Cloud Commitment Strategies: Reserved Instances, Savings Plans, and CUDs

Cloud commitment strategies reduce variable cloud costs by locking in capacity or compute for periods of time, typically one or three years. They matter when workloads are predictable—such as steady-state web services, batch ETL pipelines, or long-running databases—where upfront investment in reserved resources yields better unit economics than on-demand pricing. This page details the core mechanisms of Reserved Instances, Savings Plans, and Compute Unit Days (CUDs), and guides engineers through the operational decisions that make these commitments effective.

Committing to Predictable Workloads

When a workload runs consistently over multiple months—e.g., a PostgreSQL instance serving a production API, or a daily Spark job processing 10TB of logs—committing to compute becomes more valuable than paying per-second. The key insight is that commitment reduces cost volatility and improves pricing predictability. The following decisions determine how well the commitment performs:

Choosing the right commitment duration (1-year vs. 3-year)
Matching instance types and regions to actual usage patterns
Scheduling commitments in advance to avoid underutilization
Managing the lifecycle of commitments: renewal, modification, and termination

These strategies are not one-size-fits-all. The same infrastructure may require a mix of Reserved Instances for core services and Savings Plans for burstable workloads, with CUDs as the unit of measurement for cross-platform, cross-service commitments.

Reserved Instances: Fixed Capacity, Fixed Cost

Reserved Instances (RIs) provide a fixed compute instance—CPU, memory, and network—for a defined period, typically 1 or 3 years. Each RI is tied to a specific instance type and region, and incurs a fixed upfront cost plus a recurring hourly rate.

Key Configuration and Lifecycle

A reservation is defined by:

instance_type (e.g., m5.xlarge)
region (e.g., us-east-1)
duration (1 year or 3 years)
payment_option (all upfront, partial upfront, no upfront)

To create a reservation using a standard cloud API, send an HTTP POST request to /reservations:

POST /reservations
{
  "instance_type": "m5.xlarge",
  "region": "us-east-1",
  "duration": "3",
  "payment_option": "all_upfront",
  "count": 5
}

The response returns a reservation_id and start_time in ISO 8601 format:

{
  "reservation_id": "ri-42a7b1c9",
  "instance_type": "m5.xlarge",
  "region": "us-east-1",
  "duration": "3",
  "payment_option": "all_upfront",
  "count": 5,
  "start_time": "2025-04-01T00:00:00Z",
  "status": "active"
}

The reservation becomes active on the start_time, and the instance is allocated from that moment. The cost is charged per hour, even if the instance is idle.

Common Pitfalls

Instance type mismatch: A reservation for m5.xlarge does not apply to m5.2xlarge unless explicitly configured. Using m5.2xlarge with an m5.xlarge reservation results in underutilization and wasted capacity.
Region drift: A reservation in us-east-1 does not automatically cover us-west-2. Moving instances across regions requires manual re-allocation or cross-region reservations.
Overcommitment: Creating 10 RIs for m5.xlarge but only using 6 instances leads to 40% underutilization. The unused RIs continue charging hourly, even when idle.
Failure to renew: RIs expire after the commitment period. If not renewed, the instance becomes on-demand again, and the savings are lost.

Managing RIs

Renewal is triggered by checking the status of reservations. To renew a reservation, send a PATCH request:

PATCH /reservations/ri-42a7b1c9
{
  "duration": "3",
  "payment_option": "partial_upfront",
  "count": 8
}

If the instance type or region changes, the system must either:

Apply a new reservation and manually migrate instances, or
Use a reservation exchange: swap an existing RI for another instance type or region.

A reservation exchange is an HTTP POST to /reservations/exchange:

POST /reservations/exchange
{
  "original_reservation_id": "ri-42a7b1c9",
  "new_instance_type": "c5.2xlarge",
  "new_region": "us-west-2",
  "new_duration": "1"
}

The exchange process is asynchronous. The response includes a task_id and status tracking:

{
  "task_id": "exchange-1a2b3c",
  "status": "processing",
  "progress": 0.42,
  "estimated_completion": "2025-04-15T14:30:00Z"
}

The system polls this task until completion, at which point the old reservation is retired and the new one becomes active.

Savings Plans: Flexible Compute Commitment Across Instance Families

Savings Plans offer a commitment to a fixed amount of compute usage, measured in Compute Unit Days (CUDs), across a defined region and instance family. Unlike RIs, Savings Plans apply to a broader set of instance types, enabling flexibility in workload placement.

Core Mechanics

A Savings Plan is defined by:

compute_units (e.g., 1 CUD = 1 vCPU × 1 hour)
region (e.g., eu-west-1)
duration (1 or 3 years)
payment_option (all upfront, partial, no upfront)

The commitment is expressed in total CUDs over the duration. For example, a 3-year Savings Plan with 1,000 CUDs per month provides 36,000 CUDs over three years.

To activate a Savings Plan:

POST /savings-plans
{
  "compute_units": 1000,
  "region": "us-east-1",
  "duration": "3",
  "payment_option": "all_upfront"
}

The response returns a plan_id and commitment_start:

{
  "plan_id": "sp-8d9e0f1a",
  "compute_units": 1000,
  "region": "us-east-1",
  "duration": "3",
  "payment_option": "all_upfront",
  "commitment_start": "2025-04-01T00:00:00Z",
  "status": "active"
}

How Savings Plans Apply Costs

When an instance runs, it consumes CUDs based on its vCPU count and runtime. For example:

A m5.xlarge instance (4 vCPUs) running for 24 hours uses:

4 vCPUs × 24 hours = 96 CUDs

A c5.4xlarge instance (16 vCPUs) running for 6 hours uses:

16 vCPUs × 6 hours = 96 CUDs

Both consume 96 CUDs. The Savings Plan applies cost reduction to any instance that contributes CUDs to the plan.

