.png)
This guide compares those tradeoffs, shows how utilization changes commitment economics, and provides a framework for combining all three options.
The Short Answer
Use On-Demand Instances for short-term, irregular, or changing workloads that cannot be interrupted. Use Reserved Instances when eligible EC2 usage is predictable enough to match the reservation throughout its one- or three-year term.Use Spot Instances for flexible, fault-tolerant workloads that can recover from interruption. Most production environments benefit from combining commitments for a stable baseline, On-Demand for uncertain demand, and Spot for suitable elastic work.
On-Demand, Reserved, and Spot Compared
These options differ in what you trade for a lower rate. On-Demand preserves purchasing flexibility, Reserved Instances require a term commitment, and Spot requires architectural flexibility.| Dimension | On-Demand Instances | Reserved Instances | Spot Instances |
|---|---|---|---|
| Commitment | None | One or three years | None |
| Advertised savings | No discount | Up to 72% versus On-Demand | Up to 90% versus On-Demand |
| Price behavior | Fixed listed rate while running | Discount applies to matching On-Demand usage | Price varies by instance type and Availability Zone |
| Interruption | You control the lifecycle | You control the underlying instance lifecycle | EC2 can interrupt when it needs capacity back |
| Capacity | Launch depends on available capacity | Regional RIs provide a discount; Zonal RIs also reserve matching capacity | Runs only while Spot capacity is available |
| Best fit | Irregular or evolving workloads | Predictable, matching usage | Fault-tolerant, flexible workloads |
For a durable baseline that may change, compare Savings Plans first: Compute Savings Plans cover eligible EC2 usage across families and Regions, plus eligible Fargate and Lambda usage, while EC2 Instance Savings Plans remain within one family and Region. Savings Plans discount usage but do not reserve EC2 capacity.
On-Demand Instances
With EC2 On-Demand Instances, you pay for compute capacity by the second with no long-term commitment. A 60-second minimum applies, and AWS recommends the model for short-term, irregular workloads that cannot be interrupted.On-Demand does not guarantee that every launch will succeed. Requests remain subject to available capacity and account quotas, so a separate Capacity Reservation may be appropriate when launch assurance matters.
Reserved Instances
An EC2 Reserved Instance is not a separate physical instance. It is a billing discount that applies when On-Demand usage matches the RI’s attributes, including factors such as instance type, Region, platform, and tenancy.AWS offers one- and three-year terms, with All Upfront, Partial Upfront, and No Upfront payment options. Standard RIs provide the highest advertised RI savings, while Convertible RIs can be exchanged for other Convertible RIs under AWS’s exchange rules.
Scope matters. Regional RIs can apply across Availability Zones and eligible Linux/Unix Regional RIs support size flexibility within an instance family; Zonal RIs provide matching capacity in one Availability Zone. See the complete EC2 Reserved Instances guide before choosing an offering class, scope, or payment option.
Spot Instances
EC2 Spot Instances use spare EC2 capacity below On-Demand prices. Amazon EC2 sets a Spot price for each instance type and Availability Zone and adjusts it gradually based on long-term supply and demand.EC2 can terminate, stop, or hibernate a Spot Instance when it needs the capacity back. For stop and termination events, AWS provides a two-minute interruption notice on a best-effort basis.
Before using Spot, verify that workloads can checkpoint and recover; deploy across multiple compatible instance types and Availability Zones; use EC2 Auto Scaling or EC2 Fleet to replace capacity; and test the interruption-notice and rebalance-recommendation path.
What Each Option Costs
The advertised maximum is not the rate every workload receives. Actual EC2 cost depends on the instance, Region, operating system, tenancy, term, payment option, matching usage, andon Spot, the current capacity pool price.Use the current AWS price pages or APIs for a production calculation. Our broader EC2 pricing guide covers instance rates and bill components in more detail.
A utilization example
Assume a stable fleet would cost $10,000 per month at On-Demand rates. Also assume its applicable one-year RI offering costs 40% less when matching usage fully consumes the reservation; these are hypothetical inputs, not quoted AWS rates.| Scenario | Actual workload at On-Demand equivalent | RI payments | Realized savings |
|---|---|---|---|
| Workload remains stable for 12 months | $120,000 | $72,000 | $48,000, or 40% |
| Workload falls by 50% after six months | $90,000 | $72,000 | $18,000, or 20% |
This is why commitment evaluation must model the applicable rate and expected matching usage across the full term. For Spot, use recent Spot price history and interruption data for the instance types and Availability Zones you plan to use rather than relying on a fleet-wide average.
Which Model Fits Your Workload?
