AKS Rightsizing Tools: Azure Kubernetes Service Rightsizing Software and AKS Cost Optimization Tools Compared
Rightsizing on Azure Kubernetes Service means two different jobs, and most tool lists blur them. The first is pod rightsizing: setting CPU and memory requests close to what each container really uses, so the scheduler stops reserving capacity nobody touches. The second is node rightsizing: picking VM sizes and counts that fit the requests you end up with, because Azure bills you for the virtual machines, not for the requests. AKS now ships a free answer to both, the Vertical Pod Autoscaler add-on and node auto-provisioning, and both come with limits Microsoft documents but vendors rarely repeat. This page lists nine tools that rightsize AKS, what each one changes, the AKS-specific catch for each, and what they cost where a price is published, so you can tell whether you need to buy anything at all.
projected this month if unattended
Spend by team
Budget forecast
The short answer
The best AKS rightsizing tool depends on which layer is wasting money. If pod requests are inflated and your workloads are Linux, stateless and under about 1,000 pods, the free AKS Vertical Pod Autoscaler in InPlaceOrRecreate mode (AKS 1.34 and later) plus node auto-provisioning covers most of it at no license cost. Buy a commercial tool such as CAST AI, ScaleOps, StormForge or PerfectScale when you run Java services, Windows nodes, more than 1,000 tracked pods, or need VPA and HPA on the same workload, which the native stack does not handle.
CostAnalyst is on this list, so here is the plain version first. We do not install a controller in your AKS clusters and we do not change pod requests or node pools. We read your Azure bill and Kubernetes cost data read-only, allocate AKS spend per team and product, and show what the rightsizing tool you pick actually saved on the invoice. We publish our prices. Everything below about other vendors comes from their own documentation or AWS Marketplace listings, cited where it matters.
How we compared
Five things that actually separate these tools
Which layer is wasting the money: pod requests or nodes
Start by splitting your AKS bill into the two problems, because they need different tools and the cheapest fix is often the one you already have. Inflated pod requests show up as nodes that are full on paper and idle in practice: the scheduler cannot place new pods because every CPU is requested, yet average utilization sits at 20 or 30 percent. That is a pod rightsizing problem, and the tools that fix it change requests (VPA, ScaleOps, StormForge, PerfectScale, CAST AI's workload autoscaler). Badly sized nodes show up differently: requests are reasonable, but node pools are built on one VM size chosen a year ago, so pods fragment across machines and a large share of each node is left over. That is a node problem, and it is fixed by node auto-provisioning, which runs Karpenter on AKS, or by CAST AI, which replaces the node layer entirely. The trap Microsoft's own cost analysis view makes visible is "idle charges", defined as the cost of available capacity no workload is using. If shrinking requests does not reduce idle charges, nothing consolidated the pods onto fewer VMs, and the invoice stays flat. You need both layers handled, either by the native pair or by one vendor that does both.
What the free AKS stack actually does, and where it stops
The native answer is better than it was a year ago. The AKS Vertical Pod Autoscaler add-on is based on the upstream Kubernetes VPA and sets container requests from past usage. Since AKS 1.34 it supports an InPlaceOrRecreate mode that changes requests on a running pod without restarting it when it can, and falls back to evicting the pod when it cannot. Microsoft also states that the old Auto mode is deprecated since VPA 1.4.0 and is now an alias for Recreate. The limits are specific and written down: optimal support for up to 1,000 pods per cluster tracked by VPA objects, the recommender stores only eight days of history, JVM-based workloads are not supported because VPA cannot see real memory use, Windows containers are not supported, and Microsoft does not recommend VPA together with an HPA that scales on the same CPU and memory metrics. Node auto-provisioning deploys and manages Karpenter for you, picking VM sizes from pending pod requirements, and it is preconfigured on AKS Automatic. Its documented limits: no Windows node pools, no IPv6 clusters, no service principals, and a cluster with NAP enabled cannot be stopped. If none of those limits applies to your estate, you may not need a paid rightsizing product.