Critical Failure Modes

Underutilized CUDs: A Savings Plan with 1,000 CUDs/month but only 800 used each month leaves 200 CUDs unused, which are not charged but not applied to the commitment. The system does not auto-roll over unused CUDs.
Instance type mismatches: A Savings Plan for m5 family does not apply to c5 instances unless explicitly configured. The system tracks CUDs per instance family, so a c5.2xlarge contributes to the c5 family plan but not to the m5 plan.
Plan overlap and priority: When multiple Savings Plans apply to the same instance, the system must resolve the order of application. The rules are:
Plans are applied in order of commitment_start (earliest first).
Later plans take precedence over earlier ones.
CUDs are consumed from the latest active plan first.

To query applied plans for an instance:

GET /instances/i-1a2b3c/usage?from=2025-04-01T00:00:00Z&to=2025-04-02T00:00:00Z

Response:

{
  "instance_id": "i-1a2b3c",
  "region": "us-east-1",
  "start": "2025-04-01T00:00:00Z",
  "end": "2025-04-02T00:00:00Z",
  "total_cuds": 144,
  "applied_plans": [
    {
      "plan_id": "sp-8d9e0f1a",
      "family": "c5",
      "cuds_applied": 96,
      "rate": "0.08/hour"
    },
    {
      "plan_id": "sp-7c8d9e0f",
      "family": "m5",
      "cuds_applied": 48,
      "rate": "0.10/hour"
    }
  ]
}

Managing Savings Plans

Plans can be modified or extended:

PATCH /savings-plans/sp-8d9e0f1a
{
  "compute_units": 1500,
  "duration": "3",
  "payment_option": "partial_upfront"
}

The system recalculates the CUD consumption and applies the new commitment from the commitment_start time.

To create a new plan with overlapping commitments, the system must track all active plans and apply CUDs in order. A common failure is plan crosstalk: a c5 instance running during a period when both a c5 and m5 plan are active, but the system applies the m5 plan’s rate to the c5 instance, even though c5 is the preferred family.

Compute Unit Days (CUDs): The Currency of Commitment

CUDs are the standardized unit for measuring compute commitment. One CUD is defined as one vCPU running for one hour. It is the fundamental building block across both RIs and Savings Plans.

CUD Calculation and Usage

For any compute instance, the number of CUDs consumed is:

CUDs = (number of vCPUs) × (runtime in hours)

This applies uniformly across instance types and platforms.

To calculate total CUD usage over a period:

List all running instances and their durations.
For each instance, compute CUDs.
Sum across all instances.

Example: Over 7 days, a cluster runs:

3 m5.xlarge instances (4 vCPUs) for 24 hours each → 3 × 4 × 24 = 288 CUDs
2 c5.2xlarge instances (8 vCPUs) for 168 hours total → 2 × 8 × 168 = 2,688 CUDs

Total: 2,976 CUDs over 7 days.

Tracking and Reporting CUDs

CUDs are reported via a /cuds endpoint:

GET /cuds?from=2025-04-01T00:00:00Z&to=2025-04-08T00:00:00Z&group_by=region

Response:

[
  {
    "region": "us-east-1",
    "total_cuds": 2976,
    "instance_count": 5,
    "average_cuds_per_day": 425.14
  },
  {
    "region": "eu-west-1",
    "total_cuds": 1542,
    "instance_count": 3,
    "average_cuds_per_day": 220.29
  }
]

CUDs are also used to validate commitments. A Savings Plan with 10,000 CUDs per month should see at least 10,000 CUDs in the monthly report. If only 8,500 are reported, the system flags a CUD shortfall.

Silent Failures in CUD Accounting

Instance startup delay: An instance started at 2025-04-01T01:00:00Z and stopped at 2025-04-01T23:59:59Z contributes 22.999 hours, but CUD calculation assumes full hours. The system may round to 23 hours, leading to a 1-hour CUD shortfall.
Overlapping instances: Two instances start at the same time but are not synchronized. Their CUDs are counted separately, but the system may not align time intervals, leading to double-counting or undercounting.
CUD roll-over: Unused CUDs from one month do not automatically carry over to the next unless explicitly configured. A plan with 1,000 CUDs/month and only 900 used in March has 100 CUDs left, but those are lost if not enabled.

Decision Flow: Choosing the Right Commitment Type

When setting up a new workload, engineers follow this flow:

Analyze usage patterns over the past 6 months using the /usage endpoint.
Determine stability:

If instance type and size are stable, use Reserved Instances.
If instance types vary but compute is consistent, use Savings Plans.

Calculate required CUDs to determine plan size.
Choose duration:

1-year commitment for medium-term stability.
3-year commitment for long-term, predictable workloads.

Schedule commitments in advance, using commitment_start to align with forecasted traffic peaks.
Monitor CUD usage monthly, and adjust plans or renew RIs as needed.

Failure occurs when commitments are created too late, under-sized, or not aligned with actual usage. The most common silent failure is CUD accounting mismatch, where the commitment appears fulfilled, but actual CUDs fall short due to unaligned time windows or startup delays.

Commitment strategies are not a one-time setup. They require ongoing monitoring, reporting, and iteration. The key to success is treating CUDs as the shared currency across all compute commitments, and managing them with the same rigor as financial budgets.

This page was rewritten on 17 October 2026. It replaced a templated version that was not specific to its own title. How these pages are written and checked. If anything here is wrong, tell us at [email protected] and we will correct it.