Amazon EC2 Pricing | The Complete 2026 Guide (Every Model, Hidden Costs & Real Examples)

A clear example of Amazon EC2 Pricing complexity is a simple t3.large instance in us-east-1, which costs $0.0832/hour — about $60.74 per month when run continuously. That seems straightforward at first. However, once you add a 200GB EBS volume, transfer 5TB of data out to the internet, keep three idle Elastic IPs, and deploy an Application Load Balancer, that initial $60 estimate quickly balloons to $612 per month. This gap between expected costs and actual bills is exactly why Amazon EC2 pricing remains one of the most misunderstood cost centers in cloud computing.

EC2 (Elastic Compute Cloud) typically represents 35–50% of total AWS spend for most organizations. According to Stacklet’s 2024 State of Cloud Usage Optimization survey, 78% of organizations estimate that 21–50% of their cloud budgets are wasted — with incorrect resource sizing, outdated instance types, and misconfigured networking among the top culprits. Understanding EC2 pricing isn’t just about knowing hourly rates; it’s about mastering a complex billing model that spans five pricing strategies, seven instance families, and at least six hidden cost categories that most teams discover only after the bill arrives.

This guide covers:

• All 5 EC2 pricing models with real cost numbers and when to use each
• Instance family pricing matrix for 2026, including Graviton4-based M8g, C8g, and R8g instances
• The 6 hidden costs most teams miss during estimation (EBS storage, data transfer, NAT Gateways, load balancers, Elastic IPs, and monitoring)
• 3 real calculation examples from different workload types with line-by-line breakdowns
• 6 optimization strategies that can cut costs by 40–70% without changing your architecture

The EC2 Billing Model — How AWS Charges You

AWS built EC2 on a pay-as-you-go model that eliminates upfront capital expenditure (CapEx) and converts infrastructure costs into operational expenditure (OpEx). You pay nothing until you launch an instance, and you stop paying compute charges when you terminate it — but understanding the specifics matters.

Billing granularity varies by operating system. Linux instances bill per second with a 60-second minimum, which means if you run a test instance for 3 minutes and 15 seconds, you pay for exactly 195 seconds. Windows, Red Hat Enterprise Linux (RHEL), and SUSE Linux Enterprise Server (SLES) bill per hour with full-hour rounding. That same 3-minute test on Windows bills for a full hour.

Billing starts the moment an instance enters the “running” state and stops when the instance terminates. Here’s the critical distinction: stopping an instance (not terminating) halts compute charges but continues billing for attached EBS (Elastic Block Store) storage, allocated Elastic IPs, and any snapshots. Many teams assume stopped instances cost nothing and discover hundreds of dollars in monthly EBS charges for instances they haven’t used in months. 

AWS bills monthly with charges aggregated across all services in your account. Most EC2 costs appear under the “EC2-Instances” line item in AWS Cost Explorer, but data transfer appears under “EC2-Other,” EBS storage under “EC2-EBS,” and load balancers under their own line items — making it easy to underestimate your true EC2-related spend by 30–40% if you only look at instance hours.

The 5 EC2 Pricing Models

1. On-Demand Instances

On-Demand instances offer maximum flexibility with zero commitment. You launch an instance, AWS bills you per second (Linux) or per hour (Windows), and you terminate it when finished. No contracts, no capacity reservations, no questions asked.

This flexibility comes at a premium — On-Demand represents the most expensive EC2 pricing option per unit of compute. It makes sense for unpredictable workloads that spike randomly, short-term projects lasting days or weeks, and development environments where you need instant access without planning capacity weeks in advance.

On-Demand pricing varies significantly by operating system. Windows adds approximately 30% to the Linux baseline, while RHEL adds 40–50%. SQL Server licenses on EC2 can double or triple the hourly cost, making Amazon RDS with Bring Your Own License (BYOL) worth investigating for database workloads.

Note that Windows licensing adds about 29% over Linux, and RHEL adds 48% in this example. Regional pricing differences can reach 20–25%, with us-east-1 (N. Virginia) typically offering the lowest rates and newer regions like ap-east-1 (Hong Kong) charging premium rates.

