Reducing Cloud Costs: The Complete Guide for IT Leaders, Architects, and Engineers

In most organizations, five categories account for the bulk of the expenses:

• Compute: virtual machines, containers, and serverless.
• Databases: managed databases, often generously sized and rarely decommissioned.
• Storage: volumes, object storage, snapshots, and backups.
• Network: data traffic to the Internet, between zones and regions, and NAT gateways.
• Managed services: logging, monitoring, load balancers, and, increasingly, AI services.
• Start with the five largest services and the five largest accounts or subscriptions. That's almost always where the biggest savings are to be found.

Data traffic is often the least understood cost item, because you pay for it by the gigabyte and it’s generated in many different places.

What helps:
Gateway endpoints for S3 and DynamoDB instead of routing traffic through NAT; limiting traffic between zones where availability permits; using a CDN and caching for traffic to users; compression; and processing data in the region where it is stored.

The EU Data Act.
As of September 12, 2025, cloud providers may only charge costs they actually incur when a customer switches to another provider or to an on-premises solution. Starting January 12, 2027, these switching charges will be prohibited, including egress fees for transferring your data. However, standard data traffic during use—such as traffic to your users or between regions—will still be subject to charges. AWS, Azure, and Google each offer their own procedure for waiving egress fees during a migration, with their own terms and deadlines. For the assessment in Section 13, this means that exit costs will decrease while ongoing data traffic costs will remain.

FinOps is an approach in which finance, engineering, procurement, and management work together to drive the value of technology. Those who incur costs are also the ones who see them, and decisions are based on the value they deliver. The FinOps Foundation (part of the Linux Foundation) manages the framework.

Three phases that keep recurring:

1. Inform: Show who is spending what, and allocate costs to teams and products.
2. Optimize: Save money through usage (rightsizing, streamlining, architecture) and through pricing (commitments, negotiation).
3. Operate: Embed budgets, alerts, regular reviews, and ownership within the organization.

Since 2025, FinOps has extended beyond the public cloud. The framework now includes the following scopes: public cloud, SaaS, AI, licenses, and data centers. According to the State of FinOps 2026 (1,192 respondents, with a combined cloud budget of over $83 billion), 98% of FinOps teams manage AI costs, 90% manage SaaS, and 64% manage licenses. 78% report to the CTO or CIO. FinOps for AI is the top priority for 2026.

Maturity: Organizations grow from crawling to walking to running. Start with insight and ownership; perfection comes later.

Cost allocation starts with tagging: each resource is assigned tags such as team, application, environment, and cost center. Define required tags in policies and enforce them in code using AWS Tag Policies and Service Control Policies, Azure Policy, or Google Organization Policies, along with checks in your Terraform pipeline.

• Showback: Teams see how much they cost, but they don't pay for them themselves. A good first step.
• Chargeback: Costs are charged to the team’s budget. This fosters a stronger sense of ownership but requires reliable allocation.
• Shared Costs (network, support, shared clusters) are allocated according to a fixed formula, such as based on usage or revenue.

FOCUS.

The FinOps Open Cost and Usage Specification is an open standard for cost and usage data, currently at version 1.4. AWS, Microsoft, Google, and Oracle provide cost data in this format. European providers such as OVHcloud and STACKIT, as well as SaaS providers such as Snowflake and Databricks, are also on the FinOps Foundation’s list. This allows you to compare costs across multiple clouds and services without having to convert the data yourself.

Costs that you avoid during the design phase won’t need to be cut later on. That’s why mature organizations shift their focus on cost awareness to the design phase:

• Cost as a non-functional requirement in architectural decisions, alongside performance and security.
• Cost estimates in the pipeline: Tools like Infracost show the monthly cost of every Terraform change.
• Reference architectures with cost-effective standard choices, such as ARM, gateway endpoints, and lifecycle rules.
• The "Cost Optimization" pillar of the AWS Well-Architected Framework—or the comparable guidelines from Azure and Google—as a standard benchmark.

In Kubernetes, applications share the same nodes, so the cloud bill doesn't show which team or application is responsible for which costs. Utilization is also low. We often see an average CPU usage of 8% and memory usage of 20%.

What helps:

• Allocate costs by namespace or label using OpenCost (open source, CNCF) or a commercial tool.
• Realistic requests and limits based on measured usage, for example, using the Vertical Pod Autoscaler in recommendation mode. Overly generous requests are the biggest source of waste, because the scheduler reserves capacity that is never used.
• Bin packing and node autoscaling using Karpenter or the cluster autoscaler, so that nodes run at full capacity and idle nodes are removed.
• Spot nodes for stateless and interruptible workloads, with pod disruption budgets.
• Fewer isolated clusters: every control plane and every set of system components costs money. See the guide on Kubernetes.

AI adds two new cost items: the use of language models (per token) and GPU capacity.

Reducing token costs:

• Allocate costs by application and by team, using separate API keys or projects. Without allocation, no one knows how much each application costs.
• Choose the smallest model that's good enough, and route simple queries to a less expensive model.
• Use prompt caching for recurring contexts. At Anthropic, a cache read costs 10% of the normal input price; other providers offer similar discounts.
• Use batch processing for tasks that don't need to be completed immediately: many providers offer a 50% discount.
• Measure costs per outcome— for example, per completed customer interaction—and compare them to the value.

Reducing GPU costs:

• Measure the utilization. Average GPU utilization is 5%.
• Disable GPU capacity when it is not in use, and share GPUs across workloads whenever possible.
• Reserve this term only for demonstrably consistent inference.
• Consider an on-premises solution or a European provider for stable, large-scale AI workloads, especially if the data cannot leave the country.

Licenses can represent a larger expense than the underlying infrastructure. Windows Server and SQL Server are billed per vCPU in the cloud; with the Azure Hybrid Benefit or your own licenses (BYOL), you only pay for them once. Even smaller instances with the same performance provide immediate savings on per-core licensing. The same principle applies to SaaS as to the cloud: gain insight into usage per user, cancel unused licenses, and align contracts with actual usage. 90% of FinOps teams now include SaaS in their management.