Quick Answer
Amazon EKS charges about $73 a month per cluster for the control plane (a $0.10 per hour fee), before a single pod runs. It jumps 6x to about $438 a month if your cluster falls to extended support. On top of that you pay for compute (EC2 nodes, Auto Mode, or Fargate), storage, and data transfer. For AI workloads on GPU nodes, utilization decides the real cost.
Amazon EKS pricing has two components: a fixed control plane fee per cluster ($0.10/hr on standard support, $0.60/hr on extended support) and variable charges for the EC2 nodes, storage, and data transfer your workloads consume.
The most expensive number in Amazon EKS pricing is also the smallest one on the page. The control plane fee is $0.10 an hour, about $73 a month per cluster, and you pay it before a single container runs. Let the cluster fall behind on its Kubernetes version and that same fee jumps to $0.60 an hour, about $438 a month, a 6x increase that happens automatically, without warning, and without anything else changing.
Now multiply by reality. A team running dev, staging, and production pays roughly $219 a month in control plane fees with zero workloads deployed. Let those versions lapse and the same three clusters cost about $1,314 a month, purely on the version number.
For a finance leader, that is the problem with EKS pricing in one paragraph: the costs that surprise you are not the ones on the headline rate. And the stakes are rising, because Kubernetes is where AI workloads are moving (GPU node groups for training and inference, and a control plane tier AWS built specifically for AI and ML). 40% of companies now spend more than $10M a year on AI, and most cannot tell whether it is worth it. EKS is increasingly where that money goes.
This guide is for the person who has to forecast and defend that number. It covers every layer of Amazon EKS pricing, including the three layers AWS added in late 2025, what AI workloads actually cost on EKS, and how to tie the spend to the return. CloudZero is the AI ROI company, so we end on the part finance owns.
Key takeaways
- The control plane fee is fixed overhead you pay per cluster, per hour, forever. About $73 a month standard, about $438 on extended support, multiplied by every cluster and environment you run.
- Compute is where EKS actually scales. Three ways to run nodes (EC2, Auto Mode, Fargate), each priced differently, plus storage, public IPv4, and cross-AZ transfer.
- EKS added three new billing layers in 2025. Provisioned Control Plane (for AI and ML), EKS Capabilities, and Hybrid Nodes. Most pricing guides have not caught up.
- ECS has no control plane fee. That free-versus-$73 gap is the core of the EKS-versus-ECS cost decision.
- AI workloads change the math. GPU node groups and idle GPUs drive cost to serve, and utilization, not the node rate, decides your margin.
- Forecasting and ROI are the finance job. Tie every cluster, pod, and GPU to the customer, feature, and model it serves.
Report
Finance needs to prove AI’s return: CloudZero report
260 senior finance leaders (more than half CFOs) told us why the speed of seeing AI spend, not the size of it, separates who pulls ahead on AI from who gets burned.
Amazon EKS pricing at a glance
Amazon EKS pricing has more than five distinct layers. Here is the full structure, US East (N. Virginia), at publication:
| Billing layer | What it costs | Notes |
| Control plane, standard support | $0.10/hr (~$73/mo per cluster) | First 14 months a Kubernetes version is supported |
| Control plane, extended support | $0.60/hr (~$438/mo per cluster) | Months 15 to 26; a 6x jump, automatic |
| Provisioned Control Plane | XL $1.65, 2XL $3.40, 4XL $6.90, 8XL $13.90/hr | For AI and ML training, multi-tenant SaaS, high API volume; on top of the cluster fee |
| Compute: EC2 nodes | Standard EC2 + EBS rates, no EKS surcharge | You manage the nodes; plus IPv4 $0.005/hr per address and cross-AZ transfer |
| Compute: EKS Auto Mode | EC2 cost plus ~12% management charge | AWS manages nodes via Karpenter; per-second, 1-minute minimum |
| Compute: AWS Fargate | Per vCPU and per GB of memory used | Per-second, 1-minute minimum; no Fargate Spot for EKS pods |
| EKS Capabilities (2025) | Per managed capability | Managed Argo CD, ACK, and KRO |
| Hybrid Nodes (2025) | Per vCPU-hour, tiered by volume | On-premises or edge VMs attached to an EKS control plane |
| Storage and registry | EBS by volume and snapshots; ECR $0.10/GB-mo | ECR to EKS transfer in-Region is free |
The headline rate is the cluster fee. The bill is everything below it.
