Virtual Machines

Last reviewed: September 2026

Overview

To run servers in your own datacenter, you have to purchase hardware, install it, configure the OS, and set up networking yourself. Equipment procurement takes weeks, and changing specifications requires physical replacement.

A virtual machine (VM) replaces this process with software. You can create a server with your desired specifications in minutes and delete it immediately when no longer needed. It is the most fundamental cloud service that solved the flexibility problem of physical servers.

However, with VMs users must still manage OS patching, security configuration, and failure recovery themselves. To reduce this management burden, higher-level abstractions like containers and serverless emerged.

Product comparison

General-purpose VM

Vendor	Product	Notes
AWS	EC2 (Elastic Compute Cloud)	400+ instance types. Categorized into general purpose, compute, memory, GPU, etc.
Azure	Virtual Machines	Strength in Windows workloads (Hybrid Benefit)
Google Cloud	Compute Engine	Freely combine CPU/memory with Custom Machine Types
OCI	OCI Compute	Flexible Shape allows freely combining CPU/memory. Ampere A1 free tier

Arm-based VM

Vendor	Product	Notes
AWS	EC2 Graviton (latest m9g/c9g/r9g, m8g, t4g, etc.)	In-house designed Arm processors (latest Graviton5; earlier Graviton4/3/2 still offered). Better price-performance than x86
Azure	Cobalt 100 (Dpsv6, etc.) / Dpsv5 series	Based on in-house Cobalt 100 and Ampere Altra
Google Cloud	Axion (C4A) / Tau T2A	In-house designed Axion and Arm-based general purpose
OCI	Ampere A1 Compute	Based on Ampere Altra. Free tier available

GPU / AI accelerators

Vendor	Product	Notes
AWS	P5, G7, G5, Inf2, Trn1/Trn2	NVIDIA (H100, RTX PRO 4500 Blackwell) + in-house chips (Inferentia, Trainium). G7 is the first major cloud offering of RTX PRO Blackwell
Azure	NC, ND, NV series	NVIDIA A100, H100
Google Cloud	A3, A2, G2 series + TPU (v5p/v6e/Ironwood)	NVIDIA H100 + in-house TPU (latest 7th-gen Ironwood/TPU7x)
OCI	GPU Instances (A10, A100, H100)	NVIDIA GPU. Bare Metal option available

Images (OS templates)

Vendor	Product	Notes
AWS	AMI (Amazon Machine Image)	3rd-party images available on Marketplace
Azure	VM Image / Shared Image Gallery
Google Cloud	Image / Image Family
OCI	Custom Image / Platform Image	3rd-party images available on Marketplace

Key differences

AWS — Has the widest variety of instance types, and its in-house Arm processor (Graviton) enables cost savings.
Azure — Hybrid Benefit lets you leverage existing Windows licenses to reduce cost.
Google Cloud — Offers Custom Machine Type for freely combining CPU and memory.
OCI — Flexible Shape lets you freely combine CPU/memory, and offers an Ampere A1 free tier.

What to choose when

In this situation	Choose this
Windows workload + leveraging an existing license	Azure VM (Hybrid Benefit)
You need the widest selection of instance types	AWS EC2
You want to freely combine CPU/memory in 1-vCPU increments	Google Cloud Compute Engine (Custom Machine Type) or OCI (Flexible Shape)
Your goal is Arm-based cost savings	AWS EC2 Graviton
You want to use an Arm VM on the free tier	OCI Ampere A1
You need GPU/AI accelerators + in-house chips (Inferentia, Trainium)	AWS EC2 (P5, Inf2, Trn1)
You need a bare metal server	OCI Bare Metal or AWS Bare Metal

Purchasing options (pricing models)

For the same VM, price can vary greatly depending on the commitment model.

Option	Description	Discount rate	Risk
On-Demand (Pay-As-You-Go)	Billed by the second/hour based on usage	0% (baseline)	None
Reserved	1- or 3-year commitment	Up to 72%	Charged even if unused
Savings Plans / Savings Plan / CUD	Commit to a spending amount per hour	Up to 72%	Flexible, but requires a usage commitment
Spot / Preemptible	Use spare capacity at a low price	Up to 90%	Can be reclaimed at any time
Reserved Capacity	Reserve capacity in a specific AZ/zone	—	Guarantees capacity availability

Purchasing options comparison

Vendor	On-demand	Long-term commitment	Spot
AWS	On-Demand	Reserved Instance + Savings Plans	Spot Instance
Azure	Pay-As-You-Go	Reserved VM Instance + Savings Plan	Spot VM
Google Cloud	On-Demand	Committed Use Discount (CUD)	Preemptible / Spot VM
OCI	Pay-As-You-Go	Monthly Flex / Annual Flex / Universal Credits	Preemptible Instance

The figures above are as of the time this document was written and are subject to change. Check each vendor’s official pricing page for the latest prices.

Placement groups and dedicated hosts

You can physically control VM placement to meet high-performance/compliance requirements.

Feature	Description	AWS	Azure	Google Cloud	OCI
Same-rack placement	Low-latency communication (HPC, distributed DBs)	Placement Group (Cluster)	Proximity Placement Group	Compact Placement	Cluster Network
Spread placement	Eliminate a single point of failure	Placement Group (Spread)	Availability Set	Spread Placement	Fault Domain
Dedicated physical server	Licensing/compliance	Dedicated Host	Dedicated Host	Sole-tenant Node	Dedicated VM Host

Common mistakes

Running everything on-demand and then complaining about cost — For stable, predictable workloads, Reserved Instances or Savings Plans can save up to 72%. Analyze your usage pattern.
Choosing an instance type once and never revisiting it — Vendors continuously release new generations. You can get higher performance for the same cost, so review regularly.
Manual configuration every time, without a golden image — Repeating package installs and configuration every time you create an instance breaks consistency and wastes time. Manage AMIs/images through a build pipeline.

Checklist

Have you chosen an instance family that fits the workload’s characteristics (CPU/memory/GPU)?
Have you applied Reserved Instances or Savings Plans to stable, steady-state workloads?
Is OS patching and security update automation (SSM, Update Management, etc.) configured?