900

# With 3-year Savings Plan (52% savings)
savings_plan_cost = baseline_yearly * 0.48  # $17,712/year

Best Practices
- Analyze usage patterns before committing
- Start with 1-year commitment
- Use Savings Plans for flexibility
- Combine with Spot for variable workloads
- Review and adjust annually
- Use AWS Cost Explorer / GCP Recommender