What Amazon EKS is, and why it shows up on your bill
Amazon EKS is AWS’s managed Kubernetes service. It runs and maintains the Kubernetes control plane (the part that schedules containers, stores cluster state, and keeps everything highly available across Availability Zones) so your team can run containerized apps without operating Kubernetes themselves. Your workloads run on worker nodes, which you provide as EC2 instances, Auto Mode, or Fargate.
For finance, EKS matters for two reasons. It carries a fixed per-cluster fee that scales with how many clusters and environments you run, not with usage, and it is increasingly the home of AI and ML workloads, where the compute underneath it gets expensive fast.
That combination (fixed overhead plus volatile AI compute) is what makes a Kubernetes line on the AWS bill hard to forecast.
See also:
What is the EKS control plane fee and what triggers the extended-support jump?
Every EKS cluster pays a flat hourly fee for its managed control plane, whether or not any nodes are running. It is fixed overhead, metered from the moment the cluster exists. The number depends entirely on your Kubernetes version support status:
| Support tier | Rate | Monthly per cluster | When it applies |
| Standard support | $0.10/hr | ~$73 | First 14 months a version is supported in EKS |
| Extended support | $0.60/hr | ~$438 | Month 15 to 26 on an aging version |
Across the full 26-month version lifecycle, a cluster left to drift averages about $0.33 an hour.
The $73-to-$438 delta is where most surprise EKS bills come from, because the jump is automatic: a version ages out of standard support and the fee multiplies 6x without anyone touching the cluster.
The fix is almost free. Upgrade the Kubernetes version before the 14-month standard-support window closes, and track version lifecycles the way you track commitments. One line in your infrastructure code (a version pin with a reminder to upgrade) prevents a roughly $365-a-month-per-cluster surprise.
The multiplier is the part finance should watch. The fee is per cluster, so it compounds across environments and teams. Three environments is about $219 a month before any pods exist; the same three on extended support is about $1,314. At an enterprise running dozens of clusters, control plane fees alone become a real line item, and idle or forgotten non-production clusters are pure waste.
What does Amazon EKS actually charge for in 2025?
EKS pricing is now a stack of layers, and AWS added three of them in late 2025 that most pricing guides still do not cover. Knowing they exist is the difference between a forecast that holds and one that does not.
The compute layer is the dominant cost for almost every cluster, and there are three ways to run it:
| Compute mode | How it is priced | Cheapest for | The catch |
| EC2 nodes (managed or self-managed) | Standard EC2 plus EBS, no EKS surcharge | Steady, high-utilization workloads at scale | You manage patching, scaling, upgrades; plus IPv4 $0.005/hr per address and cross-AZ transfer |
| EKS Auto Mode | EC2 cost plus a roughly 12% management charge | Teams wanting Kubernetes without managing node groups | The premium adds up at scale; over 150 nodes, contact AWS for volume pricing |
| AWS Fargate | Per vCPU and per GB memory, from image pull to pod termination | Bursty, sporadic, or short-lived workloads | No Fargate Spot for EKS pods (it is ECS-only); steady workloads are cheaper on EC2 |
A note that trips up teams moving from ECS: Fargate Spot does not exist for EKS pods. If a cost plan assumes Fargate Spot on EKS, that option is not available. Compute Savings Plans do apply to Fargate (up to about 50%) and to EC2 nodes, which is the lever for predictable workloads.
The three 2025 additions:
- EKS Provisioned Control Plane (GA November 2025) gives a cluster pre-allocated control plane capacity in tiers: XL about $1.65/hr, 2XL $3.40, 4XL $6.90, 8XL $13.90, on top of the cluster fee. AWS positions it for AI and ML training, multi-tenant SaaS, and high API-call volumes, which is where heavy Kubernetes workloads now live. For an AI team, this is a real new line to budget.
- EKS Capabilities (GA November 2025) are managed versions of Argo CD, ACK, and KRO, billed per capability, trading operational effort for a fee. See the add-ons landscape at https://www.cloudzero.com/blog/eks-add-ons/.
- EKS Hybrid Nodes let you attach on-premises or edge VMs to an EKS control plane, billed per vCPU-hour and tiered by monthly volume.
Add storage and registry on top: EBS volumes and snapshots, and Amazon ECR at $0.10 per GB-month (transfer from ECR to EKS in the same Region is free). For the networking side, see CloudZero’s guide to Amazon VPC pricing and CloudTrail vs Cloudwatch.
