AI Cost Allocation Intermediate

Cost allocation is the foundation of FinOps. Without knowing who is spending how much on what, optimization is impossible. This lesson covers implementing tagging strategies for AI resources, choosing between chargeback and showback models, attributing GPU costs accurately, and handling shared infrastructure costs.

Tagging Strategy for AI Resources

Every AI resource should have these mandatory tags:

Chargeback vs Showback

• Showback - Show teams their costs without billing them. Good for starting FinOps culture; low friction, high visibility
• Chargeback - Actually charge team budgets for their AI resource usage. Stronger incentive to optimize, but requires mature cost allocation
• Hybrid - Showback for shared infrastructure (monitoring, control plane), chargeback for direct GPU usage

GPU Cost Attribution

GPU cost attribution requires correlating GPU usage metrics with Kubernetes pod ownership:

• Namespace-based - Each team gets a namespace; all GPU costs in that namespace are attributed to the team
• Label-based - Use Kubernetes labels to attribute GPU pods to teams and projects
• Time-weighted - For shared GPUs (MIG, time-slicing), attribute costs based on actual usage time

Handling Shared Costs

Some infrastructure costs cannot be directly attributed to a single team:

• Platform overhead - Kubernetes control plane, monitoring stack, GitOps controllers; split proportionally by team GPU usage
• Shared storage - Common datasets and model registries; split by storage consumption or access frequency
• Networking - Shared InfiniBand fabric; split proportionally by data transfer volume