How to Reduce Cloud Costs: AWS, Azure, Google Cloud, Kubernetes and AI

Give every line of the bill an owner, delete what nobody uses, resize what's left, stop test environments at night, move cold data to cheaper storage classes, cap log retention, and only then commit to the usage that never drops. AWS, Azure and Google Cloud each ship a native tool for every one of those steps. Cost Optimization Hub and Compute Optimizer on AWS, Azure Advisor, and the FinOps hub on Google Cloud list the recommendations; Savings Plans, reservations and committed use discounts cut the rate. On Kubernetes, the cost starts with each container's requests. For AI APIs, OpenAI and Anthropic let you cap spend per project or workspace, and Bedrock and Google's Agent Platform take tags or labels that carry through to billing.

Prices below are list prices in US dollars, before tax and discounts, taken from each provider's price data on October 28, 2026, with a link to the page. Other regions and machine types charge different rates.

Why does the cloud bill keep growing?

In Flexera's 2026 State of the Cloud survey of more than 750 cloud decision-makers and users, 85% name managing cloud spend as a top challenge, and wasted cloud spend rose to 29%, the first increase in five years, with AI workloads behind it. Flexera 2026 State of the Cloud. The FinOps Foundation's State of FinOps 2026 drew 1,192 respondents who together spend more than $83 billion a year on cloud, and they rank workload optimization and waste reduction as their single top priority today. State of FinOps 2026.

Teams past the first clean-up find smaller work left. One practitioner told the Foundation: "We have hit the 'big rocks' of waste and now face a high volume of smaller opportunities that require more effort to capture." AI spend joined the list fast: 98% of respondents now manage it, up from 31% two years ago, and many say they're asked to fund AI projects out of what optimization saves. State of FinOps 2026.

Where do you start cutting cloud costs?

With allocation, a capability of its own in the FinOps Framework. Until you know which team, product or environment each line of the bill belongs to, every saving turns into an argument about who pays. On AWS, user-defined and AWS-generated cost allocation tags appear in Cost Explorer only after you activate them, and Cost Categories sit on top. Azure offers tag inheritance and cost allocation rules for shared costs. Google Cloud uses labels and the billing export to BigQuery. Our primer What is FinOps? covers the rest of the framework.

Then open your provider's recommendation hub. AWS Cost Optimization Hub consolidates and prioritizes cost recommendations across your accounts and Regions. Azure gathers its own in Advisor cost recommendations. The Google Cloud FinOps hub shows every active savings opportunity in one dashboard, built from Cloud Billing and the recommenders Google groups under Active Assist.

Which levers work on every cloud?

Two FinOps Framework capabilities cover most rows of the table below. Usage Optimization matches resources to actual usage patterns. Rate Optimization lowers what you pay for the same usage through negotiated discounts, commitment discounts and other pricing mechanisms. Each cell links to the provider's own feature.

How do you pay less for the same usage?

Commit to the usage that runs all month. On AWS, Compute Savings Plans take up to 66% off and EC2 Instance Savings Plans up to 72% (Savings Plans pricing), and Reserved Instances are still on offer. Azure reservations cut up to 72% from pay-as-you-go prices and work as a billing discount that leaves the resource untouched (reservations); the savings plan for compute reaches 65% and the one for databases 35% (Azure savings plans). Google Cloud committed use discounts run for one or three years and come spend-based or resource-based (CUDs). Resource-based CUDs take up to 55% off most machine types and up to 70% off some (resource-based CUDs), and you can't cancel a spend-based one (spend-based CUDs). On self-serve (online) Cloud Billing accounts, Google also applies sustained use discounts on its own to resources used more than 25% of a month, up to 30% net for a VM that runs the whole month (SUDs).

Rightsize first, then commit to the floor: the usage you never drop below. Our guide Savings Plans vs Reserved Instances works through the sizing on AWS.

Work that can stop and restart, such as batch jobs and CI runners, belongs on spot capacity. AWS advertises Spot "for up to 90% off" and warns you two minutes before it stops or terminates the instance (EC2 Spot · interruption notices). Azure Spot VMs reach 90% off pay-as-you-go too, with 30 seconds of eviction notice (Azure Spot · eviction). Google Cloud Spot VMs go up to 91% off for many machine types, GPUs, TPUs and Local SSD, and you set the preemption notice when you create the VM; Google recommends 120 seconds for workloads that need more than 30 (Spot VMs).

How do you cut waste without touching production?

Start with idle resources, test environments, cold data and old logs, which sit away from the traffic your customers send.

What's running with nobody using it?

