Cloud Cost Optimisation in 2026: FinOps Moves That Actually Cut the Bill
By Stackvora Team · 2026-09-07 · 9 min read
Cloud bills grew a new line item in the last two years: inference. Between GPU hours, vector databases and data movement, the old cost playbook — right-size instances, buy reservations — no longer covers the largest growth areas. Here is where spend actually accumulates in 2026 and what reliably reduces it.
Find the money before you optimise
Most organisations cannot answer "what does this service cost per month?" Tagging is the unglamorous prerequisite. Enforce a small mandatory tag set — owner, service, environment, cost-centre — through policy at resource-creation time, not through cleanup scripts. Then publish per-team showback. Visibility alone routinely removes 10–15% of spend because teams delete things they forgot they owned.
The five biggest line items
| Category | Typical waste | First move |
|---|---|---|
| Compute | Over-provisioned, always-on non-prod | Schedule non-prod off outside work hours |
| Storage | Old snapshots, un-tiered object data | Lifecycle rules to infrequent/archive tiers |
| Data transfer | Cross-AZ and egress chatter | Co-locate chatty services, add caching/CDN |
| Databases | Sized for peak, idle 80% of the time | Serverless or autoscaling tiers |
| AI/GPU | Idle reserved GPUs, oversized models | Batch queues, smaller models, caching |
Non-production is the easiest win
Development and staging environments typically run 168 hours a week to serve maybe 45 hours of use. A scheduler that stops non-prod compute at night and weekends cuts those environments by roughly 60–70% with no architectural change. Pair it with ephemeral environments that expire after a pull request merges.
Commitments after right-sizing, never before
Reserved instances and savings plans lock in a discount on a baseline. If you commit before right-sizing, you have locked in your waste for one to three years. Sequence it: measure, right-size, stabilise for a month, then commit to the floor — not the average. Keep 20–30% of steady-state on-demand so you retain the freedom to re-architect.
Egress and cross-zone traffic
Data transfer is the cost that hides inside architecture diagrams. A microservice mesh spread across availability zones can spend more on cross-AZ traffic than on the compute doing the work. Practical fixes: zone-aware routing so callers prefer same-zone endpoints, a CDN in front of anything public, compression on internal APIs, and keeping analytics pipelines in the same region as their source data.
The new one: AI spend
Inference cost behaves differently from infrastructure cost — it scales with user behaviour, not with provisioned capacity. Controls that work:
- Model routing. Use a small model for classification and extraction; reserve large reasoning models for the minority of hard requests.
- Prompt caching. Stable system prompts and repeated context should not be re-billed every call.
- Output limits. Cap max tokens; verbose answers are pure cost with no quality gain.
- Batch what is not interactive. Offline jobs belong on discounted batch tiers.
- Per-tenant budgets. Meter usage per customer so one heavy account cannot erase your margin.
Kubernetes specifics
Cluster bills are usually a bin-packing problem. Requests set far above real usage leave nodes half empty while the autoscaler adds more. Use actual usage data to set requests, enable a bin-packing-aware autoscaler, and adopt spot capacity for stateless and batch workloads with a small on-demand floor for stability.
Make savings stick
Reductions decay. Put cost into the same review loop as reliability: a monthly FinOps review with per-team numbers, anomaly alerts on daily spend, and a budget check in the infrastructure pipeline that flags a pull request adding an expensive resource type. A one-off optimisation project saves money for a quarter; a feedback loop saves it permanently.
Key takeaways
- Enforce tagging first — you cannot reduce what you cannot attribute.
- Shut down non-prod outside working hours before anything more complex.
- Right-size, then commit to the floor, not the average.
- Cross-AZ and egress traffic is an architecture cost, not a billing accident.
- AI spend needs routing, caching, output caps and per-tenant budgets.
Want a review of your own cost profile? Describe your stack to Stackvora AI, or read the cloud guides in the knowledge base.