Serverless Kubernetes costs for EKS, AKS, GKE, and OKE

This post is an update to the original post from March 2023, “Kubernetes cost comparison: Who provides the best value?” We're revisiting the topic because, while much has changed with capabilities and customers’ use case priorities, such as AI model training and inferencing being top of mind, cloud providers’ pricing for serverless Kubernetes services hasn't budged in the last fifteen months.

Whether for x86- or Arm-based Linux images, Oracle Cloud Infrastructure (OCI) Kubernetes Engine (OKE) still offers the lowest cost for most enterprise use cases. Even when compared to competitors’ lowest cost regions, OKE is about one third the cost of comparable offerings from Amazon Elastic Kubernetes Service (EKS) with Amazon Web Services (AWS) Fargate, Azure Kubernetes Service (AKS) with virtual nodes, and Google Kubernetes Engine (GKE) Autopilot for a serverless experience.

Kubernetes services from AWS, Azure, Google Cloud, and OCI

For the uninitiated, Kubernetes is an open source container orchestration solution for automating the deployment, scaling, and management of containerized software. Gartner estimates that by 2029, more than 95% of global organizations will be running containerized applications in production, which is a significant increase from less than 50% in 2023.

AWS, Azure, Google Cloud, and OCI all offer managed Kubernetes services, each aiming to help reduce your time, cost, and the operational burden of managing Kubernetes infrastructure. The services have no minimum charges, no upfront commitments required, and you pay only for the underlying resources and applicable management fees.

All four services manage the availability and scalability of Kubernetes control plane nodes and offer varying levels of choice and control for worker nodes. Customers often opt for serverless worker nodes to focus on higher-value priorities, which is a macro trend for cloud services in general. In this post, we compared the following Kubernetes serverless options for each service:

• Amazon Elastic Kubernetes Service (EKS) with AWS Fargate
• Azure Kubernetes Service (AKS) virtual nodes (using Azure Container Instances)
• Google Kubernetes Engine (GKE) Autopilot
• OCI Kubernetes Engine (OKE) virtual nodes

As Kubernetes usage grows and budgets receive more scrutiny, customers look for ways to reduce costs. Given that Kubernetes is an open source framework, the managed offerings are largely similar across providers, yet what you pay for these services varies tremendously. Let’s look at how these services are priced.

Comparing prices for serverless Kubernetes services

The following table contains serverless Kubernetes service pricing for production workloads using Linux images and current Kubernetes versions for both x86 and Arm processors. We’re using the lowest published regional rates for AWS, Azure, and Google to show them at their best. In other words, while OCI prices are consistent regardless of region, the others’ pricing varies quite widely, but we're comparing against their lowest. Observe the following substantial price differences:

• x86 and Linux: OKE is 70% lower for CPU and 65% lower for memory
• Arm and Linux: OKE is 70% lower for CPU and 58% lower for memory

These published rates have the following details:

• All calculations use prices published as of August 1, 2024. We have attempted to use the closest comparisons in terms of Kubernetes version and processor configurations, such as manufacturer, type, generation, speed, memory, bandwidth, and so on.
• Common cloud industry practice is to define compute instances based on the number of virtual CPUs they include. Each vCPU provides the capacity for one thread of runs. A vCPU doesn't provide a whole physical compute core but is instead part of a core. In contrast, Oracle’s x86 Compute shapes use OCPUs, which equate to physical CPU cores, each of which provides for two threads. So, two x86 vCPUs on AWS, Azure, or Google Cloud are comparable to one OCPU on OCI. Arm vCPUs and OCPUs are comparable across all providers: One Arm vCPU equals one Arm OCPU.
• Prices reflect each service’s recommended options for production workloads, such as Standard tier for AKS, Standard edition for GKE, and three virtual nodes for OKE. For simplicity, we also assume the use of general purpose processors with the included ephemeral storage allocation.
• GKE offers a free tier providing $74.40 in monthly credits. We apply this credit in the scenario comparisons.
• AKS virtual nodes aren't available with Arm processors.
• GKE Autopilot currently supports Arm processors in only three of their 34 global regions (Iowa, Netherlands, and Singapore).