What does a typical EKS cluster cost per month?
A single small production cluster, three m5.large nodes running 24/7 on-demand in us-east-1, illustrative:
| Line item | Example monthly |
| Control plane (standard support) | ~$73 |
| EC2 compute (3x m5.large at ~$0.096/hr) | ~$210 |
| EBS storage (100 GB gp3) | ~$8 |
| Public IPv4 (3 addresses at $0.005/hr) | ~$11 |
| Cross-AZ and misc data transfer | ~$10 |
| Total | ~$312 |
Move one node to Spot (up to about 90% off the compute portion) and the total drops toward $260. Let the cluster slip to extended support and the control plane line alone goes from $73 to about $438, taking the total past $670 with no change to the workload. Same cluster, different version number.
Is EKS or ECS cheaper? The control plane cost compared
The single biggest cost difference between Amazon EKS and Amazon ECS is the control plane fee: EKS charges about $73 a month per cluster; ECS charges nothing for its control plane. On the EC2 launch type, ECS adds no orchestration fee at all; you pay only for the compute.
| Amazon EKS | Amazon ECS | |
| Control plane fee | ~$73/mo per cluster (more on extended support) | Free |
| Density | High pod density per node, varies by instance type and CNI configuration | Lower density, simpler task packing |
| Portability | Standard Kubernetes, portable across clouds | AWS-native, higher lock-in |
| Best for | Teams already on Kubernetes or needing portability | AWS-only workloads wanting simplicity |
So when is the EKS premium worth it? When you need the Kubernetes ecosystem, multi-cloud portability, or the higher pod density, the per-cluster fee is a rounding error against the compute.
When you are AWS-only and want the simplest path, ECS removes a fixed cost entirely.
How AI changes EKS pricing: from cluster cost to cost per outcome
AI is now one of the biggest drivers of EKS spend, and it has changed the question finance has to answer. Before AI, an EKS cluster mostly ran APIs, web apps, microservices, and background jobs on relatively inexpensive CPU instances, and the cost conversation was about right-sizing containers, Spot, and cluster utilization.
Now teams run LLM inference, vector databases, RAG pipelines, fine-tuning jobs, and AI agents on the same clusters, and those need expensive GPU nodes. A handful of GPUs can cost more than an entire traditional EKS environment. Six shifts explain why, and each is a place finance loses visibility.
- GPU cost dominates the bill. A small GPU footprint can outweigh every CPU workload in the cluster combined, so the cost center moves from many cheap nodes to a few expensive ones.
- Overprovisioning gets worse. AI workloads are hard to predict, so teams reserve more CPU, more memory, and dedicated GPU capacity “just in case.” The result is clusters with low utilization and high spend.
- Inference runs around the clock. Training is expensive, but inference is often the larger long-term cost, because teams keep AI services running continuously to avoid startup latency. That becomes idle GPU and node cost and excess capacity during off-peak hours, and idle GPU is the most expensive idle there is.
- Data movement adds a second bill. AI systems constantly move data between Amazon S3, databases, vector stores, model endpoints, and the cluster, which raises network transfer, storage, and cross-AZ costs. Teams fixate on GPU spend and miss these.
- Clusters multiply. Instead of one production cluster, AI initiatives spin up development, experimentation, training, and inference clusters, then forget to clean them up, so the per-cluster control plane fee from the trap above compounds with every forgotten environment.
- Visibility breaks. CPU and memory utilization no longer tell the story. The numbers finance now needs are cost per inference, cost per AI request, cost per customer conversation, and cost per generated document, which is the unit economics of the workload.
The result is that the AI workload becomes the dominant cost center even when it serves fewer applications. An illustrative monthly split:
| Workload | Illustrative monthly EKS cost |
| Traditional microservices | $8,000 |
| AI inference service | $35,000 |
| Idle GPU time | $12,000 |
| Data transfer and storage | $5,000 |
In that picture, idle GPU time alone outweighs the entire microservices footprint. The same rule from EC2 applies inside the cluster: for AI, the node rate is not the cost to serve, utilization is.
A GPU node costs the same per hour whether it is fully loaded or idle, so utilization decides your cost per million tokens, and idle GPU capacity quietly multiplies your effective cost several times over.