2. EC2 Savings Plans

Savings Plans let you commit to a consistent dollar-per-hour spend for one or three years in exchange for discounts up to 72% compared to On-Demand pricing. Unlike traditional Reserved Instances, Savings Plans focus on spending commitment rather than specific instance reservations, offering significantly more flexibility.

AWS offers two Savings Plan types:

EC2 Instance Savings Plans provide up to 72% discount but lock you to a specific instance family within a specific region. If you commit to M7g instances in us-east-1, your discount applies only to M7g instances running in us-east-1 — but you can freely change instance size (medium to xlarge), operating system (Linux to Windows), and tenancy (shared to dedicated) within that family. This model works perfectly for predictable workloads where you know you’ll need compute-optimized or memory-optimized instances consistently.

Compute Savings Plans provide up to 66% discount with maximum flexibility. Your commitment applies across any EC2 instance family, any region, and even extends to AWS Fargate and AWS Lambda. You can migrate from C5 instances in Virginia to M7g instances in Ireland to Lambda functions in Tokyo and maintain the same discounted rate. The flexibility costs you about 6 percentage points in maximum discount compared to EC2 Instance Savings Plans, but the operational freedom often justifies the tradeoff. 

Both types offer three payment models: All-Upfront (pay the entire commitment upfront for maximum discount), Partial-Upfront (pay roughly half upfront and the rest monthly for mid-tier discount), and No-Upfront (pay monthly with smallest discount but zero capital outlay).

Example: A t3.large Linux instance costs $0.0832/hr On-Demand. With a 1-year All-Upfront EC2 Instance Savings Plan, that rate drops to approximately $0.052/hr — a 37% discount that saves $445.92 annually per instance.

3. Reserved Instances (RIs)

Reserved Instances represent AWS’s original discount model: commit to a specific instance type in a specific region for one or three years, receive up to 72% off On-Demand rates. While AWS now recommends Savings Plans over RIs, Reserved Instances still offer two capabilities Savings Plans don’t: capacity reservations in specific Availability Zones (AZs) and slightly different pricing structures for certain workload patterns.

Standard RIs provide up to 72% savings with modification capabilities limited to Availability Zone, instance size within the same family, and networking type. You can convert a c5.xlarge RI to two c5.large RIs, but you cannot convert a C5 RI to an M5 RI without purchasing a new reservation.

Convertible RIs offer up to 66% savings with the ability to exchange your reservation for different instance families, operating systems, or tenancy models during the term. If your workload shifts from compute-optimized to memory-optimized requirements mid-contract, Convertible RIs let you adapt without losing your discount entirely.

Reserved Instances also offer volume discounts at enterprise scale: 5% additional discount on purchases between $500,000 and $4 million annually, and 10% additional discount beyond $4 million. The RI Marketplace lets you sell unused reservations to other AWS customers if your capacity needs change, providing an exit option Standard Savings Plans lack.

The key difference between RIs and Savings Plans: Reserved Instances reserve actual capacity in a specific Availability Zone (if you choose zonal RI), guaranteeing that capacity exists when you need to launch instances. Savings Plans offer billing discounts without capacity guarantees. For workloads requiring guaranteed capacity during peak demand — such as tax processing systems in March or retail platforms in November — zonal Reserved Instances provide insurance that Savings Plans cannot.

4. Spot Instances

Spot Instances let you bid on unused AWS capacity at discounts up to 90% compared to On-Demand pricing. When AWS has spare capacity in a particular Availability Zone and instance type, it offers that capacity as Spot at heavily discounted rates. When AWS needs that capacity back for On-Demand or Reserved Instance customers, it reclaims your Spot instance with a two-minute interruption notice.

This model works brilliantly for stateless, fault-tolerant, and flexible workloads: batch processing jobs that checkpoint progress every few minutes, CI/CD build pipelines that can restart failed jobs automatically, big data analytics using frameworks like Spark that handle node failures gracefully, machine learning training runs with periodic model checkpoints, and containerized microservices behind load balancers where losing one container doesn’t impact the application.

5. Dedicated Hosts

Dedicated Hosts allocate an entire physical server exclusively to your AWS account. Rather than billing per instance, AWS bills per host-hour regardless of how many instances run on that server. A c5.9xlarge Dedicated Host supports up to eight c5.xlarge instances, four c5.2xlarge instances, or any combination that fits the host’s physical capacity.