What seeing the waste costs on Azure
Before any tool rightsizes anything, somebody has to show which namespaces are oversized, and on AKS that has a cost of its own. The AKS cost analysis add-on is built on OpenCost and reconciles usage with your Azure invoice, but Microsoft's prerequisites say the cluster must be on the Standard or Premium tier, not Free. Azure's retail price list charges the Standard tier at $0.10 per cluster per hour, which is $876 a year per cluster, and the Premium tier with long-term support at $0.60 per cluster per hour. Two more conditions rule some buyers out: the Kubernetes cost views are only available on Enterprise Agreement and Microsoft Customer Agreement offer types, and the add-on supports roughly 7,000 containers per cluster under its current 4 GB memory limit, with memory use of about 200 MB plus 0.5 MB per container. Production clusters are usually on Standard already for the uptime SLA, so for many teams the view is effectively free. Teams on pay-as-you-go or buying through a CSP need Kubecost, OpenCost or a bill-level tool instead.
What the commercial tools cost, where they publish it
Most AKS rightsizing vendors show a quote form on their own site and a full price on AWS Marketplace, which applies to their product in general rather than to one cloud. CAST AI, which supports AKS alongside EKS, GKE and Oracle Cloud, lists Growth at $1,000 a month for up to 500 CPUs and four clusters, plus $0.00694444 per managed CPU per hour, which is $5.00 per CPU for a 720-hour month, and Enterprise at $5,000 a month. ScaleOps lists a platform fee of $50,000 a year plus $9.00 per vCPU, with the metering period not stated on the listing. StormForge Optimize Live, now owned by CloudBolt, lists $2,700 a month per 1,000 vCPUs on contract and $0.0041 per requested vCPU-hour pay-as-you-go with no minimum, about $2.99 per vCPU a month. PerfectScale by DoiT publishes a free tier up to 200 vCPUs a month with unlimited clusters and quotes the paid tiers. Kubecost is free to 250 cores and $15,000 a year for Enterprise on its public offer. Ask each vendor whether an Azure Marketplace private offer is available, because buying through Azure Marketplace can let the spend count toward an existing Azure consumption commitment if the offer is eligible.
What access the tool needs inside your AKS cluster
Security review is where AKS rightsizing projects stall, so ask for the access model in writing before the demo. Recommendation-only tools need read access and an in-cluster agent or Prometheus; the AKS cost analysis add-on, for example, deploys an agent and creates a managed identity with read access to the node resource group. Automation tools need write access: VPA and pod-level products mutate pod specs through an admission webhook and patch running pods, and node-level products such as CAST AI and NAP create and delete VMs in your subscription. CAST AI starts with a read-only agent that produces a savings estimate before you grant anything else, and Sedai describes connecting to each AKS cluster through Azure Active Directory authentication with a copilot mode that waits for approval. If your security team will not approve a controller with write access this quarter, the realistic shortlist is VPA in Off mode, the cost analysis add-on, Kubecost or OpenCost, and a bill-level view, and the automation decision waits for the next review.
At a glance
9 AKS rightsizing tools compared
← Scroll to see all columns →
| Tool | Best for | What it rightsizes | The AKS catch | Pricing |
|---|---|---|---|---|
| AKS Vertical Pod Autoscaler add-on | Linux, stateless AKS workloads under about 1,000 tracked pods | Pod CPU and memory requests; in place on AKS 1.34 and later | No JVM, no Windows containers, 8 days of history, not with a CPU or memory HPA | Free; included in AKS Automatic and AKS Standard |
| AKS node auto-provisioning (Karpenter) | Clusters whose node pools were sized once and never revisited | VM sizes and node counts from pending pod requests | No Windows node pools, no IPv6, no service principals, cluster cannot be stopped | Free feature; you pay for the VMs. Preconfigured on AKS Automatic |
| AKS cost analysis add-on | Seeing idle and unallocated AKS spend per namespace inside the Azure portal | Nothing; it measures and shows idle charges | Standard or Premium tier only; EA and MCA offers only; about 7,000 containers per cluster | No add-on fee; needs the Standard tier at $0.10 per cluster-hour ($876 a year) |
