Compute Engine Overview

Google Cloud Compute Engine is GCP's infrastructure-as-a-service offering, providing virtual machines that run on Google's global infrastructure — the same infrastructure that powers Google Search, YouTube, and Gmail. This infrastructure advantage translates to consistently low network latency and high throughput between GCP regions.

Compute Engine VMs are available in 40+ regions worldwide, with per-second billing, automatic sustained use discounts, and the unique ability to create custom machine types. Unlike AWS and Azure, where you must choose from predefined instance sizes, GCP lets you specify exact CPU and memory amounts.

GCP's pricing philosophy emphasizes simplicity and automatic savings. While AWS and Azure require you to actively purchase reservations for discounts, GCP's sustained use discounts apply automatically when you run instances for significant portions of a billing month.

Custom Machine Types — GCP's Unique Advantage

Custom machine types are arguably GCP's most compelling feature for cost optimization. Instead of choosing from predefined configurations (like AWS's m7i.xlarge with fixed 4 vCPU / 16 GB), you can specify exact vCPU and memory amounts:

• vCPUs: 1 to 96 (must be 1 or an even number)
• Memory: 0.9 GB to 6.5 GB per vCPU (standard), up to 8 GB per vCPU (extended memory)
• Granularity: Memory in 256 MB increments

Why this matters: Consider an application that needs 4 vCPUs and 12 GB of RAM. On AWS, you'd choose between m7i.xlarge (4 vCPU, 16 GB — overpaying for 4 GB extra RAM) or c7i.xlarge (4 vCPU, 8 GB — insufficient memory). On GCP, you simply create a custom machine with exactly 4 vCPU and 12 GB, paying only for what you use.

In practice, custom machine types save 15-30% compared to the nearest predefined alternative, especially for workloads with non-standard resource requirements. This makes GCP particularly attractive for containerized workloads where resource limits are precisely defined.

Extended memory: For applications needing higher memory-to-CPU ratios (like in-memory databases), extended custom machines allow up to 624 GB of memory per vCPU — far exceeding the standard 6.5 GB limit. Extended memory incurs a premium, but it's still cheaper than provisioning an oversized predefined instance.

Predefined Machine Families

E2 — Cost-Optimized General Purpose

E2 instances are GCP's most cost-effective option for general workloads. They use a dynamic resource manager that allocates CPU time efficiently, providing reliable performance for non-compute-intensive applications at the lowest price point.

E2 instances are available with shared-core configurations (e2-micro, e2-small, e2-medium) that cost as little as $6-7/month, making them comparable to AWS t3.micro but without the complexity of CPU credit management.

N2/N2D — Balanced General Purpose

N2 instances use Intel Cascade Lake or Ice Lake processors, while N2D uses AMD EPYC Rome/Milan. These are GCP's closest equivalents to AWS M-series instances, offering balanced CPU-to-memory ratios for production workloads.

N2D instances typically cost 10-13% less than N2 while delivering comparable performance for most workloads, making them the better default choice unless you specifically need Intel-specific features.

C2/C3 — Compute Optimized

C2 instances deliver the highest single-thread performance on GCP, using Intel Cascade Lake processors at sustained 3.8 GHz. C3 instances introduce Intel Sapphire Rapids with DDR5 memory and PCIe Gen5 for improved memory bandwidth and I/O throughput.

C3 instances also support hyperdisk (high-performance block storage) and offer up to 176 vCPUs per instance, significantly more than C2's maximum of 60.

M1/M2/M3 — Memory Optimized

GCP's memory-optimized instances offer massive RAM configurations: M1 up to 4 TB, M2 up to 12 TB, and M3 up to 4 TB with newer processors. These are designed for SAP HANA, large in-memory databases, and real-time analytics platforms.

A2/A3/G2 — Accelerator Optimized

A2 instances feature NVIDIA A100 GPUs, A3 instances use NVIDIA H100 GPUs, and G2 instances provide NVIDIA L4 GPUs for cost-effective inference. Google's TPU (Tensor Processing Unit) instances offer an alternative to GPU-based ML training with potentially better cost-efficiency for TensorFlow workloads.

Sustained Use Discounts — Automatic Savings

Sustained Use Discounts (SUDs) are one of GCP's most attractive features. They work automatically — no purchases, commitments, or reservations required. The discount increases as your instance runs for more of the billing month:

• 0-25% of the month: Full on-demand price
• 25-50% of the month: 20% discount on incremental usage
• 50-75% of the month: 40% discount on incremental usage
• 75-100% of the month: 60% discount on incremental usage

Net effect: An instance running 100% of the month automatically receives approximately 30% discount compared to straight on-demand pricing. This means even without any commitment, GCP's effective pricing is already 30% below its listed hourly rate for always-on workloads.

Important note: SUDs apply to N1, N2, N2D, C2, and custom machine types. E2 instances and newer families (C3, M3) use a different pricing structure and do not receive SUDs but are priced competitively to account for this.

Committed Use Discounts (CUDs)

For workloads requiring deeper savings beyond SUDs, Committed Use Discounts offer significant reductions:

• 1-year commitment: Up to 37% discount
• 3-year commitment: Up to 70% discount

GCP CUDs have a unique advantage: they're resource-based, not instance-based. You commit to a certain amount of vCPUs and memory in a region, and the discount applies flexibly across any qualifying instances. This means you can change instance types within the same family without losing your commitment discount.

Example: If you commit to 16 vCPUs and 64 GB RAM in us-central1, that commitment could cover one n2-standard-16, two n2-standard-8, or four n2-standard-4 instances — providing much more flexibility than AWS Reserved Instances which lock you to a specific instance type.

Preemptible and Spot VMs

GCP Spot VMs (the successor to Preemptible VMs) offer up to 60-91% discount for fault-tolerant workloads. Key differences from AWS Spot Instances:

• Pricing: GCP Spot prices are more stable than AWS (less market-driven fluctuation)
• Availability: GCP provides 30-second termination notice (vs AWS's 2 minutes)
• Duration: No maximum runtime limit (Preemptible VMs had a 24-hour limit; Spot VMs do not)
• Pricing model: GCP offers a dynamic or fixed discount model

Best for: Distributed data processing (Dataproc), CI/CD pipelines, rendering farms, fault-tolerant web applications behind load balancers, and development/testing environments.

Sole-Tenant Nodes

For workloads requiring physical isolation (compliance requirements, bring-your-own-license scenarios, or consistent performance guarantees), GCP offers sole-tenant nodes. You get dedicated physical servers that host only your VMs, with the flexibility to use any machine type on the node.

Sole-tenant nodes support Windows Server BYOL (bring your own license), which can provide significant cost savings for organizations with existing per-core Windows licenses — similar to Azure's Hybrid Benefit but available on GCP infrastructure.

GCP vs AWS vs Azure — Where GCP Wins