Visibility built, now the levers. They come in two families that teams persistently confuse: commercial levers that lower the price of what you run, and engineering levers that reduce what you run. Both matter, and the ordering rule is simple: never commit commercially to waste, so rightsizing evidence comes before reservation purchases.

The Commitment Portfolio

Reservations discount specific resource footprints, a VM size family in a region, a database tier, storage capacity, deepest discounts (up to the seventy percent range over three years) in exchange for specificity, softened by instance size flexibility within the family and exchanges within the rules. Savings plans for compute discount a committed hourly spend across almost all compute anywhere, shallower discounts for far more freedom, the workload can change size, family, region, even service, and the plan follows. The decision is a portfolio, not a choice: your stable, known floor (the domain controllers, the production database tier, the baseline AKS node count from months of the metrics yesterday built) goes into reservations; the flexible middle of compute that exists but shifts shape goes into a savings plan; the spiky top stays pay as you go or goes to spot. Buy in layers quarterly rather than one annual big bang, watch utilization of every commitment (an unused reservation is a subscription to nothing), and centralize purchasing with shared scope so finance sees one portfolio while amortized cost allocates it fairly, yesterday’s plumbing paying off.

Azure Hybrid Benefit stacks on top for Windows Server and SQL Server: bring licenses with Software Assurance and the license component of the price disappears, roughly forty percent on Windows VMs and more on SQL. It applies to the Managed Instance and SQL VM estate from the earlier posts, and the audit habit is checking that eligible resources actually have it enabled, the number of estates paying twice for licenses they own is embarrassing.

Engineering Levers, In Order of Leverage

Delete: orphaned disks, unattached public IPs, stopped but allocated VMs, empty App Service plans, the aged snapshots the Automation post’s runbooks were built for; every estate’s first sweep finds real money. Schedule: dev and test compute off nights and weekends is a seventy percent reduction on those environments, and the runbook already exists. Rightsize with evidence: Advisor’s recommendations seeded by actual utilization, validated against p95 not average, and executed through the IaC pipeline so the change persists. Rearchitect where the pattern fits: the VMSS spot mixing and Container Apps scale to zero posts are cost posts wearing architecture costumes, and the biggest single line items usually fall to design changes, the Log Analytics table plans and ingestion transforms, Cosmos autoscale versus manual math, storage lifecycle tiering, each covered in its September post and each worth more than a discount. The meta lever is making cost a nonfunctional requirement in design reviews, one line in the template, expected monthly cost and unit metric, so the bill is designed rather than discovered.

The Operating Rhythm

Monthly: commitment utilization and coverage review, anomaly postmortems, the unallocated bucket trend. Quarterly: portfolio rebalance with exchanges where drift demands, the deletion sweep, and one deliberate optimization project chosen by expected value. The maturity marker is when coverage of the stable floor sits above eighty percent, utilization of commitments above ninety five, and the conversation shifts from “why is the bill this big” to “is this unit cost acceptable”, which is the question a business can actually answer.

Cheers
Osama