This model serves two primary use cases. First, Bring Your Own License (BYOL) scenarios where your Microsoft Windows Server, SQL Server, or Oracle Database licenses are tied to physical cores or sockets rather than virtual machines. Running on Dedicated Hosts lets you leverage existing licenses instead of paying AWS’s license-included pricing, potentially saving 40–60% on Windows and SQL Server workloads with existing Enterprise Agreements.

EC2 Instance Families — Complete 2026 Pricing Matrix

EC2 offers over 700 instance types across six major families, each optimized for specific workload characteristics. Selecting the appropriate family dramatically impacts both performance and cost — a memory-optimized r7g.large costs $0.1512/hr while a general-purpose m7g.large with less RAM costs $0.0816/hr. Running memory-intensive workloads on under-provisioned general-purpose instances creates performance problems; running CPU-bound workloads on over-provisioned memory-optimized instances wastes 40–50% of your budget.

General Purpose (T4g, M7g, M8g)

General-purpose instances provide balanced CPU, memory, and networking resources suitable for web servers, application servers, development environments, and small-to-medium databases. This family splits into two distinct categories with different billing characteristics.

T-series instances (T4g, T3, T3a) use a CPU credit model for burstable performance. You accumulate credits when CPU utilization stays below the baseline (20% for t3.large, 40% for t3.xlarge) and spend credits when bursting above baseline. For low-utilization workloads — development servers active 2–3 hours daily, batch processing jobs with idle periods, staging environments — T-series delivers the same performance as M-series at 30–40% lower cost. For steady-state workloads using consistent CPU, T-series instances exhaust credits and either throttle performance or charge $0.05 per vCPU-hour for surplus credits in Unlimited mode, often costing more than M-series.

M-series instances (M8g, M7g, M7i, M6i) provide consistent performance without credit systems. M8g instances powered by AWS Graviton4 processors deliver the best price-performance in this family, offering 40% better performance per dollar than comparable x86 instances. Most web applications, API servers, and general workloads run perfectly on M8g with zero code changes for common languages (Python, Java, Node.js, Go, .NET Core, Ruby).

Compute Optimized (C7g, C8g)

Compute-optimized instances maximize CPU performance for CPU-bound workloads: high-traffic web servers handling thousands of requests per second, video encoding and transcoding pipelines, scientific modeling and simulation, batch processing that’s CPU-constrained rather than I/O-constrained, and high-performance computing (HPC) applications that need maximum single-threaded performance.

C8g instances (Graviton4) represent the 2026 best-practice recommendation, delivering up to 30% better performance than Graviton3-based C7g while maintaining similar pricing. C7i instances (Intel Xeon 4th gen) serve workloads requiring x86-specific instruction sets or legacy software incompatible with ARM architecture, though this compatibility requirement shrinks monthly as cloud-native tools mature.

Memory Optimized (R7g, R8g, X-series)

Memory-optimized instances provide high memory-to-vCPU ratios for in-memory databases (Redis, Memcached, SAP HANA), big data analytics, real-time data processing streams, and large caching tiers. R-series instances offer ratios around 8 GiB RAM per vCPU; X-series pushes to 16–32 GiB per vCPU for the most memory-intensive workloads.

R8g instances (Graviton4) deliver the best price-performance in this family, with r8g.large at $0.1512/hr providing 8 GiB RAM per vCPU. For extremely large in-memory datasets, X2gd instances scale to 1 TiB RAM, and x2iedn.32xlarge reaches 3.8 TiB — enough to hold massive Redis clusters or entire relational database working sets in memory.

Accelerated Computing (P5, G6, Trn2, Inf2)

Accelerated computing instances integrate GPUs, FPGAs, or custom AI accelerators for machine learning, 3D rendering, video processing, and high-performance computing requiring massive parallel processing. This family spans from inference to training to general GPU compute.

P5 instances with NVIDIA H100 Tensor Core GPUs target large-scale AI training and high-performance computing, starting around $27.39/hr for p5.48xlarge — suitable only for enterprise AI initiatives with substantial budgets. G6 instances with NVIDIA L4 GPUs provide cost-effective ML inference and graphics workloads starting at $0.8048/hr for g6.xlarge.