| CAST AI | AKS estates that want nodes, spot VMs and pod requests managed by one controller | Nodes, spot capacity, bin packing and pod requests | Takes over node provisioning, so it replaces NAP or the cluster autoscaler | Free monitoring plan; Growth $1,000 per month plus $5.00 per CPU per month on AWS Marketplace; Enterprise $5,000 per month |
| ScaleOps | Platform teams that want continuous pod rightsizing without touching node tooling | Pod requests, replicas and pod placement | Metering period for the $9.00 per vCPU charge is not stated on the listing | $50,000 per year platform fee plus $9.00 per vCPU on AWS Marketplace |
| StormForge Optimize Live (CloudBolt) | Teams that want ML-based pod rightsizing priced per vCPU with no minimum | Pod requests and limits, including Java heap settings | Priced on requested vCPUs, including HPA replicas | $0.0041 per requested vCPU-hour pay-as-you-go, or $32,400 per year per 1,000 vCPUs on AWS Marketplace; 30-day trial |
| PerfectScale by DoiT | Smaller AKS estates that want automated rightsizing with a free entry tier | Pod requests, limits and node utilization | Paid tiers are quote only | Free up to 200 vCPUs per month with unlimited clusters; Advanced and Expert quoted per vCPU |
| Kubecost (IBM) | Chargeback and showback per namespace across AKS and other clusters | Nothing; recommends requests and flags idle capacity | Recommends only; an engineer applies the change | Free to 250 cores; Enterprise $15,000 per year to 250 cores on AWS Marketplace |
| CostAnalyst | Finance and FinOps teams proving what AKS rightsizing saved on the Azure invoice | Nothing in the cluster; allocates Azure and AKS spend per team | Kubernetes allocation is on the Scale plan | $99, $299 and $799 per month, published; Kubernetes on the Scale plan |
Product facts checked July 2026. Vendors change pricing and packaging often, so confirm before you buy.
Tool by tool
What each tool is genuinely best at
AKS Vertical Pod Autoscaler add-on
Best for: Linux, stateless AKS workloads under about 1,000 tracked pods
The VPA add-on is the first thing to try and, for a lot of clusters, the last. It recommends and applies CPU and memory requests from observed usage, and the InPlaceOrRecreate mode on AKS 1.34 and later resizes running pods without a restart where the kubelet allows it. Microsoft recommends starting in Off mode to read recommendations before letting it act. Its documented limits are the reason commercial tools exist: eight days of history misses monthly batch peaks, Java services are unsupported, and it does not coexist with an HPA on the same metrics.
AKS Vertical Pod Autoscaler add-on compared to CostAnalystAKS node auto-provisioning (Karpenter)
Best for: Clusters whose node pools were sized once and never revisited
Node auto-provisioning is Microsoft's managed Karpenter, built on the open source Karpenter and the AKS Karpenter provider. It picks VM configurations from what pending pods ask for instead of from a fixed node pool, which is the other half of rightsizing: smaller requests only save money when something packs pods onto fewer VMs. AKS Automatic turns it on by default and adds a pod readiness SLA. It does not touch pod requests, so pair it with VPA or a pod-level vendor.
AKS node auto-provisioning (Karpenter) compared to CostAnalystAKS cost analysis add-on
Best for: Seeing idle and unallocated AKS spend per namespace inside the Azure portal
The cost analysis add-on is OpenCost, managed by Microsoft and reconciled against your Azure invoice. It splits cluster cost into namespaces and assets and names four charge types, including idle charges for capacity no workload uses, which is the number a rightsizing project should move. It changes nothing in the cluster. Data can take up to 24 hours to appear after you enable it, and the offer-type restriction rules out pay-as-you-go and CSP subscriptions.
AKS cost analysis add-on compared to CostAnalystCAST AI
Best for: AKS estates that want nodes, spot VMs and pod requests managed by one controller
CAST AI is the broadest option on this list for AKS: it supports AKS alongside EKS, GKE and Oracle Cloud, and it rightsizes both layers, replacing the node autoscaler, moving eligible workloads to spot capacity with fallback to on-demand, and resizing pod requests. It starts with a read-only agent that estimates savings before you grant write access. Its own pricing page is a quote form; the AWS Marketplace listing is where the numbers are, and a 500-CPU cluster on Growth lists at about $42,000 a year.