AWS, Azure, and Google Cloud prices vary widely by region

Prices for Amazon EKS with AWS Fargate vary widely across AWS regions and can be much higher than in the US East region in northern Virginia—72% higher in São Paulo. Likewise, CPU prices for AKS virtual nodes in many other Microsoft Azure regions are substantially higher than in the East US region—a whopping 100% higher in Brazil South. Similarly, CPU prices for GKE Autopilot in other Google Cloud regions can be much higher than their lowest rates in the Iowa region—59% higher in São Paulo, for instance.

In contrast, OCI offers consistent pricing and availability for all services across all global regions, including OKE virtual nodes. This predictability makes it easy for you to plan and budget for rapid geographic expansion regardless of where you need to consume the cloud services.

Calculating Kubernetes costs

The total monthly price for a Kubernetes service is the sum of fees for cluster, node, CPU, and memory, calculated with the following formula:

• Cluster: (# clusters) * (price-per-hour) * (# daily-cluster-hours) * (# days)
• Node: (# nodes) * (price-per-hour) * (# daily-node-hours) * (# days)
• CPU: (# nodes) * (# vCPU) * (price-per-hour) * (# daily-CPU-hours) * (# days)
• Memory: (# nodes) * (# GiB) * (price-per-hour) * (# daily-memory-hours) * (# days)

To compare the true cost of real-world use cases, we use two moderately sized serverless scenarios that represent common workloads: One using x86 and Linux images and one using Arm and Linux images.

Scenario A: Serverless Kubernetes cluster for x86 and Linux

In this scenario, we have a single serverless cluster of 20 pods running constantly for 31 days, each configured with 16 vCPU and 64 GiB memory using x86 and Linux images. We show all the calculations and then populate the comparison table with rounded totals.

EKS with Fargate

• Cluster: (1 cluster) * ($0.10/cluster-hour) * (24 hours) * (31 days) = $74.40
• Node: No charge
• CPU: (20 pods) * (1 node/pod) * (16 vCPU/node) * ($0.04048/vCPU-hour) * (24 hours) * (31 days) = $9,637.48
• Memory: (20 pods) * (1 node/pod) * (64 GiB/node) * ($0.004445/GiB-hour) * (24 hours) * (31 days) = $4,233.06
• Total: $74.40 + $0 + $9,637.48 + $4,233.06 = $13,944.94 for EKS with Fargate

AKS virtual nodes

• Cluster: (1 cluster) * ($0.10/cluster-hour) * (24 hours) * (31 days) = $74.40
• Node: No charge
• CPU: (20 pods) * (1 node/pod) * (16 vCPU/node) * ($0.0405/vCPU-hour) * (24 hours) * (31 days) = $9,642.24
• Memory: (20 pods) * (1 node/pod) * (64 GiB/node) * ($0.00445/GiB-hour) * (24 hours) * (31 days) = $4,237.82
• Total: $74.40 + $0 + $9,642.24 + $4,237.82 = $13,954.46 for AKS Virtual Nodes

GKE Autopilot

• Cluster: No charge because the GKE free tier provides one free cluster per month
• Node: No charge
• CPU: (20 pods) * (1 node/pod) * (16 vCPU/node) * ($0.0445/vCPU-hour) * (24 hours) * (31 days) = $10,594.56
• Memory: (20 pods) * (1 node/pod) * (64 GiB/node) * ($0.0049225/GiB-hour) * (24 hours) * (31 days) = $4,687.80
• Total: $74.40 + $0 + $10,594.56 = $15,282.36 for GKE Autopilot

OKE virtual nodes

• Cluster: (1 cluster) * ($0.10/cluster-hour) * (24 hours) * (31 days) = $74.40
• Node: (3 virtual nodes) * ($0.015/virtual-node-hour) * (24 hours) * (31 days) = $33.48
• CPU: (20 pods) * (1 node/pod) * (16 vCPU/node) * ($0.0125/vCPU-hour) * (24 hours) * (31 days) = $2,976.00
• Memory: (20 pods) * (1 node/pod) * (64 GiB/node) * ($0.0015/GiB-hour) * (24 hours) * (31 days) = $1,428.48
• Total: $74.40 + $33.48 + $2,976.00 + $1,428.48 = $4,512.36 for OKE Virtual Nodes

For Scenario A, EKS with Fargate, AKS virtual nodes, and GKE Autopilot are each nearly three times as expensive as OKE virtual nodes, even using prices from the lowest cost region for all services.