Karpenter consolidation and Spot help, but only if someone is watching utilization, and pod-level GPU cost is invisible on a standard AWS bill. AWS even built the new Provisioned Control Plane tier (XL through 8XL) for AI and ML training, which is another line to budget.
This is the AI ROI question finance now owns, and it is a different question than before. The conversation has moved from “how much does my cluster cost” to “how much does each AI outcome cost.” Answering it means comparing self-hosting on EKS GPU nodes against managed APIs at your real volume.
Related Reads:
How to forecast and reduce your EKS spend
Estimate with AWS pricing calculator then add the layers calculators miss: the per-cluster control plane fee across every environment, idle GPU time, and cost per inference. Then reduce, in rough priority order:
- Stay on supported Kubernetes versions. The single cheapest win. Upgrading before the 14-month window closes avoids the 6x extended-support jump on every cluster.
- Kill idle and forgotten clusters and nodes. Non-production clusters left running are pure control-plane-plus-compute waste, and idle GPU nodes are the most expensive idle of all.
- Right-size and move to Graviton. ARM-based Graviton delivers roughly 10 to 20% better price-performance for most containerized workloads, often with no code changes.
- Cover steady compute with commitments. Compute Savings Plans cover EC2 nodes and Fargate; track coverage and utilization as the finance KPIs, not just total spend.
- Match the compute model to the workload. Steady on EC2 nodes, bursty on Fargate, batch and fault-tolerant on Spot.
- Right-size the control plane. Use Provisioned Control Plane tiers only where AI, ML, or high API volume actually needs them.
All of this controls cost. In the AI era, that is only half of finance’s job.
Why controlling cost is only half of finance’s job
Controlling cost is the floor; proving return is the job. The old Kubernetes playbook (right-size, commit, use Spot, stay current on versions) still works, but AI changed the terms.
The unit changed, because for AI a high node rate can be the cheapest option if it serves more tokens. Waste got brutal, because an idle GPU node wastes far more than an idle web server.
And forecasting broke, because AI spend on the cluster scales with product adoption, not a fixed node count.
The result is a measurement gap. Only about half of organizations (51%) are confident they can measure the ROI of their AI investments (CloudZero, State of AI Costs 2025). On EKS, where pod and GPU costs hide behind a single Kubernetes line, that gap is wider than most teams realize. Closing it takes unit cost intelligence, not a rate card.
How CloudZero turns EKS and AI spend into provable ROI
CloudZero gives finance the numbers to forecast and defend Kubernetes and AI spend, mapped to the problems in this guide. CloudZero Kubernetes Cost Analysis combines metrics from AWS Container Insights with billing data to calculate Kubernetes costs automatically, then allocates them the way the business thinks.
| The finance problem | What CloudZero does |
| One Kubernetes line, no idea what drives it | Cost per pod, namespace, cluster, deployment, and per team, product, customer, and AI model |
| Control plane fees multiplying across clusters | Surfaces every cluster and its fees, so idle and forgotten non-production clusters are visible |
| Surprise extended-support jumps and runaway GPU jobs | Real-time, hourly visibility with anomaly detection that flags the spike before the invoice closes |
| AI spend that scales with adoption | Budgets and forecasting built for spend that moves with usage, not a fixed node count |
| Idle GPU nodes eroding margin | Surfaces idle and oversized capacity; CloudZero Advisor (https://advisor.cloudzero.com/) recommends cheaper fits, including Graviton |
| Commitment decisions made on guesswork | Tracks Savings Plan coverage and utilization across EC2 nodes and Fargate |
| AI spend split across EKS GPUs and managed APIs | AI Hub unifies cluster GPU spend with OpenAI, Anthropic, and Amazon Bedrock into total AI spend and cost per model |
| No way to prove return to the board | Ties spend to revenue-driving units, so you can show cost per customer and gross margin |
The results show up on the P&L. Symphony Talent cut AWS costs by 48% after cost-per-client visibility revealed it had been losing money on a client for years, which reshaped its next contract negotiation. Diaceutics reduced its AWS bill by 41% and beat its S3 savings target by 70%. Drift improved COGS by $2.4M. And Upstart reduced cloud costs by $20M through engineering visibility and accountability. Finance and engineering teams at Toyota, Skyscanner, Grammarly, Duolingo, and Upstart run from the same numbers.
or take the self-guided product tour to explore at your own pace.
Amazon EKS pricing FAQs