Trainium2 (Trn2) and Inferentia2 (Inf2) are AWS-designed chips purpose-built for AI workloads. Trn2 delivers up to 4× better price-performance than GPU-based training for deep learning models, while Inf2 provides the lowest cost per inference for deploying trained models at scale. Organizations training transformers, large language models, or computer vision models should evaluate Trainium pricing before defaulting to GPU instances — monthly savings often exceed 60–70%.

Storage Optimized (I4g, Im4gn)

Storage-optimized instances deliver high sequential read/write performance and extreme IOPS using local NVMe SSD storage. NoSQL databases (Cassandra, MongoDB, ScyllaDB) that need terabytes of fast local storage, data warehouses running analytics across petabyte-scale datasets, distributed file systems, and log processing pipelines use this family.

I4g instances powered by Graviton processors provide up to 30 TB of local NVMe storage per instance with up to 250,000 IOPS per volume. Im4gn instances optimize for memory-to-storage ratio with up to 32 TiB RAM and 30 TB NVMe storage on the largest sizes.

Critical consideration: Local NVMe storage is ephemeral. When the instance stops or terminates, data disappears. Storage-optimized instances work for distributed systems that replicate data across multiple nodes (Cassandra replicates 3×; MongoDB replica sets maintain 3 copies) where losing one node doesn’t lose data. Single-instance databases requiring data persistence need EBS-backed storage instead.

Key Factors That Affect Your EC2 Bill

Beyond instance type selection, several variables significantly impact your monthly EC2 costs. Understanding these factors lets you optimize spending without changing your architecture or instance families.

Region and Availability Zone selection creates 15–25% pricing variance. US East (N. Virginia / us-east-1) typically offers the lowest rates globally due to AWS’s highest concentration of capacity and infrastructure investment. US West (N. California / us-west-1) costs 15–20% more; Asia Pacific regions (Singapore, Mumbai, Hong Kong) add 20–30%. Newer regions command premium pricing — ap-east-1 (Hong Kong) and ap-southeast-3 (Jakarta) rank among the most expensive globally. For latency-insensitive workloads (batch processing, data analytics, non-customer-facing systems), consolidating in us-east-1 cuts costs substantially. Customer-facing applications requiring low latency may justify premium regional pricing, but dev/test environments rarely need regional distribution.

Operating system licensing creates dramatic cost differences. Linux and Amazon Linux 2 represent baseline pricing. Windows Server adds 30–40% due to Microsoft licensing fees embedded in hourly rates. Red Hat Enterprise Linux adds 40–50% for subscription access. SQL Server on Windows roughly doubles instance costs — a t3.large jumps from $0.0832/hr (Linux) to $0.1072/hr (Windows) to $0.2656/hr (Windows with SQL Server Web). If you run SQL Server workloads, evaluate Amazon RDS with BYOL licenses or migrating to open-source PostgreSQL; monthly savings often reach $500–$2,000 per database.

Processor architecture impacts both performance and cost. Graviton (ARM) instances cost 10–20% less than equivalent x86 instances while delivering equal or better performance for most cloud-native workloads. M7g (Graviton3) costs about 15% less than M7i (Intel 4th gen) with better single-threaded performance. M8g (Graviton4) extends this advantage with 40% better price-performance than previous generations. Most applications built on Java, Python, Node.js, Go, Rust, .NET Core, Ruby, or PHP migrate to Graviton with zero code changes — simply rebuild your container images for ARM64 architecture and redeploy. Legacy x86-only dependencies are shrinking rapidly; evaluate Graviton for new workloads and consider migration paths for existing systems to capture 15–20% immediate savings.

Tenancy models create cost layers. Shared tenancy (default) provides the lowest cost with instances running on shared physical hardware. Dedicated Instances add a 10% per-instance surcharge for hardware isolation, billed separately from instance hours. Dedicated Hosts bill per physical server rather than per instance, costing 20–40% more than shared tenancy unless you can maximize host utilization or leverage BYOL licensing that offsets infrastructure premiums. Most workloads run perfectly on shared tenancy; reserve dedicated options for genuine compliance requirements or BYOL scenarios with proven ROI.

