Compute on AWS: one set of building blocks, three ways to buy them
Compute is the category you'll come back to most often. AWS gives you raw virtual machines (EC2), a way to keep the right number of them running (Auto Scaling), a simple bundled server (Lightsail), job schedulers (Batch, Parallel Computing Service) and several specialised options. It's easier to think of these as one engine, several dashboards. Nearly everything here runs on EC2 underneath. The services differ in how much of the wheel you hold, and in how you pay for the hours.
Service names and how they map change over time. The tables below were checked in October 2026. Where a mapping is my own judgement rather than a documented equivalence, the notes say so.
Table 1: Every current AWS compute service
| AWS service | What it's for, in one line | Notes / former name |
|---|---|---|
| Amazon EC2 | Virtual machines where you pick the instance type, OS image and network | Instance families: general purpose, compute-, memory-, storage- and accelerator-optimised |
| Amazon EC2 Auto Scaling | Keeps a group of EC2 instances at the size you want and replaces unhealthy ones | Can mix on-demand and Spot capacity in one group |
| Amazon Lightsail | A bundled virtual server with storage, data transfer and a static IP for a flat monthly price | Made for small sites and simple apps |
| AWS Batch | Queues and runs batch jobs on EC2 or Fargate, and scales them for you | No cluster to manage |
| AWS Parallel Computing Service (PCS) | Managed Slurm clusters for HPC and engineering work | Can use EC2 Capacity Blocks for ML |
| AWS ParallelCluster | Open-source tool for building and running your own HPC clusters | You run the cluster yourself |
| EC2 Mac instances, Dedicated Hosts | Specialised hardware: macOS builds, and licences tied to a physical server | Billed differently from shared instances |
| EC2 Capacity Blocks for ML | Reserve GPU instances in advance for a fixed date range | A way to buy capacity, not a separate service |
Purchase options (how you pay for EC2):
- On-Demand: pay by the second or hour, with no commitment.
- Savings Plans: a 1- or 3-year commitment to spend a set amount per hour.
- Compute Savings Plans give up to 66% off and also cover Fargate and Lambda.
- EC2 Instance Savings Plans give up to 72% off but tie you to one instance family in one Region.
- Spot: spare capacity at a large discount. AWS can take it back after a two-minute interruption notice.
How to pick
Start from the shape of the workload, then choose how to pay:
- Long-running server you manage yourself → EC2, usually in an Auto Scaling group.
- Small, predictable site where a flat bill matters more than fine control → Lightsail.
- Jobs that queue, run and finish (rendering, nightly ETL) → Batch. Use PCS if your team already thinks in Slurm.
- How to pay: cover the steady baseline with Savings Plans, the unpredictable peaks with On-Demand, and fault-tolerant work that can restart with Spot.
Here is the kind of record I'd keep after a selection review:
# Compute selection record (worked example)
workload: nightly-video-transcode # queued, restartable jobs
service: AWS Batch # queue + scaling handled for us
capacity:
baseline: none # nothing runs between batches
burst: Spot # jobs checkpoint, so a 2-minute notice is fine
fallback: On-Demand # used when Spot capacity is short
commitment: Compute Savings Plan # covers the always-on API tier, not the batch jobs
azure_equivalent: Azure Batch (Spot nodes)
gcp_equivalent: Google Cloud Batch (Spot VMs, 120 s notice requested)
Table 2: The same jobs on AWS, Azure and Google Cloud (checked October 2026)
| Job | AWS | Azure | Google Cloud | Where the match breaks |
|---|---|---|---|---|
| Virtual machines | EC2 | Virtual Machines | Compute Engine | Families line up by purpose (general, compute, memory, storage, accelerator), but sizes and names don't map one to one. Compare vCPU, memory and price per hour, not names. |
| Keep a VM fleet at size | EC2 Auto Scaling | Virtual Machine Scale Sets | Managed instance groups | Scaling rules and health checks behave differently on each cloud. |
| Simple flat-priced server | Lightsail | No direct equivalent | No direct equivalent | Judgement: the nearest option is a small VM, but you're billed per resource instead of a flat bundle. |
| Batch jobs | AWS Batch | Azure Batch | Google Cloud Batch | Azure Batch retired low-priority nodes on 30 Sep 2025. Use Spot nodes. |
| Managed HPC / Slurm | PCS, ParallelCluster | Azure CycleCloud (Workspace for Slurm) | Cluster Toolkit | Judgement: only PCS is a managed service. The others deploy clusters that you then run yourself. |
| Spare-capacity discount | Spot Instances | Spot Virtual Machines | Spot VMs | AWS gives 2 min notice. Azure gives up to 30 s, best effort, through Scheduled Events. Google defaults to no notice with a best-effort shutdown of up to 30 s, but you can ask for a 120 s notice. |
| Commitment discount | Savings Plans (Compute / EC2 Instance) | Savings plan for compute; Reserved VM Instances | Committed use discounts | Google also applies sustained use discounts automatically once eligible VMs run for more than 25% of a month. AWS and Azure have nothing like this. |
| Reserved GPU window | EC2 Capacity Blocks for ML | No direct equivalent | No direct equivalent | Judgement: Azure and Google have other ways to reserve GPUs, but none that maps cleanly to a fixed date range. |
Gotcha: don't reuse Spot shutdown handling between clouds. A handler that relies on AWS's two minutes will lose work on Azure, where the notice can be 30 seconds or less.
Trade-off: EC2 Instance Savings Plans give the deeper discount, but they lock you to one family in one Region. If you might move workloads to another cloud, the more flexible Compute Savings Plan (or the equivalent on the other cloud) is usually safer.
Retiring or closed
- AWS SimSpace Weaver stopped taking new customers on 20 May 2025, and support ended on 20 May 2026.
- Amazon Linux 2 reached end of support on 30 June 2026. Check older EC2 images and AMIs, and move to Amazon Linux 2023.
- Azure Batch low-priority VMs were retired on 30 September 2025 and replaced by Spot.
Going deeper
For instance sizing, placement and cost modelling in depth, take Cloud Pro: AWS Compute.
Key takeaways
- EC2 is the engine underneath. Auto Scaling, Batch and PCS manage it for you, and Lightsail bundles it at a flat price.
- Pay for the steady baseline with Savings Plans, peaks with On-Demand, and restartable work with Spot.
- Spot notice differs by cloud: AWS 2 min; Azure up to 30 s; Google none by default, or 120 s if you ask for it.
- Google's sustained use discounts are automatic and have no match on AWS or Azure.
- Lightsail and Capacity Blocks have no direct equivalent on the other two clouds.
- SimSpace Weaver and Amazon Linux 2 have reached end of support. Check for both.