Why compute is the first category
In a multi-cloud review, compute is usually where you decide first and pay for longest. All three clouds sell virtual machines. The differences that cost you money are in the details: how you pick a machine shape, how a cheap interruptible VM is taken away from you, and how discounts are paid for. This lesson covers every current Google Cloud compute service, then maps each one to AWS and Azure, one row per job.
Think of Compute Engine as a car dealership with one showroom. The machine families are the models (saloon, van, sports car). The series (N4, C4, M4 and so on) are the model years. Spot VMs, sole-tenant nodes and the specialised hosts are the ways you can pay for or park the car.
Google Cloud compute services (checked October 2026)
| Service | What it is for | Former name / note |
|---|---|---|
| Compute Engine | Virtual machines, grouped into six machine families | — |
| General-purpose family (E2, N4, N4D, N4A, C4, C4D, C4A, N2, N2D, T2D, T2A) | Everyday web, app and database servers | E2 and the N-series accept custom machine types |
| Compute-optimized (H3, H4D, C2, C2D) | HPC and CPU-heavy work | — |
| Memory-optimized (M3, M4, X4, X5) | SAP HANA and large in-memory databases, up to 12 TB of memory and beyond on the X series | — |
| Storage-optimized (Z3, Z4M, Z4D) | High-IOPS local SSD work such as scale-out databases | — |
| Network-optimized (C4N, M4N) | Workloads that need very high network bandwidth | Newer family |
| Accelerator-optimized (A3, A4, A4X, G2, G4) | GPU training, inference and graphics | A4X cannot run as Spot |
| Spot VMs | Cheap, interruptible capacity, up to 91% off | Replaces the older preemptible VMs |
| Managed instance groups (MIGs) | Fleets of identical VMs with autoscaling and autohealing | Part of Compute Engine |
| Sole-tenant nodes | A physical server that runs only your project's VMs | Used for per-core licences (BYOL) and isolation |
| Bare metal instances | Compute Engine machines with no hypervisor | Some series only |
| Bare Metal Solution | Dedicated servers for specialised workloads such as Oracle | Managed hardware with local SAN |
| Google Cloud VMware Engine | A VMware private cloud on dedicated nodes | — |
| Batch | Queues and runs batch jobs on Compute Engine | No charge for Batch itself, you pay for the VMs |
Older series (N1, E2's predecessors, M1/M2, A2) still run, but the documentation labels them earlier generations. Choose from the 3rd and 4th generations for new work.
How to pick
Start with general-purpose. E2 is the budget choice for small or bursty servers. N4/N4D/N4A suit steady workloads and allow custom shapes. C4/C4D/C4A are for performance-sensitive production. Move to a specialised family only when one resource clearly limits you: CPU (H-series), memory (M/X), local disk (Z) or GPUs (A/G). Use Spot for anything that can restart, use MIGs for anything you need more than one of, and use sole-tenant nodes only when a licence or a compliance rule requires them. VMware Engine and Bare Metal Solution exist so you can lift a workload across without rewriting it. They are not a default.
Gotcha: custom shapes. On E2 and the N-series you can ask for, say, 6 vCPU and 20 GB of memory. AWS and Azure sell fixed sizes only, so the same workload maps to the next size up on those clouds. Compare costs using the shape you will actually buy, not the shape you would ideally want.
Equivalence table (dated October 2026)
| Job | Google Cloud | AWS | Azure | Where the match breaks |
|---|---|---|---|---|
| General-purpose VM | Compute Engine E2/N4/C4 | EC2 M and T families | Virtual Machines D and B series | Only Google offers custom shapes |
| Arm VM | C4A, N4A (Axion), T2A | EC2 Graviton | Cobalt 100 VMs | Per-core prices differ, so benchmark before you assume parity |
| Compute-heavy / HPC | H3, H4D, C2D | EC2 C and Hpc families | F and HB/HC series | Close match |
| Large memory | M3, M4, X4 | EC2 R, X, U families | E and M series | Upper memory limits differ; check against SAP certification lists |
| Storage-optimized | Z3, Z4M, Z4D | EC2 I family | L series | Close match |
| GPU | A3, A4, G2, G4 | EC2 P and G families | ND, NC, NV series | GPU models and regional stock differ most here |
| Interruptible VM | Spot VMs | EC2 Spot Instances | Azure Spot VMs | Notice period differs (see below) |
| Autoscaled fleet | Managed instance groups | EC2 Auto Scaling groups | Virtual Machine Scale Sets | Close match |
| Dedicated host | Sole-tenant nodes | EC2 Dedicated Hosts | Azure Dedicated Host | Licensing rules are the vendor's, not the cloud's |
| Bare metal | Bare metal instances; Bare Metal Solution | EC2 .metal instances | BareMetal Infrastructure | Judgement: Bare Metal Solution is closer to a managed hosting contract |
| VMware in the cloud | Google Cloud VMware Engine | Amazon Elastic VMware Service (GA August 2025) | Azure VMware Solution | Node types and licence portability differ |
| Batch scheduling | Batch | AWS Batch | Azure Batch | Close match; all three charge only for the underlying compute |
| Commitment discount | Committed use discounts (CUDs) | Reserved Instances, Savings Plans | Reservations, savings plan for compute | See below |
| Automatic discount | Sustained use discounts | No direct equivalent | No direct equivalent | Applies to older series only |
The two rows that cause the most trouble
Preemption. Compare how much warning each cloud gives before it takes a VM back:
# How much warning before the cloud reclaims an interruptible VM
google_spot:
preemption_notice: 0s # the default; you can set 120s
shutdown_window: up to 30s # best effort, after the soft-off signal
on_preempt: STOP | DELETE # STOP is the default
sla: none # outside the Compute Engine SLA
aws_spot:
notice: 120s # sent via instance metadata and EventBridge
azure_spot:
notice: 30s # best effort, via Scheduled Events
If your shutdown script was written for AWS's two minutes, set Google's 120-second preemption notice. Otherwise it gets 30 seconds at most.
Discounts. Google has two kinds of CUD, each for 1 or 3 years. Resource-based CUDs give up to 55% off, or up to 70% on memory-optimized series, and are tied to a region and project. They cannot be cancelled. Compute flexible CUDs give 17–63% off and apply across the whole billing account. These correspond roughly to AWS's Reserved Instances and Savings Plans and to Azure's reservations and savings plan. Sustained use discounts are automatic, but they apply only to N1, N2, N2D, C2, M1, M2 and sole-tenant nodes. They do not combine with CUDs. Don't build a cost model for a new C4 or N4 fleet that counts on them.
Retiring and legacy
No Google compute service in this category is announced for retirement as of October 2026. Preemptible VMs are the legacy form of Spot VMs, and 1st-generation series such as N1 are earlier-generation hardware. Avoid both for new designs.
Going deeper
For sizing, MIG design and commitment planning, continue with Cloud Pro: Google Cloud Compute.
Key takeaways
- Compute Engine has six machine families. Start with general-purpose and move to a specialised family only when one resource limits you.
- Custom machine types (E2 and N-series) have no direct equivalent on AWS or Azure.
- Spot preemption notice: Google defaults to 0s (120s if you set it) plus up to 30s to shut down; AWS gives 2 minutes; Azure gives 30 seconds.
- Sole-tenant nodes map to EC2 Dedicated Hosts and Azure Dedicated Host. MIGs map to Auto Scaling groups and Scale Sets.
- Resource-based CUDs cannot be cancelled. Sustained use discounts cover only older series and have no AWS or Azure equivalent.