The cloud bill arrives every month. And every month it is higher than expected. A study by Flexera (State of the Cloud Report 2026) shows: 67% of companies exceed their cloud budget. By an average of 30%. For companies with AI workloads, by as much as 45%.
The problem is not the cloud itself. It is the lack of transparency and governance.
Why Costs Are Exploding
- AI Compute: GPU instances for inference and training are 10-50x more expensive than standard VMs. A single GPT-4-based Copilot can generate 5,000 EUR/month in API costs at high usage. Multiplied by 200 employees: 1 million EUR/year.
- Zombie Resources: VMs running that nobody uses. Snapshots nobody deletes. Test environments that keep running after the project ends. According to Flexera: 27% of all cloud spending is waste.
- Overprovisioning: VMs with 8 vCPUs and 32 GB RAM running at 5% utilization. Because the developer chose "one size bigger just to be safe."
- Data Egress: Uploading data to the cloud is free. Downloading data costs money. For data-intensive applications (analytics, AI training, backup), egress can account for 10-20% of the total bill.
- Unused Reservations: Companies buy Reserved Instances or Savings Plans but do not fully utilize them. 35% of all reservations are underutilized according to AWS data.
FinOps as a Discipline
FinOps (Financial Operations) is not a tool but a practice: Engineering, finance, and business work together to understand, optimize, and forecast cloud costs. The FinOps Foundation (part of the Linux Foundation) has over 10,000 certified practitioners worldwide in 2026.
The three phases:
- Inform: Create transparency. Who is consuming what? Which team? Which project? Enforce tagging standards. Build dashboards.
- Optimize: Right-sizing (downsizing VMs), purchasing reservations, using spot instances, eliminating zombie resources, configuring auto-scaling.
- Operate: Continuous process. Monthly cost reviews, budget alerts, governance policies, showback/chargeback to teams.
AI Workloads: The New FinOps Challenge
AI workloads present new problems for FinOps: Costs are difficult to predict (dependent on prompt length, token consumption, model choice), usage varies widely, and ROI measurement is harder than with traditional workloads.
Best practices for AI FinOps:
- Define token budgets per team/project
- Model routing: Route simple tasks to cheaper models (GPT-4o-mini instead of GPT-5.4)
- Use prompt caching (saves 50-80% on repeated queries)
- Use batch APIs instead of real-time APIs where possible (50% cheaper)
- Local models for standardized tasks (Ollama, see our separate article)
Cloud costs are the third-largest IT cost block in 2026, after personnel and licenses. Those who do not actively manage them are burning money. FinOps is no longer optional.
Sources: Flexera State of the Cloud Report 2026, FinOps Foundation Annual Report, AWS Cost Explorer Benchmark Data, Gartner Cloud Cost Management MQ.