Azure Advisor flags resources that saw no use at all over the last seven days and recommends shutting them down (Advisor). On AWS, Compute Optimizer lists idle resources and Trusted Advisor runs cost checks. Google Cloud makes recommendations for idle VMs and idle disks, IP addresses and images. Ask the owner before you delete: a detached disk can be somebody's only backup.

Do test environments run all night?

A VM that runs 12 hours a day on weekdays bills about 260 hours a month instead of 730. In US East, an AWS m7i.large at $0.1008 an hour drops from $73.58 to $26.21 a month, and an Azure D2s v5 at $0.096 an hour drops from $70.08 to $24.96: 64% less on both (EC2 On-Demand pricing · Azure Linux VM pricing). To schedule it, AWS has Instance Scheduler on AWS, Azure has VM auto-shutdown, and Google Cloud has instance schedules. Azure moved its older Start/Stop VMs solution to the previous-versions docs with no further development planned (Microsoft's note), so start from auto-shutdown.

Is cold data in the right storage class?

Ten TiB in Cloud Storage Standard costs about $204.80 a month in Iowa (us-central1) (Cloud Storage pricing). Data nobody has read in months can move to Nearline, Coldline or Archive; Nearline carries a 30-day minimum (storage classes). Lifecycle rules move objects on a schedule, and Autoclass moves each one by its access pattern. Azure Blob Storage has hot, cool, cold and archive tiers, and data in cool should stay at least 30 days (access tiers · lifecycle management). On AWS, an S3 Lifecycle rule can move objects to Standard-IA after 30 days or to Glacier Flexible Retrieval after a year, and S3 Intelligent-Tiering moves objects nobody reads to cheaper tiers for a small monthly monitoring charge.

How long do you keep logs?

CloudWatch Logs keeps log data indefinitely by default, until you set a retention period on each log group (log groups). Azure Log Analytics tables keep data for 30 days by default, 90 for some, and the Basic and Auxiliary table plans store verbose data for less (retention · table plans). Google Cloud Logging charges $0.01 per GiB per month for logs kept past 30 days and nothing for logs kept the default period (Observability pricing), with custom retention from 1 to 3,650 days per bucket (log buckets). Five TiB held past 30 days costs $51.20 a month, every month, until somebody shortens the retention.

What does NAT traffic cost you?

An AWS NAT gateway bills for every hour it's available and every GB it processes (NAT gateway pricing), and traffic to S3 and DynamoDB can go through gateway endpoints, which carry no additional charge (gateway endpoints). Azure NAT Gateway bills resource hours plus data processed per GB, with bandwidth charges on top; data coming in and traffic within the same Availability Zone are free (NAT Gateway pricing · bandwidth pricing), and Azure also offers virtual network service endpoints. On Google Cloud, Cloud NAT has its own price list, and Private Google Access lets VMs with no external IP reach Google APIs. How to Reduce Your AWS Bill prices the NAT fix in dollars.

Does the region change the price?

Yes. At list price, the same VM costs 59.4% more in São Paulo than in US East on both AWS and Azure, and Google Cloud Storage Standard costs 75% more.

The gap holds on other AWS types: a t3.medium costs $0.0416 an hour in us-east-1 and $0.0672 in sa-east-1, 61.5% more (EC2 On-Demand pricing). Latency and data residency rules limit what can move. Backups, batch jobs and test environments often can; add the cross-region transfer before you move them. If you run in Brazil, AWS billing in Brazil covers how AWS invoices in reais.

How do you reduce Kubernetes costs?

Kubernetes cost starts with requests. The kube-scheduler reads each container's requests to decide which node gets the pod (resource management), so an inflated request reserves CPU and memory that nothing uses and keeps extra nodes running. GKE Autopilot makes the link explicit: its price depends on the compute resources your workloads request, while GKE Standard node pools bill the Compute Engine VMs underneath (GKE pricing).

• Fit requests to real usage. The Vertical Pod Autoscaler updates a workload's resource to adjust its requests.
• Scale replicas with the Horizontal Pod Autoscaler and nodes with a cluster autoscaler: Karpenter or Cluster Autoscaler on EKS, the AKS cluster autoscaler, the GKE cluster autoscaler.
• Put interruptible workloads on spot nodes. Microsoft's own AKS cost list names AKS Automatic, VPA, HPA and KEDA, Spot VMs (up to 90% savings), Arm64 and Reserved Instances (up to 72% savings) (AKS cost best practices).
• See cost per namespace. AWS split cost allocation data adds Kubernetes pod and ECS task costs to the CUR, which used to stop at the EC2 instance. The AKS cost analysis add-on shows cost by cluster and namespace and is built on OpenCost. GKE cost allocation breaks the bill down by namespace and label.