Instance generation significantly affects price-performance. AWS releases new instance generations every 12–24 months with better performance at similar or lower pricing compared to previous generations. M7g (7th gen) typically costs the same per hour as M6g (6th gen) while delivering 20–25% more performance, effectively reducing cost-per-transaction by 20%. Always select the latest generation available in your region unless you have Reserved Instances or Savings Plans locked to older generations — and when those commitments expire, migrate immediately to capture price-performance improvements.

The 6 Hidden EC2 Costs (That Blow Up Your Estimate)

Most EC2 cost estimates focus exclusively on instance hours, missing six additional charges that collectively add 30–60% to your actual monthly bill. These “hidden” costs aren’t secret — AWS documents them thoroughly — but they appear on separate line items in billing reports, making them easy to overlook during planning.

1. EBS Storage

EC2 instances need storage volumes, and unless you explicitly select instance store (ephemeral local storage that disappears when the instance stops), you’re paying for Elastic Block Store (EBS) volumes. EBS storage bills independently from compute hours — if you launch a t3.large with a 200GB root volume, stop the instance after one hour of testing, and forget about it for a month, you pay $0.08 for compute and $16.00 for the storage volume sitting idle.

gp3 SSD volumes cost $0.08 per GB-month with 3,000 IOPS and 125 MB/s throughput included free. This represents AWS’s best general-purpose storage option — 20% cheaper than legacy gp2 volumes ($0.10 per GB-month) with better baseline performance. Additional IOPS cost $0.005 per provisioned IOPS-month above the 3,000 baseline, and additional throughput costs $0.04 per provisioned MB/s-month above 125 MB/s baseline.

Example: A production t3.large with 200GB gp3 storage costs $60.74/month for compute and $16/month for storage — 26% more than compute alone. Add daily snapshots with 14-day retention (approximately 80GB total snapshot storage) for another $4/month, bringing total cost to $80.74/month.

2. Data Transfer / Egress

Data transfer into AWS from the internet is free. Data transfer out from AWS to the internet (egress) costs substantial money on tiered pricing that catches unprepared teams off-guard.

Internet egress pricing: First 100 GB per month free (aggregated across all AWS services globally), $0.09 per GB for the next 10 TB, $0.085 per GB for the next 40 TB, $0.07 per GB for the next 100 TB, $0.05 per GB beyond 150 TB total. A busy API server transferring 5 TB monthly to external clients costs approximately $382.50 just for egress (after the 100GB free tier).

Cross-AZ data transfer within the same region costs $0.01 per GB each direction. If your application tier in us-east-1a queries a database in us-east-1b, each query’s response data costs $0.01 per GB from database to application, and if replication runs from us-east-1b back to us-east-1a, another $0.01 per GB. Multi-AZ architectures with high internal traffic accumulate hundreds of dollars monthly in cross-AZ charges that same-AZ deployments avoid entirely.

Cross-region transfer between AWS regions costs $0.02 per GB for most region pairs. Replicating 1 TB daily from us-east-1 to eu-west-1 for disaster recovery costs $20/day or $600/month in data transfer alone, independent of storage or compute costs.

EC2 to S3 in the same region is free. EC2 to RDS in the same region and same AZ is free; different AZs charge $0.01 per GB each direction. Architectures that minimize cross-AZ database queries and consolidate data-intensive operations within single AZs reduce transfer costs substantially.

Organizations often discover egress as their third-largest AWS line item after compute and storage. A microservices architecture with 50 services making millions of API calls daily across AZs can accumulate $2,000–$5,000 monthly in data transfer charges that architectural consolidation or same-AZ deployment reduces by 60–80%.

3. Elastic IP Addresses

AWS provides one free Elastic IP address per running instance in your account. Additional Elastic IPs and all idle (unassociated) Elastic IPs cost $0.005 per hour ($3.60 per month) per address.

This creates two cost traps. First, teams allocate Elastic IPs for temporary testing, terminate the instances, forget to release the IPs, and pay $3.60/month indefinitely for unused addresses. Ten idle Elastic IPs scattered across accounts from old POCs cost $36/month for resources providing zero value.