Start with two questions: Is the usage predictable across the commitment term, and can the application tolerate interruption? Those answers narrow the choice quickly.Use On-Demand when
The workload is temporary, irregular, or still being tested.
Demand changes too quickly to establish a reliable baseline.
Instance families, Regions, or architecture may change soon.
The application cannot tolerate Spot interruption and does not justify a commitment yet.
Use Reserved Instances when
Matching EC2 usage is expected to persist throughout the selected term.
The required Region, platform, tenancy, and instance characteristics are understood.
Capacity assurance from a Zonal RI in a specific Availability Zone is required.
The team can monitor utilization and adjust future purchases as the fleet changes.
Use Spot Instances when
Jobs can checkpoint, retry, or resume on replacement capacity.
Services are stateless and can redistribute work across instances.
Batch processing, CI/CD, rendering, analytics, or distributed training can tolerate capacity changes.
The architecture can diversify instance types and Availability Zones.
| Workload pattern | Starting choice | Why |
|---|---|---|
| New or unpredictable service | On-Demand | Preserve flexibility while usage develops |
| Durable matching baseline | RI or Savings Plan | Discount the usage expected to continue |
| Restartable batch or elastic work | Spot | Trade interruption tolerance for lower cost |
| Stable baseline with variable peaks | Commitment plus On-Demand | Cover the floor without committing the ceiling |
| Fault-tolerant baseline with bursts | Commitment, Spot, and On-Demand | Assign each demand layer to its suitable model |
Combine On-Demand, Reserved, and Spot
A blended strategy starts with workload classification, not a universal coverage percentage. The appropriate allocation depends on demand variability, architecture, planned migrations, and confidence in the forecast.Commit the durable floor. Use an RI or Savings Plan only for the hourly usage you reasonably expect to remain eligible throughout the term.
Leave uncertainty On-Demand. Keep new services, temporary projects, variable peaks, and likely migrations outside the committed floor until their usage becomes predictable.
Move suitable work to Spot. Use Spot for jobs that can recover, checkpoint, and spread across multiple capacity pools.
For sizing and coverage details, see the AWS Savings Plans guide.
Size each commitment in four steps: normalize historical hourly usage; remove demand likely to change because of announced migrations, rightsizing, service retirements, seasonal patterns, or architecture changes; subtract usage already covered by current commitments; and commit only the remaining durable hourly baseline. Historical usage is evidence, not a guarantee that the same baseline will continue.
How Usage.ai Fits the Strategy
We built Flex Insured Commitments for teams that want commitment savings without carrying the full downside of a changing forecast. At the billing layer, we analyze usage, make recommendations, and use the cloud provider’s API to purchase and manage commitments after customer approval.With Flex Insured Commitments, eligible teams can get up to 57% savings associated with a three-year AWS commitment with none of the commitment risk.
If eligible usage drops and a qualifying Flex Commitment costs more than equivalent On-Demand usage, we provide cashback protection on the covered difference, subject to current eligibility and program terms.
Our pricing is a percentage of realized savings. If we do not generate savings through the program, there is no savings-based fee.
Conclusion: Choose the Right EC2 Mix
Choose On-Demand for uncertainty, an RI or Savings Plan for a defensible baseline, and Spot for work engineered to tolerate interruption. Reassess the mix whenever demand or architecture changes rather than treating the initial purchase as permanent.Frequently asked questions
What is the main difference between the three?
On-Demand has no long-term commitment; RIs discount matching usage for a one- or three-year term; Spot uses spare capacity at a variable discounted price and can be interrupted. The choice is primarily between purchasing flexibility, forecast confidence, and interruption tolerance.
Do Reserved Instances guarantee EC2 capacity?
Only Zonal RIs include a capacity reservation for their specified Availability Zone. Regional RIs provide the billing discount across the Region but do not reserve capacity.
Can Spot Instances run production workloads?
Yes, when the production component can tolerate interruption and replace capacity automatically. Use diversification, health checks, checkpointing where applicable, and On-Demand or committed capacity for components that require continuity.
Can all three models be combined?
Yes. A common design uses commitments for predictable baseline usage, On-Demand for uncertain or peak demand, and Spot for fault-tolerant work; the proportions should reflect the actual environment rather than a fixed industry target.
What happens when RI usage drops?
The RI cost remains for the full term: an All Upfront RI has already been paid for, while Partial Upfront and No Upfront RIs continue generating discounted hourly charges even when matching usage disappears.
Depending on the RI, AWS may allow modification, a Convertible RI exchange, or a Standard RI Marketplace sale, but those options have conditions and do not amount to a general cancellation right.
