Kubernetes Rightsizing Tools: Pod Rightsizing Software and Container Rightsizing Platforms Compared

Best for: Every cluster, before any tool is purchased

The honest first answer, and it moved a long way in the last two releases. The Vertical Pod Autoscaler produces CPU and memory request recommendations from observed usage and can apply them for you. Its update modes matter more than its recommendations: Off produces numbers and changes nothing, Recreate evicts pods to apply them, Auto was deprecated in VPA 1.4.0 and is now an alias for Recreate, and InPlaceOrRecreate attempts an in-place update first and falls back to eviction when it cannot complete one in time. InPlaceOrRecreate is an alpha feature gate that defaults to disabled and has to be enabled on both the updater and the admission controller. Underneath it sits InPlacePodVerticalScaling, which reached stable in Kubernetes v1.35. Goldilocks from Fairwinds wraps VPA in a dashboard so you can read recommendations per namespace without adopting automation at all. Start here for a fortnight. You will learn how much waste you actually have, and the number usually justifies whatever you buy next.

Best for: Continuous automated rightsizing that stays inside your cluster

Positioned as autonomous rather than advisory: it continuously right-sizes CPU, memory, and nodes across VPA, HPA, and Karpenter, then adds smart pod placement and replica optimization so the freed capacity turns into fewer nodes instead of more idle. That combination is the reason it belongs near the top of a shortlist, because rightsizing without consolidation does not move the invoice. It installs with a single helm command and is available self-hosted, as ScaleOps Cloud, or air-gapped, which matters for security reviews that will not accept an external control plane reading production telemetry. A separate GPU platform covers fractional GPU sharing, GPU replica optimization, and batch inference optimization. Pricing is not published. Vendor savings claims run to eighty percent, with named case studies at sixty-two percent CPU and forty percent memory, so treat the headline as marketing and the case studies as directional.

Best for: Teams that want to tune the HPA and the requests together

Now sold as StormForge by CloudBolt following CloudBolt acquisition of StormForge, announced 31 March 2025 with terms undisclosed. The distinguishing idea is bi-dimensional autoscaling: instead of treating vertical rightsizing and horizontal pod autoscaling as separate problems, it adjusts resource requests and limits alongside the HPA target utilization so the two do not fight each other. That is a real gap in most tooling, where an aggressive rightsizer lowers requests, the HPA sees higher utilization, and replicas climb until the savings evaporate. Machine learning drives the recommendations, and the vendor describes the patents as patent-pending. The trial is unusually concrete for this category: full optimization on one cluster for thirty days, which is enough to produce an honest before-and-after on real workloads. Headline savings claims of up to eighty percent are vendor claims.

Best for: Aggressive automation across rightsizing, bin-packing, and spot

The widest automation surface here. Alongside workload rightsizing it runs cluster optimization, Karpenter cluster optimization, spot instance handling with automated fallback, a GPU product, and separate database and storage optimization. If your goal is the largest possible reduction and you are prepared to let software make production changes without a human in the loop, this is the category benchmark and has been for several years. Two things to plan for. First, that automation needs meaningful write access, which is the longest conversation in most security reviews. Second, CAST AI no longer publishes pricing: the pricing page now states that the model depends on factors specific to your environment and that they need to get in touch to give accurate pricing. Any per-CPU figure you find in an older comparison is stale, including ones still circulating in competitor blog posts.

Best for: Starting free and scaling by vCPU with a published entry tier

The most transparent entry point in the automated group and now part of DoiT International. Pricing is based on monthly vCPU consumption with a genuinely usable free tier up to 200 vCPU per month, then Advanced and Expert tiers quoted per vCPU. It covers cost optimization and waste detection alongside performance and resilience issues, which is a useful framing because the fastest way to lose a rightsizing program is one OOM kill in production that nobody predicted. It works with Karpenter, Cluster Autoscaler, and Node Auto Provisioning, and is described as agnostic to all Kubernetes distributions and cloud environments including private data centers, with a self-hosted option. Worth noting that the per vCPU model scales with the cluster you are trying to shrink, so model the cost at your current size and at your target size before signing.

