Suboptimal Use of Compute Savings Plans for Specialized Instances

When organizations rely solely on CSPs, these lower priority instances are typically the last to benefit from the plan, especially if other instance types consume most of the discounted hours.

As a result, p5 usage may fall through the cracks and be billed at full On-Demand rates despite an active CSP. This dynamic makes CSPs a potentially inefficient choice for workloads that heavily or predictably rely on these instance types. EC2 Instance Savings Plans provide better discount targeting for known usage patterns, and EC2 Instance Savings Plans for P5 and P5en are available in 3-year terms and, since June 2025, 1-year terms with up to 40% savings compared to On-Demand. AWS also reduced P5 and P5en On-Demand and Savings Plan prices in June 2025, so older CSP discount comparisons may be out of date. Additionally, Capacity Blocks provide reserved GPU capacity for fixed windows at reservation prices that AWS updates based on supply and demand, but they come with operational rigidity and inflexible scheduling constraints that limit their applicability.

AWS EC2 instances can be purchased via multiple pricing models: On-Demand, Compute Savings Plans (CSPs), EC2 Instance Savings Plans, Reserved Instances, and Capacity Blocks (for specific accelerated instance types).

Compute Savings Plans apply dollar-based discounts based on discount percentage priority, with savings first applied to usage with the highest discount percentage. If multiple instances have the same discount percentage, the plan applies to usage with the lowest rate. This allocation methodology can result in accelerated instances like p5.48xlarge and p5en.48xlarge receiving lower priority if its discount percentages are less competitive than other instance types in the environment.

For accelerated instances like `p5.48xlarge` and `p5en.48xlarge`, CSPs may offer relatively low discount rates, which makes them both less cost-effective and deprioritized in CSP allocation. This can lead to uncovered spend on these expensive instances unless the total CSP commitment is large enough to cover all usage.

• Review usage patterns of high-cost accelerated instances such as `p5.48xlarge` and `p5en.48xlarge`
• Determine if these instances are being run consistently at high utilization levels
• Assess how much of their usage is actually covered under existing Compute Savings Plans
• Verify current CSP discount rates for various instances in your specific regions, as rates vary by region, term, and payment option
• Evaluate whether other instance types with higher CSP discount rates are exhausting the plan allocation
• Compare costs against dedicated EC2 Instance Savings Plans (available in 1-year and 3-year terms for P5 and P5en) for predictable workloads
• Confirm whether an EC2 Instance Savings Plan would have better alignment with observed usage patterns
• Identify potential overcommitment risks if CSPs were scaled up solely to include p-type instance usage

• Consider using dedicated EC2 Instance Savings Plans instead of CSPs for predictable, high-utilization p5 workloads
• Model CSP allocation on actual discount percentages in order to determine whether p-type instances are likely to be left uncovered
• Compare total cost of ownership between CSPs and dedicated instance savings plans for your specific usage patterns
• Evaluate potential fit for Capacity Blocks only if workloads are short-term, highly predictable, and operationally isolated
• Ensure any long-term commitment aligns with actual usage patterns and instance mix
• Review and update savings plan strategy as AWS introduces new instance-specific savings plan options

• Understanding how Savings Plans apply to your usagedocs.aws.amazon.com
• Compute Savings Plansaws.amazon.com
• One Year EC2 Instance Savings Plans are now available for P5 and P5en instancesaws.amazon.com
• Pricing and usage model updates for Amazon EC2 instances accelerated by NVIDIA GPUsaws.amazon.com
• Amazon EC2 Capacity Blocks for ML Pricingaws.amazon.com