Rightsizing — IT definition

Rightsizing is the FinOps practice of continuously adjusting IT resource capacity to actual measured need, to eliminate over-provisioning without degrading performance. It is one of the most immediate and universal optimization levers: per Flexera, 35 % on average of an enterprise's cloud compute is over-provisioned, which represents hundreds of thousands of euros wasted annually per million euros spent.

Initially applied to cloud (VMs too large relative to actual CPU/RAM usage), rightsizing now extends to software licenses, SaaS subscriptions, storage, containers, and even AI model allocation.

The principle

Rightsizing rests on three elements:

• •Actual usage measurement: fine-grained monitoring of effective consumption (CPU, RAM, IOPS, active seats, API requests).
• •Comparison with provisioned capacity: "I paid for X, I use Y".
• •Adjustment recommendation: move to lower (or higher) capacity matched to usage profile.

Classic trade-off: tolerate a 20-30 % safety margin to absorb peaks, but no more.

Cloud rightsizing

In public cloud, rightsizing applies to:

• •Compute: move from m5.4xlarge to m5.2xlarge if CPU < 30 %.
• •Storage: shift from SSD gp3 to st1 for low-traffic volumes.
• •Managed databases: downgrade RDS, Cosmos DB, BigQuery slots.
• •Kubernetes: adjust pod requests/limits based on metrics.
• •Lambda / Cloud Functions: tune allocated memory (impacts CPU and cost).

Tooling: AWS Compute Optimizer, Azure Advisor, GCP Recommender, Cloudability, Vantage, Kubecost (Kubernetes).

SaaS and license rightsizing

Rightsizing extends to software:

• •Active seats: paying for 100 Salesforce licenses while only 60 are used → cancel 40 seats.
• •Tiers: user on Enterprise tier (€50/month) only using Business-tier features (€15) → downgrade.
• •Personal subscriptions: teams using individual accounts where an enterprise account would be cheaper (and safer).
• •Commitment reservations: annual adjustment of Reserved Instances or Savings Plans based on real usage.

Kabeen provides application and SaaS rightsizing by cross-referencing purchased licenses, assigned seats, and real usage measured by agents and SSO.

AI rightsizing

New 2025-2026 perimeter: tuning LLM models to business needs.

• •Model too powerful: using GPT-4 for simple tasks a smaller model would handle at 10× lower cost.
• •Context too wide: paying for 200k token contexts when 8k would suffice.
• •Useless inferences: calling the LLM on every keystroke when caching would solve it.

AI optimization is going to become one of the major FinOps topics in coming years.

Barriers to rightsizing

Rightsizing is conceptually simple but stays underused:

• •Performance fear: "what if we downgrade and production crashes?" — answer: testing, safety margin, monitoring.
• •Action cost: downgrading requires downtime or a migration operation.
• •Governance: who decides? Who executes? Who owns if it breaks?
• •Lack of data: without fine-grained monitoring, impossible to measure over-provisioning.
• •Cumulative effect: 35 % savings across 1,000 resources is massive. On one resource, it's invisible.

Hence the importance of industrializing via a FinOps or ITAM platform.

Rightsizing vs autoscaling vs scaling

Three complementary approaches:

• •Rightsizing: static sizing adjusted periodically (monthly).
• •Autoscaling: automatic real-time adjustment to load.
• •Vertical vs horizontal scaling: instance size vs number of instances.

A mature strategy combines all three: rightsizing for the baseline, autoscaling for peaks, vertical/horizontal scaling depending on application architecture.

Measurable benefits

• •20 to 35 % cloud savings: in year one.
• •15 to 25 % SaaS license savings: in year one.
• •Reduced carbon footprint: less compute = less energy.
• •FinOps maturity: structuring practice that anchors the allocation and measurement discipline.