CAST AI compared to CostAnalystScaleOps
Best for: Platform teams that want continuous pod rightsizing without touching node tooling
ScaleOps publishes an AKS cost optimization guide and runs on AKS, EKS and GKE, self-hosted, as ScaleOps Cloud or air-gapped. It focuses on the pod layer: continuous request rightsizing, replica tuning and placement, starting from read-only recommendations. The public price is steep at the bottom end, $50,000 a year before usage, so it fits estates where the pod-level waste is clearly larger than that. Get the metering period of the per-vCPU charge in writing, because it moves the bill by more than the platform fee.
ScaleOps compared to CostAnalystStormForge Optimize Live (CloudBolt)
Best for: Teams that want ML-based pod rightsizing priced per vCPU with no minimum
StormForge rightsizes pod requests and limits with machine learning over weeks of usage and can resize pods in place. It is cluster-agnostic and the vendor has published a walkthrough for running it alongside Karpenter on AKS, which maps directly to AKS node auto-provisioning. It is the cheapest paid automation on this list at small scale because the pay-as-you-go meter has no minimum, and it handles Java workloads, which the AKS VPA does not.
StormForge Optimize Live (CloudBolt) compared to CostAnalystPerfectScale by DoiT
Best for: Smaller AKS estates that want automated rightsizing with a free entry tier
PerfectScale supports AKS, EKS, GKE, OpenShift and Rancher and is the only automated vendor here with a published free tier sized for real clusters: up to 200 vCPUs a month. That makes it the easiest paid-grade tool to trial on one AKS cluster without a procurement cycle. Above the free tier the price is per vCPU and quoted, so ask for the rate before the trial ends.
PerfectScale by DoiT compared to CostAnalystKubecost (IBM)
Best for: Chargeback and showback per namespace across AKS and other clusters
Kubecost is the cost allocation tool most AKS teams already know, built on the same OpenCost engine as the Azure add-on but without the tier and offer-type restrictions, and with one view across AKS, EKS and GKE. It produces rightsizing recommendations and leaves them to you. For Azure customers on pay-as-you-go or CSP billing, where the native add-on is not available, it is the obvious measurement layer.
Kubecost (IBM) compared to CostAnalystCostAnalyst
Best for: Finance and FinOps teams proving what AKS rightsizing saved on the Azure invoice
CostAnalyst reads your Azure bill and Kubernetes cost data read-only, allocates AKS spend to teams and products next to the rest of your cloud and SaaS spend, and flags anomalies. It sits outside the cluster the rightsizing tool changes, so the before-and-after number finance sees does not come from the vendor being paid on it. It does not change requests or nodes.
See how CostAnalyst worksHow to choose
Pick by the problem you actually have
Your AKS workloads are Linux, stateless and under 1,000 pods per cluster
Use the free stack first. Turn on the VPA add-on in Off mode for two weeks, read the recommendations, then move to InPlaceOrRecreate on AKS 1.34 or later. Enable node auto-provisioning, or move new clusters to AKS Automatic where both are preconfigured. Measure idle charges before and after. Only buy a tool if idle charges do not fall.
You run Java services or Windows node pools
The native stack does not cover you: VPA does not support JVM workloads or Windows containers, and node auto-provisioning does not support Windows node pools. StormForge handles Java heap settings, and CAST AI or PerfectScale cover the broader cluster. Price StormForge pay-as-you-go against your requested vCPUs first, because it has no minimum.
You want one vendor to handle nodes, spot VMs and pod requests
That is CAST AI. It replaces node auto-provisioning or the cluster autoscaler, so decide that before the trial. Run its read-only agent to get a savings estimate, then price it off the AWS Marketplace card: Growth at $1,000 a month plus $5.00 per CPU, which is about $42,000 a year at 500 CPUs.
You need VPA-style rightsizing and horizontal autoscaling on the same workloads
Microsoft does not recommend VPA with an HPA on the same CPU and memory metrics, so this is the most common reason AKS teams outgrow the add-on. ScaleOps and StormForge both manage requests while HPA scales replicas. ScaleOps suits large estates that can absorb a $50,000 platform fee; StormForge suits smaller ones.
You are on pay-as-you-go or buy Azure through a CSP
The AKS cost analysis view is limited to Enterprise Agreement and Microsoft Customer Agreement offers, so it will not show you the waste. Use Kubecost, free to 250 cores, or OpenCost for namespace allocation, and a bill-level tool for the invoice view, then choose the automation.
