Compute pricing
AWS compute from EC2 instances and families to containers and serverless, and Google Cloud Compute Engine machine types and series, priced in every region each provider sells them.
AWS
EC2, containers and serverless
All EC2 Instances
Search and filter 1,268+ instance types with live multi-attribute filtering by vCPU, memory, hourly cost, CPU architecture, and GPU presence.
- Linux & Windows on-demand pricing across all regions
- x86_64 and arm64 Graviton architecture filters
- Live unit economics ($/vCPU-hr, $/GiB-hr)
EC2 Instance Families
Browse 195 families organized into 6 workload-optimized categories with single-row tab navigation and generational comparison.
- 6 core architectural workload categories
- Detailed family descriptions and generation specs
- GPU, storage, and networking capability badges
Containers: ECS or EKS
Choose between ECS and EKS: AWS's simpler native orchestrator or its managed Kubernetes. See what each costs, with Fargate as a compute option under either.
- ECS: no platform fee on your own EC2
- EKS: cluster, Auto Mode and Hybrid Nodes fees
- Fargate rates by region for either platform
AWS Lambda
Serverless functions billed per request and per GB-second of run time. Estimate a function's month in any region, with the free tier and provisioned concurrency.
- Duration volume tiers for x86 and Graviton
- SnapStart, Lambda@Edge and other meters by region
- Managed Instances fee beside the EC2 price
AWS Savings Plans
Price a workload with AWS's published Compute and EC2 Instance Savings Plans rates, then size the hourly commitment and see what an over- or under-commitment costs.
- Compute and EC2 Instance plans, 1 or 3 years
- No, Partial and All Upfront rates side by side
- Break-even utilization and unused commitment
Google Cloud
Compute Engine
Also on Google Cloud: Machine types, Machine series →
Google Cloud machine types
Filter 578 Compute Engine machine types by vCPU, memory, machine series and price, with a listing for each of the 43 regions they run in.
- On-demand, Spot and resource-based CUD rates
- Linux and Windows, composed from per-vCPU and per-GiB rates
- Live unit economics ($/vCPU-hr, $/GiB-hr)
Google Cloud machine series
Browse 32 machine series, such as N2 and C4A, grouped into Google's 6 machine families by workload.
- General-purpose to accelerator-optimized
- CPU vendor and architecture per series
- Committed use discount range per series
Compare Instances
Two different tools for two different questions: build your own side-by-side for the instances you're actually deciding between, or read a pre-built breakdown of the classic architecture tradeoffs.
Family Comparisons & Silicon Benchmarks
26 engineered comparison guides covering the Silicon Triangle (Intel vs AMD vs Graviton), generational upgrades, and NVMe right-sizing — with a verdict, not just a table.
C7i vs C7a vs C7g vs C7i-flex
C7g (Graviton3) delivers the highest CoreMark performance-per-dollar and lowest raw hourly price (~19% cheaper than C7i). C7a (AMD Genoa) leads on raw x86 single-thread throughput. C7i-flex offers a 5% discount over C7i for unconstrained burstable web services.
M7i vs M7a vs M7g vs M7i-flex
For modern containerized web fleets, M7g is the runaway value leader with ~20% lower cost than M7i. For enterprise Java or memory-bound x86 monoliths, M7a delivers blistering cache and memory performance.
R7i vs R7a vs R7g
R7g offers the lowest cost per GiB of RAM on AWS for standard relational and in-memory databases. R7a offers up to 1.5 TiB of memory on top sizes with AMD Genoa's exceptional memory throughput.
T4g vs T3 vs T3a
T4g is roughly 20% cheaper than T3 and outperforms both T3 and T3a on raw CPU benchmarks. It is the definitive default choice for dev/test and burstable production workloads.
Instance Categories Quick Jump
View all categories →Balanced compute, memory, and networking for web servers, microservices, and general workloads.
High clock-speed processors for CPU-bound applications, batch processing, and media transcoding.
High RAM-to-vCPU ratios for in-memory databases, Redis caches, and real-time big data processing.
Low latency, high NVMe I/O throughput for data warehousing, Elasticsearch, and high-frequency OLTP.
NVIDIA GPUs, AWS Trainium, and Inferentia hardware acceleration for AI/ML and 3D rendering.
Clusters built for computational fluid dynamics, weather forecasting, and complex simulations.
