FinOps

What is FinOps?

FinOps is the operating model for cloud cost: treating spend as a first-class engineering signal with owners, budgets, and feedback loops, rather than a surprise invoice finance escalates quarterly. The cultural core is moving cost accountability to the teams who create the cost — the people who chose the instance type, left the NAT gateway chatty, or forgot the dev cluster — because only they can fix it. The FinOps Foundation frames the loop as inform → optimize → operate, run continuously.

The mechanics that actually move bills

Visibility first: cost allocation tags/labels enforced at provision time (untagged = unattributable = unfixable), per-team dashboards, and anomaly alerts that fire in days, not at month-end. The big three levers, in typical order of yield: rightsizing (the fleet is provisioned for a peak that never came), commitment discounts (Savings Plans/Reserved capacity for the stable baseline — 30–60% off compute you were buying anyway), and waste deletion (idle load balancers, unattached volumes, forgotten environments — every audit finds them). Unit economics is the maturity marker: cost per request, per customer, per token — absolute spend growing while unit cost falls is success; the reverse is the real alarm.

The 2026 wrinkle: AI spend

Inference costs broke traditional FinOps assumptions — spend scales per-call, varies 100× by model choice, and sits in engineers’ prompt decisions. Token-level attribution, model-routing policies, and batch/caching discounts are now standard FinOps surface area.

What people get wrong

• FinOps as a procurement person negotiating discounts — without engineering ownership, nothing structural changes.
• Optimizing before measuring: untagged spend makes every initiative guesswork.
• Savings-plan lock-in on unstable workloads — commit to the floor, not the forecast.

Primary source: FinOps Foundation — FinOps Framework