Finance wants the savings proven by someone other than the vendor
Keep the measurement outside the cluster the vendor changes. Allocate AKS spend per team from the Azure bill before the rollout, and compare the same view after it. CostAnalyst does that read-only next to your other cloud and SaaS spend, at a published price.
Questions buyers ask
AKS rightsizing tools, answered
What is the best AKS rightsizing tool?
For Linux, stateless workloads under about 1,000 pods, the best AKS rightsizing tool is the free Vertical Pod Autoscaler add-on in InPlaceOrRecreate mode paired with node auto-provisioning. For Java services, Windows nodes, larger clusters or VPA plus HPA on the same workload, CAST AI, ScaleOps, StormForge and PerfectScale are the commercial options, chosen by which layer is wasting money.
Does AKS support vertical pod autoscaling?
Yes. AKS offers the Vertical Pod Autoscaler as an add-on based on the upstream Kubernetes VPA. It is preconfigured on AKS Automatic and configured explicitly on AKS Standard. Modes are Recreate, InPlaceOrRecreate (AKS 1.34 and later), Initial and Off; the old Auto mode is deprecated since VPA 1.4.0 and behaves like Recreate.
Does AKS support in-place pod resizing?
Yes, through the VPA InPlaceOrRecreate mode on AKS 1.34 and later. VPA tries to change requests and limits on the running pod without a restart and falls back to evicting it when an in-place change is not possible. Memory changes on many runtimes still end in a restart, so do not assume every change is disruption-free.
Can I use VPA and HPA together on AKS?
Not on the same metrics. Microsoft does not recommend using the Vertical Pod Autoscaler with a Horizontal Pod Autoscaler that scales on the same CPU and memory usage, because the two fight each other. An HPA on custom or external metrics can coexist. Commercial tools such as ScaleOps and StormForge are built to manage requests while HPA scales replicas.
Is Karpenter available on AKS?
Yes, as node auto-provisioning. NAP deploys, configures and manages Karpenter on AKS, based on the open source Karpenter and AKS Karpenter provider projects. It is preconfigured on AKS Automatic and enabled explicitly on AKS Standard. It does not support Windows node pools, IPv6 clusters or service principals, and a NAP cluster cannot be stopped.
How much does AKS cost analysis cost?
The add-on has no separate fee, but it requires the AKS Standard or Premium tier. Azure's retail price list charges Standard at $0.10 per cluster per hour, about $876 a year per cluster, and Premium at $0.60 per cluster per hour. The Kubernetes cost views also require an Enterprise Agreement or Microsoft Customer Agreement.
Does CAST AI support AKS?
Yes. CAST AI supports AKS alongside EKS, GKE and Oracle Cloud. On AKS it replaces node provisioning, uses spot capacity with fallback to on-demand, bin-packs pods and resizes requests. It installs a read-only agent first to estimate savings. Its AWS Marketplace listing prices Growth at $1,000 a month plus $5.00 per managed CPU per month.
Why does VPA not work for Java workloads on AKS?
Microsoft states that VPA does not support JVM-based workloads because it has limited visibility into their actual memory use: the JVM reserves heap up front, so container memory looks full whether the application needs it or not. Tools that tune the heap settings and the container request together, such as StormForge, are the usual answer.
How long does VPA need to learn on AKS?
The AKS VPA recommender stores up to eight days of historical data. That is enough for workloads with a daily or weekly rhythm and too short for monthly batch jobs, month-end finance runs or seasonal peaks. Commercial tools typically learn over weeks; StormForge, for example, is described as using 28 or more days of data.
Is AKS Automatic cheaper than AKS Standard?
Not automatically. AKS Automatic preconfigures node auto-provisioning and VPA, which removes much of the rightsizing work, but Azure's retail price list shows its own meters, including $0.16 per hour for the hosted control plane, against $0.10 per hour for the Standard tier. Whether it is cheaper depends on how much idle capacity the automation removes from your VM bill.
Keep reading
See your savings in dollars
Connect your cloud and SaaS spend read-only and get a prioritized savings plan. Money never moves. Create your account and get started.