Google's best practices for cost-effective Kubernetes on GKE go deeper, and we compare the tools in The Best Kubernetes Cost Management Tools.

How do you control AI and model API spend?

Give each team its own billing unit with each provider, and set a limit or an alert on it. When every service shares one API key, the bill arrives with no owner.

• OpenAI lets you set two controls on an organization or a project: a spend alert, which sends a notification while API traffic continues, and a hard spend limit, at which affected requests return a 429 error. OpenAI also sets an approved monthly usage limit per organization, separate from yours, that rises with the usage tier: $500 on Build, $5,000 on Launch, $200,000 on Grow (rate limits).
• Anthropic workspaces separate projects, environments or teams under one bill, each with its own spend and rate limits set below the organization's, and the Claude Code workspace is the one that supports per-user monthly spend limits (workspaces). The Usage and Cost Admin API returns historical usage and cost for your own reports.
• On Amazon Bedrock, attach tags to an application inference profile to track the cost of on-demand model calls (inference profiles). AWS Budgets and Cost Anomaly Detection cover Bedrock like any other service.
• On Gemini Enterprise Agent Platform (formerly Vertex AI), custom metadata labels on API calls go through to billing, so you can break charges down by label (labels · the rename).
• On Azure, Microsoft Foundry's cost guide covers estimating spend before deployment, tracking it in real time and setting alerts (Foundry costs).

For prices per million tokens, read LLM API Pricing. What AI Coding Tools Cost per Developer works out the bill for Claude Code, Codex, Cursor and Copilot, and How to Track AI Costs by Team walks through the setup for each provider. The tools that read AI bills are in The Best FinOps Tools for AI.

How do you keep the bill from growing back?

• Set budgets on every provider, and decide which ones should act on their own. AWS Budgets alert on actual and forecasted spend, and a budget action can apply an IAM policy or a service control policy, or target specific EC2 or RDS instances, either automatically or after your approval. Azure evaluates budgets every 24 hours and sends the email "normally within an hour of the evaluation" (Azure budgets). Google says an alerts-only budget "doesn't automatically cap" usage or spending (Google Cloud budgets). A Google spend cap budget pauses new usage of the Gemini API, Agent Platform, Cloud Run or Cloud Run functions in one project once estimated costs pass 100% of the budget.
• Turn on anomaly detection. AWS Cost Anomaly Detection uses machine learning models to detect and alert on unusual spend, Azure sends anomaly alerts at subscription scope, and Google Cloud has an anomalies dashboard.
• Name an owner for every account, subscription and project, and go through the top lines with them once a month.
• Price changes before they ship. Pre-deployment architecture costing came up as a top capability FinOps teams want from their tools (State of FinOps 2026), and Terraform cost in pull requests shows one way to do it.

Common mistakes when cutting cloud costs

• Buying a commitment before you delete idle resources and rightsize, which locks the discount onto waste. A Google Cloud spend-based CUD can't be cancelled.
• Treating an alerts-only budget as a cap. It sends an email; the resources keep running.
• Moving a workload to spot when it can't survive an interruption with 30 seconds of warning, Azure's notice.
• Sending data to a cold class while someone still reads it every week.
• Comparing regions on the VM price alone, without latency and cross-region transfer.
• Counting the saving from the recommendation instead of checking next month's bill.

Guides by provider

• AWS: How to Reduce Your AWS Bill, with us-east-1 prices for NAT, gp3, logs and Savings Plans, and The Best AWS Cost Optimization Tools.
• Azure: How to reduce Azure costs and the best Azure FinOps tools.
• Google Cloud: How to reduce Google Cloud costs and the best Google Cloud FinOps tools.
• Kubernetes: The Best Kubernetes Cost Management Tools.
• AI: LLM API Pricing, What AI Coding Tools Cost per Developer, How to Track AI Costs by Team and The Best FinOps Tools for AI.
• Across clouds: The Best Cloud Cost Optimization Tools, and FinOps consulting or a platform if your team lacks the hours.

Where LevelFour fits

The native tools in the table each cover their own cloud. LevelFour reads your Cloud, SaaS, and AI bills in one view, finds the savings, and ships each one: applied in your cloud or SaaS account with access scoped to that one change, as a pull request to your infrastructure code, or as manual steps. Savings Audit confirms each saving on the bill, and Commitments covers Savings Plans and Reserved Instances bought and renewed. Cloud Savings costs 30% of realized savings, and plans start at $30 a month (pricing). You can also buy LevelFour on the AWS, Google Cloud and Azure Marketplaces.

See what LevelFour can save you in 5 minutes, with read-only access.

Frequently asked questions