What are the main cloud pricing models such as on-demand, reserved, spot, and savings plans, and when is each most appropriate?

Question

Accepted Answer

Cloud providers offer multiple pricing models that trade commitment flexibility for cost savings. Understanding these models is essential for building cost-effective architectures, and most production environments use a combination of several models.

On-demand pricing charges you by the hour or second with no commitment. You pay the full listed price but can start and stop resources at any time with no penalty. On-demand is appropriate for unpredictable workloads where you cannot forecast usage, short-term projects or experiments, initial development and testing before understanding resource needs, and handling temporary traffic spikes above your reserved baseline. The flexibility comes at the highest per-unit cost.

Reserved Instances and Savings Plans offer significant discounts, typically 30 to 72 percent compared to on-demand, in exchange for a one-year or three-year commitment. Reserved Instances commit to a specific instance type in a specific region. Savings Plans are more flexible, committing to a dollar amount of compute usage per hour regardless of instance type or region. These are ideal for baseline production workloads with predictable, steady usage. If you know you need a certain amount of compute running continuously, reserved pricing dramatically reduces costs. The risk is over-provisioning: if you reserve more than you use, you pay for unused capacity.

Spot Instances, called Preemptible VMs on Google Cloud, offer the deepest discounts, up to 90 percent off on-demand prices, but the provider can reclaim the capacity with short notice, typically two minutes. Spot pricing is excellent for fault-tolerant workloads like batch processing, data analysis, CI/CD pipelines, and any task that can handle interruption and restart. They are not suitable for workloads requiring continuous availability.

The optimal cost strategy layers these models. Reserved capacity covers the baseline load that runs continuously. On-demand handles expected variability above the baseline. Spot instances handle batch processing and other interruptible workloads. Auto-scaling groups can be configured to use a mix, launching spot instances first and falling back to on-demand when spot capacity is unavailable.

Additionally, per-request pricing models like serverless functions charge only for actual execution, making them extremely cost-effective for sporadic workloads with long idle periods between invocations. The key is matching each workload's characteristics to the pricing model that minimizes cost while meeting availability requirements.