Best for: Policy-controlled automation with a human in the loop by default

Densify has rebranded: densify.com now redirects to kubex.ai, and the product line is split into an Automated Pod Scaler for right-sized pod requests, a Container Sizer for new containers with no usage history, an HPA Optimizer for autoscaling specs, and a Node Optimizer that recommends instance types and CPU-to-memory ratios. A Kubex Automation Controller executes changes within policy. The reason it earns a place on a shortlist is the explicit human-in-the-loop or fully automated choice, which lets you pass the security review in advisory mode and turn automation on per namespace once the recommendations have proven themselves. Densify heritage also means unusually strong node-shape analysis, which is the half of the problem most Kubernetes-only tools ignore. No published pricing; a free trial is offered. The savings figures on the site, including a twenty to sixty percent range, are vendor claims.

Best for: Buyers who want the price tied to realized savings

Zesty covers multi-dimensional autoscaling, adaptive pod placement to raise node utilization, persistent volume autoscaling to cut storage cost, and compute cost visibility, which makes it broader than a pure pod rightsizer. Persistent volume autoscaling is genuinely differentiated: over-provisioned PVCs are a quiet, permanent line item that almost no rightsizing tool touches, partly because shrinking a PVC is not a thing Kubernetes lets you do casually. Commercially it takes the savings-share position, stating that pricing is based on actual usage and realized savings and that it will never be more than the value delivered, with a claimed minimum 1:3 ROI across the customer base. No tiers, numbers, or free tier are published, and any engagement starts with an analysis of your Cost and Usage Report. If your finance team prefers a variable cost that only appears when savings do, this is the model to ask about.

Best for: Autonomous optimization beyond Kubernetes, including serverless

The broadest scope on this list and the clearest position on autonomy. Sedai uses reinforcement learning to find configurations for containers on EKS, AKS, and GKE, but also for VMs, serverless functions, storage, and data and streaming workloads, so it suits an organization where Kubernetes is one of several places money leaks rather than the only one. It ships two explicit modes: Copilot, where a human approves each optimization with a click, and Autopilot, where it executes without manual approval. That progression is the right shape for a nervous platform team, and the vendor leans on it, citing eight United States patents on autonomous action in cloud environments and a claim of zero incidents caused by its autonomous actions. No pricing is published. Named customer results include fifty percent savings and a forty-six percent Kubernetes reduction, which are vendor-published case studies rather than independent measurements.

Best for: Seeing the waste per namespace before automating anything

Kubecost is a cost allocation and visibility product with request-sizing recommendations attached, not an actuator, and it is worth being precise about that because it appears on almost every rightsizing shortlist. Built on the open-source OpenCost project and acquired by IBM in September 2024, it allocates container spend by namespace, deployment, and label, and surfaces where requests exceed observed usage. The free Foundations tier covers unlimited clusters up to 250 cores with 15-day metric retention, which is enough to quantify the opportunity across a real estate of clusters at zero cost. Use it to build the business case and to keep score afterwards, then pair it with one of the actuators above when someone approves the change. Treating it as a rightsizing tool on its own is how teams end up with a beautiful dashboard and an unchanged bill.

Best for: Finance owners who need the Kubernetes saving in the wider budget

We are on this list for completeness and we are firmly in the recommendation group. Costanalyst connects clouds, Kubernetes, and SaaS subscriptions read-only, reports waste in dollars with the underlying line items, attributes it to a team or cost center, and raises anomaly alerts before the invoice lands. It never holds write access, never resizes a pod, and never moves money, which makes the security review short and the automation column empty. Use it when the question in the room is whether the Kubernetes saving is material next to everything else in the technology budget, or when the platform team has produced a rightsizing number and finance wants it in the same view as the cloud commitments and the SaaS renewals. If the clusters need something to actually apply the change, buy ScaleOps, StormForge, CAST AI, or PerfectScale for that job and keep us for the reporting.