Guides & Analysis
Decision Guides
- The Windows licensing premium: what it actually costsWindows Server pricing on EC2 is a flat, additive per-vCPU license fee stacked on top of the Linux price, not a vague markup. Real numbers on how much, how it scales, and when it's worth avoiding.
- T-family vs. M-family: picking the right instance for variable loadA decision checklist for choosing between burstable T instances (t3, t4g) and fixed-performance M instances. Real on-demand pricing showing the T discount at baseline, and when that discount stops applying.
- Right-sizing EC2 with $/vCPU-hr and $/GiB-hrA repeatable method for picking instance size and family using cost-per-vCPU-hour and cost-per-GiB-hour instead of habit. Includes real pricing data showing whether bigger sizes get cheaper, and how compute, general-purpose, and memory-optimized families trade off.
- Choosing a GPU instance on cost, not habitA method for picking accelerated EC2 instances using $/GPU-hour and GPU memory per dollar instead of defaulting to the biggest or most familiar type. Real pricing across every GPU family in our dataset, why memory is usually the binding constraint, and when going wide beats going big.
- When a .metal instance is actually worth itAWS prices .metal instances the same as the largest virtualized size in their family, to the cent. The real cost of bare metal isn't the hourly rate, it's what you take on by running without a hypervisor. Here's how to tell if that trade is worth making.
- Graviton vs. x86: what the discount is worthAWS Graviton (arm64) instances are usually cheaper per hour than their x86_64 equivalents, but the savings only materialize if your workload can run on arm64 without a costly migration. Prices, benchmarks, vCPUs vs. cores, and a decision checklist.
- Everyone thinks their service needs a C instanceCompute-optimized instances get picked by habit, and of the three families that look like a match on a spec sheet, C is the one that demands you actually need the vCPUs. How to check whether the service touches its cores, why memory is usually the binding constraint, and how a move from C to M at the same memory cuts the bill by about 44%.
- How Compute Savings Plans decide what they coverHow AWS decides which usage a Compute Savings Plan covers each hour, with AWS's five worked scenarios, an interactive commitment slider, and what a higher discount changes.
- Managing AWS Compute Savings Plans: Elasticity, Seasonality, and Planned WasteWhy targeting 100% Savings Plan utilization leaves money on the table. How weekly diurnal cycles create unavoidable waste, the break-even math of off-peak hours, and how to size commitments for elastic workloads.
- ECS on EC2 or Fargate: finding the break-evenFargate costs $72.08 a month for the 2 vCPU / 4 GB workload a $62.05 c6i.large node runs. The crossover is utilisation, and in us-east-1 it sits at 86% of the month. Here is the arithmetic, the cell that changes it, and a checklist.
From the Blog
- Graviton versions: from experiment to defaultGraviton1 (2018) was an experiment nobody was sure about. Graviton2 made Arm a mainstream choice, and each generation since has shifted the price-performance math. The story of all five generations, with numbers from our data.
- The three-way CPU race: how AMD flipped the script on EC2The old rule of thumb was simple: Intel was default, AMD was 10% cheaper, and Graviton was 20% cheaper. In the 7th and 8th generations, AMD Genoa turned that pricing model upside down. Benchmarks, clock speeds, and real cost per compute unit.
- The RAM per dollar curve: how C, M, R, and X families price memoryCompute-optimized, general-purpose, and memory-optimized instances charge drastically different effective rates per gigabyte of memory. The math behind the memory curve, why buying RAM on M-instances wastes money, and sizing formulas for memory-bound workloads.
- The burstable CPU credit trap: when T instances cost more than MT2, T3, and T4g instances look like the cheapest compute on AWS until sustained CPU load triggers T Unlimited surcharges at $0.05 per vCPU-hour. How the credit accounting works, when credit burn breaks the budget, and the exact crossover points to M-series.
- Local NVMe vs EBS: why the 'd' suffix is AWS's best storage dealAdding local NVMe SSDs to EC2 instances costs pennies per hour while delivering hundreds of thousands of IOPS and sub-100µs latency. How instance storage economics compare against gp3 and io2, and the architectural patterns for ephemeral storage.
- The Windows license tax on EC2: when an OS doubles your compute billRunning Windows Server on EC2 adds a 70% to 103% price premium over identical Linux hardware, with zero Graviton support. The core-based licensing math, the cost of staying on legacy .NET, and the economics of modernizing to Linux containers.