Second, architectures requiring multiple Elastic IPs per instance pay $3.60/month per additional IP. A NAT instance with 5 public IPs (one free, four chargeable) adds $14.40/month in IP allocation costs before considering compute or data transfer.

Best practice: Regularly audit Elastic IPs using AWS Cost Explorer’s “EC2-Other” category filtering for “IdleAddress.” Release any unassociated addresses immediately. For production systems requiring static IPs, confirm the business genuinely needs IP stability versus using DNS with dynamic IPs that avoid allocation charges entirely.

4. NAT Gateway

Private subnet instances lacking direct internet access require Network Address Translation (NAT) to reach external services. While you can build NAT instances using EC2 (self-managed), most teams use managed NAT Gateway service for simplicity and high availability.

NAT Gateway bills $0.045 per hour ($32.40 per month) plus $0.045 per GB of data processed through the gateway. An application processing 1 TB monthly through NAT Gateway pays $32.40 for availability plus $45.00 for data processing — $77.40 total.

Multi-AZ architectures often deploy one NAT Gateway per AZ for fault tolerance, multiplying costs. Three NAT Gateways cost $97.20/month in hourly charges alone before processing any data. Organizations with dozens of VPCs and multiple production environments sometimes operate 20–30 NAT Gateways, creating $650–$970/month in baseline NAT infrastructure costs.

5. Elastic Load Balancer (ALB/NLB)

Production applications typically use load balancers for traffic distribution, health checks, and SSL termination. AWS offers three types with different pricing models.

Application Load Balancer (ALB) costs $0.0225 per hour ($16.20 per month) plus $0.008 per LCU-hour where LCU (Load Balancer Capacity Unit) measures the maximum of: new connections per second (25 connections = 1 LCU), active connections per minute (3,000 connections = 1 LCU), processed bytes per hour (2.22 GB = 1 LCU for HTTP, 1 GB = 1 LCU for HTTPS), or rule evaluations per second (1,000 evaluations = 1 LCU). A moderately busy application using 3 LCUs continuously costs $16.20 hourly plus (3 LCUs × $0.008 × 730 hours) = $33.72 — total $49.92/month.

Network Load Balancer (NLB) costs $0.0225 per hour plus $0.006 per NLCU-hour with different capacity dimensions optimized for TCP/UDP traffic. Pricing typically runs 10–20% lower than ALB for equivalent traffic due to lower NLCU rates, though the hourly charge remains identical.

Classic Load Balancer (CLB) represents legacy technology AWS no longer recommends, charging $0.025 per hour plus $0.008 per GB data processed. Migration to ALB or NLB typically reduces costs 20–30% while providing better features.

6. Premium Support & Monitoring

AWS Business Support costs 10% of monthly usage for the first $10,000, 7% for the next $80,000, 5% for the next $250,000, and 3% above $340,000. A $5,000 monthly AWS bill adds $500 for Business Support. Enterprise Support costs substantially more with minimums around $15,000/month, though it includes dedicated Technical Account Managers and architectural reviews valuable at enterprise scale.

CloudWatch detailed monitoring enables 1-minute metric resolution instead of 5-minute basic monitoring (free). Detailed monitoring costs $0.01 per metric per month. An EC2 instance with detailed monitoring emits roughly 30 metrics, costing $0.30/month per instance. Multiply by 100 instances across environments and CloudWatch detailed monitoring adds $30/month — seemingly small, but combined with CloudWatch Logs ingestion ($0.50 per GB), CloudWatch Alarms ($0.10 per alarm per month), and CloudWatch Dashboards ($3 per dashboard per month), comprehensive monitoring costs escalate to $200–$500/month for mid-size deployments.

Real-World EC2 Cost Calculations

Example 1 — Small Web Server (Single Instance)

A small business runs a WordPress site on EC2 with straightforward requirements: one always-on instance, moderate storage, minimal traffic.

