Cloud costs sprawl for predictable reasons: resources provisioned for peak and never resized, environments left running after the project ends, storage that never ages out, and a single invoice nobody can attribute to a team. Cloud cost optimization is the practice of correcting that systematically — and at Archonova it is part of how we architect and operate cloud, not a separate product.
Core practices
These are the established, vendor-neutral practices behind most cloud cost optimization work. Savings vary entirely by estate — anyone quoting a fixed percentage before looking at your bill is guessing.
Instance families and sizes are usually chosen before real load data exists, then never revisited. Rightsizing compares actual CPU, memory, network and IOPS utilisation against the provisioned shape and moves workloads to the smallest instance that still meets headroom and burst requirements. Modern generations of the same family are often cheaper and faster, so a generation upgrade is frequently the first saving available.
On-demand pricing buys flexibility you may not need. Where a baseline of compute runs continuously, one- or three-year commitments (reserved instances, savings plans, committed use discounts) trade that flexibility for a materially lower rate. The discipline is to commit only to the stable baseline, keep the variable layer on-demand, and track coverage and utilisation so commitments do not go unused.
Spare hyperscaler capacity is sold at a steep discount on the condition it can be reclaimed with little notice. That suits fault-tolerant, interruptible work — batch processing, CI runners, rendering, stateless queue consumers, many data-pipeline stages. It requires checkpointing, retry logic and diversified instance pools, and it should never carry a workload that cannot survive eviction.
Object and block storage costs are driven by both capacity and access pattern. Hot tiers price for frequent reads; cool and archive tiers cost far less per GB but add retrieval latency and per-request fees. Lifecycle policies that age data automatically from hot to cool to archive — matched to genuine access frequency and retention obligations — are one of the least disruptive optimisations available.
Cloud estates accumulate spend with no owner: volumes left unattached after an instance is terminated, idle load balancers, unassociated public IPs, forgotten snapshots and stale AMIs, non-production environments running through nights and weekends, dangling NAT gateways. None of it appears in a design document; all of it bills monthly. Periodic sweeps plus automated non-production schedules remove it.
You cannot optimise what nobody owns. A consistent tagging taxonomy — environment, application, team, cost centre — turns one aggregate invoice into per-team, per-service accountability, and makes anomaly detection and budget alerts meaningful. Tag policies enforced at provisioning time work; retrospective tagging clean-ups rarely hold.
Where a workload is genuinely portable, comparable compute, egress and storage rates differ between hyperscalers, and between hyperscaler and private or colocated infrastructure. This matters most for steady-state, high-egress or data-heavy workloads. It is worth modelling honestly — including migration effort and operational overhead — rather than assuming either that multi-cloud always saves money or that it never does.
How we approach this
Cost work that lands as a spreadsheet of recommendations tends to decay within a quarter. We treat it as part of the same Cloud Solutions engagement that designs and runs the platform, so decisions about instance shape, commitment coverage and lifecycle policy sit with the people accountable for reliability.
Pull billing and utilisation data, map spend to workloads and owners, and identify what is genuinely steady-state versus variable.
Orphaned resources, non-production schedules and storage lifecycle rules land in days. Rearchitecting for spot or a different data tier is a project — we scope them differently.
Reservations and savings plans are sized against the demonstrated baseline, with coverage and utilisation tracked so nothing is bought and forgotten.
Tag policy at provisioning time, budgets and anomaly alerts, and a recurring review in the operations cadence rather than an annual scramble.
Overview
Three causes account for most avoidable cloud spend. Each has a different fix, and confusing them is why cost programmes stall.
Capacity sized for a launch-day peak, or copied from an on-premise spec, keeps billing at that level long after real load is known.
Detached volumes, dormant environments, unused load balancers and old snapshots quietly accrue charges with no owner and no purpose.
Without tagging and allocation, spend is one aggregate number. Nobody can see their own consumption, so nobody changes behaviour.
Match instance family, size and generation to observed utilisation, with deliberate headroom rather than guesswork.
Reserved instances and savings plans priced against the stable baseline, tracked for utilisation so discounts are actually realised.
Hot, cool and archive tiers applied by real access pattern and retention rules, automated through lifecycle policies.
Process
Evidence first, then changes that are safe to make, then the controls that stop the drift returning.
Billing exports, utilisation metrics and resource inventory — the ground truth behind the invoice.
Rank actions by saving against effort and risk, separating same-week cleanup from architectural change.
Rightsizing, lifecycle policies, schedules, spot adoption and commitment purchases — executed with rollback plans.
Tag enforcement, budgets, anomaly alerts and a standing review so cost stays a tracked operational metric.
Tech stack
Native provider tooling first, augmented where an estate needs cross-cloud visibility.
Industries served
Data-heavy, seasonal and regulated estates feel cloud waste first.
Share your current environment and we'll walk through where the spend is going and which levers apply — before any commitment.