Scenario B: Serverless Kubernetes cluster for Arm and Linux

The only difference in this scenario from the previous one is that we're employing Arm and Linux images instead of x86 and Linux images.

EKS with Fargate

• Cluster: (1 cluster) * ($0.10/cluster-hour) * (24 hours) * (31 days) = $74.40
• Node: No charge
• CPU: (20 pods) * (1 node/pod) * (16 vCPU/node) * ($0.03238/vCPU-hour) * (24 hours) * (31 days) = $7,809.02
• Memory: (20 pods) * (1 node/pod) * (64 GiB/node) * ($0.00356/GiB-hour) * (24 hours) * (31 days) = $3,390.26
• Total: $74.40 + $7,809.02 + $3,390.26 = $11,273.68 for EKS with Fargate

AKS virtual nodes: AKS virtual nodes don't support Arm and Linux configurations.

GKE Autopilot

• Cluster: N charge because the GKE free tier provides one free cluster per month
• Node: N charge
• CPU: (20 pods) * (1 node/pod) * (16 vCPU/node) * ($0.0356/vCPU-hour) * (24 hours) * (31 days) = $8,475.65
• Memory: (20 pods) * (1 node/pod) * (64 GiB/node) * ($0.003938/GiB-hour) * (24 hours) * (31 days) = $3,750.24
• Total: $0 + $0 + $8.475.65 + $3,750.24 = $12,225.88 for GKE Autopilot

OKE virtual nodes

• Cluster: (1 cluster) * ($0.10/cluster-hour) * (24 hours) * (31 days) = $74.40
• Node: (3 virtual nodes) * ($0.015/virtual-node-hour) * (24 hours) * (31 days) = $33.48
• CPU: (20 pods) * (1 node/pod) * (16 vCPU/node) * ($0.01/vCPU-hour) * (24 hours) * (31 days) = $2,380.80
• Memory: (20 pods) * (1 node/pod) * (64 GiB/node) * ($0.0015/GiB-hour) * (24 hours) * (31 days) = $1,428.48
• Total: $74.40 + $33.48 + $2,380.80 + $1,428.48 = $3,917.16 for OKE Virtual Nodes

AKS virtual nodes can't support the Arm-based use case, and both EKS with Fargate and GKE Autopilot are about three times as expensive as OKE virtual nodes for Scenario B.

Conclusion: OKE offers far lower costs for serverless Kubernetes

The evidence is clear: Regardless of x86 or Arm processor type, even in competitors’ lowest-cost regions, OKE offers far lower costs than EKS, AKS, and GKE for serverless Linux scenarios. As itemized, AWS, Azure, and Google Cloud are an astounding 2.9–3.4 times higher than OCI. Put differently, OKE is about one third the cost of alternatives for serverless Kubernetes.

The cost differences are even greater in many other public cloud regions globally where AWS, Azure, and Google Cloud have much higher CPU prices than their lowest rates, up to 100% more in some cases. These financial differences can really add up, particularly when workloads scale. Think about what your organization could do with savings of many hundreds of thousands of dollars per month: Hire or retain staff, fund new projects, bolster the bottom line, or a blend of them all.

Learn more about OKE and try the service for yourself, read more about the latest set of OKE capabilities to simplify large-scale Kubernetes environments at reduced costs, and use the OCI Cost Estimator to price out your own workloads. You can also contact us to speak with a cloud architect to discover how OKE can help you accelerate and simplify your adoption of Kubernetes and scale its usage across your organization in a cost-effective way. To hear from and meet with OKE subject matter experts, attend Oracle CloudWorld in Las Vegas or virutally from September 9–12.

GARTNER is a registered trademark and service mark of Gartner, Inc. and/or its affiliates in the U.S. and internationally and is used herein with permission. All rights reserved.