Specifications:

• Instance: t3.medium (2 vCPU, 4 GiB RAM)
• Region: us-east-1
• OS: Linux (Amazon Linux 2)
• Availability: 730 hours/month (continuous operation)
• Storage: 50 GB gp3 SSD
• Egress: 100 GB/month to internet
• Pricing model: On-Demand

Line-by-Line Calculation:

• EC2 On-Demand: $0.0416/hr × 730 hrs = $30.37
• EBS gp3 Storage: 50 GB × $0.08/GB = $4.00
• Data Transfer: 100 GB free tier = $0.00
• Total On-Demand: $34.37/month

With 1-Year Reserved Instance (No-Upfront):

• EC2 Reserved: $0.026/hr × 730 hrs = $18.98
• EBS gp3 Storage: 50 GB × $0.08/GB = $4.00
• Data Transfer: 100 GB free tier = $0.00
• Total Reserved: $22.98/month
• Annual Savings: ($34.37 – $22.98) × 12 = $136.68/year (37% reduction)

For predictable workloads running continuously, Reserved Instances or Savings Plans cut costs by one-third with zero architectural changes. The business saves enough annually to fund a second environment for dev/test purposes or invest in performance optimization.

Example 2 — ML Training Workstation (Spot)

A data science team trains machine learning models intermittently, running computationally intensive training jobs that can tolerate interruptions and restart from checkpoints.

Specifications:

• Instance: g6.xlarge (4 vCPU, 16 GiB RAM, 1 NVIDIA L4 GPU)
• Region: us-east-1
• OS: Linux (Deep Learning AMI)
• Usage: 100 hours/month (training jobs)
• Storage: 200 GB gp3 SSD
• Egress: 50 GB/month (model outputs)
• Pricing model: Spot

On-Demand Comparison:

• EC2 On-Demand: $0.8048/hr × 100 hrs = $80.48
• EBS gp3 Storage: 200 GB × $0.08/GB = $16.00
• Data Transfer: 50 GB free tier = $0.00
• Total On-Demand: $96.48/month

With Spot Pricing:

• EC2 Spot (approx 65% discount): $0.28/hr × 100 hrs = $28.00
• EBS gp3 Storage: 200 GB × $0.08/GB = $16.00
• Data Transfer: 50 GB free tier = $0.00
• Total Spot: $44.00/month
• Monthly Savings: $96.48 – $44.00 = $52.48 (54% reduction)

Spot pricing transforms GPU instances from prohibitively expensive to reasonably affordable for teams without enterprise budgets. The team implements model checkpointing every 10 minutes so interruptions lose at most 10 minutes of training progress, making interruption risk minimal compared to 54% cost savings. Annual savings of $629.76 funds additional experimentation and faster iteration cycles.

How to Choose the Right EC2 Pricing Model

Selecting optimal pricing requires matching workload characteristics to commitment models. This decision framework helps navigate the five options:

Hybrid strategies often deliver optimal results. Cover your baseline steady workload with Savings Plans or Reserved Instances (capturing 50–72% discounts), handle variable load above baseline with On-Demand instances (maintaining flexibility), and run batch processing on Spot (achieving 90% discounts). This three-tier approach reduces overall compute costs by 40–60% compared to pure On-Demand while maintaining operational flexibility.

6 EC2 Cost Optimization Best Practices

1. Right-size before committing. AWS Compute Optimizer analyzes actual usage patterns and recommends optimal instance types based on CPU, memory, network, and disk metrics from the past 14 days. Before purchasing Reserved Instances or Savings Plans for t3.xlarge instances, confirm Compute Optimizer doesn’t recommend t3.large (half the cost) based on actual 30% average CPU utilization. Over-provisioned instances locked into 3-year commitments waste 40–50% of reservation value. Review Compute Optimizer recommendations before any commitment.
2. Migrate to Graviton. AWS Graviton4 (M8g, C8g, R8g) delivers 10–20% lower pricing than equivalent x86 instances with equal or better performance. Most modern applications built on Java, Python, Node.js, Go, Rust, .NET Core, Ruby, PHP, or containerized workloads migrate to Graviton with zero code changes — rebuild Docker images for ARM64, redeploy, validate performance, and capture immediate 15–20% cost reduction. Legacy applications with x86-only dependencies shrink monthly as libraries add ARM support. Evaluate Graviton for all new workloads; build migration plans for existing systems to compound savings over time.
3. Use Spot for interruptible workloads. Any workload that checkpoints progress and handles interruptions gracefully belongs on Spot instances: batch data processing (Spark, Hadoop, EMR), CI/CD build pipelines, machine learning training runs with periodic model saves, video rendering and transcoding, scientific simulations, ETL pipelines, and containerized stateless services behind load balancers with auto-scaling. Set up Spot Fleet or EC2 Auto Scaling with mixed instance types to maintain capacity even when specific instance types face interruption. Organizations running 30–40% of compute on Spot reduce overall EC2 spend by 25–35%.
4. Automate instance scheduling. Development and testing environments rarely need 24/7 availability. AWS Instance Scheduler or third-party tools automatically stop instances nights and weekends based on tags and schedules. If developers work 9 AM–6 PM weekdays, that’s 45 hours weekly versus 168 hours continuous — 73% compute cost reduction with zero functionality loss. Even conservative scheduling (stopping instances 12 hours daily on weekdays) cuts costs 50%. Tag instances with “Environment:Dev” or “Schedule:OffHours” and implement automated stopping for non-production workloads.
5. Delete orphaned resources. Unattached EBS volumes accumulate when teams terminate instances without deleting associated volumes, costing $0.08–$0.125 per GB-month indefinitely for unused storage. Idle Elastic IPs allocated for temporary testing cost $3.60/month each when not associated with running instances. Old AMIs and forgotten snapshots pile up in accounts, each costing $0.05 per GB-month. Schedule monthly audits using AWS Cost Explorer to identify “EC2-EBS” charges for unattached volumes, “EC2-Other” charges for idle addresses, and snapshot storage for outdated backups beyond retention policies. Many organizations find 15–20% of total EBS costs tie to orphaned resources providing zero value.
6. Monitor and alert proactively. Create CloudWatch billing alarms at meaningful thresholds ($X daily spend, $Y monthly forecast) to catch cost spikes before month-end surprises. Review AWS Cost Explorer weekly filtering by service, instance type, and usage type to identify anomalies. Implement cost allocation tags (Team, Project, Environment) on all resources so Cost Explorer breaks down spending by business unit or application, enabling accountability and budget enforcement. Organizations that review costs weekly instead of monthly catch misconfigurations (accidentally launching 50× p5 instances, forgetting to terminate test environments) within days instead of weeks, preventing $10,000+ billing mistakes.

EC2 Pricing FAQs

How much does Amazon EC2 cost per month?

EC2 costs vary by instance and usage. Small servers cost $30–$35/month, dev setups $50–$200/month, and most production apps $200–$2,000/month. High-end GPU instances can exceed $26,000/month.

Is Amazon EC2 free?

Yes, AWS offers a 12-month Free Tier with 750 hours/month of a micro instance plus limited storage. It’s suitable for learning, not production.

What is the cheapest EC2 instance?

The cheapest option is t4g.nano at about $3/month. For basic real-world use, t4g.micro (~$6/month) is more practical.

What is the difference between Savings Plans and Reserved Instances?

Both give up to 72% discounts with long-term commitments. Savings Plans are flexible across services, while Reserved Instances lock specific instance types and regions with capacity guarantees.

How does EC2 pricing compare to AWS Lambda?

EC2 is cheaper for always-on workloads. Lambda is cheaper for short, infrequent, event-based tasks. Choose based on usage pattern.

Conclusion

Understanding Amazon EC2 Pricing goes far beyond checking an hourly rate. Real-world EC2 costs are influenced by pricing models, instance families, regions, operating systems, and — most importantly — hidden charges such as storage, data transfer, NAT gateways, load balancers, and idle resources. As shown throughout this guide, workloads that appear inexpensive at first can quickly scale into significantly higher monthly bills if these factors are overlooked.

The key to cost-efficient EC2 usage in 2026 is alignment: matching workload behavior to the right pricing model, selecting the correct instance family, adopting Graviton where possible, and combining On-Demand, Spot, and commitment-based discounts strategically. For predictable workloads, EC2 Savings Plans are one of the most effective ways to reduce long-term compute costs, and readers who want a deeper understanding of commitment-based pricing can explore our detailed guide on EC2 Savings Plans.

At GoCloud, we help organizations turn complex AWS pricing into predictable, optimized cloud spend. Whether you’re estimating your first EC2 deployment or optimizing a large, multi-account AWS environment, the right pricing strategy is the difference between cloud efficiency and unnecessary cloud waste.