In brief
Cloud cost optimization is the practice of matching what you pay for cloud infrastructure to the value it produces. Most companies can cut 20–35% of their AWS, Azure or Google Cloud bill within six months, without slowing delivery, by doing four things in order: make every cost visible and owned, remove waste, commit to the steady baseline at a discount, and fix the architecture that drives cost per customer. The goal is not the smallest bill. It is a cost per customer, per transaction or per AI request that falls as the business grows.
Every growing company reaches the same meeting. The finance lead puts the cloud bill on the screen, points at a line that has doubled in eighteen months, and asks a simple question: is that growth, or is that waste? Engineering usually cannot answer with confidence, because nobody designed the platform to answer it.
That gap is what this playbook closes. It is written for CTOs, CFOs, founders and engineering leaders who need to control cloud spend without turning engineers into accountants or freezing the roadmap. You will find the metrics that matter at board level, the ten levers ranked by savings and effort, real use cases by company stage, a worked example with numbers, and a 90-day plan you can start on Monday.
- ~27% - of cloud spend is wasted, by organisations’ own estimate (Flexera State of the Cloud)
- $723B - forecast worldwide public cloud spending in 2025 (Gartner)
- Up to 72% - discount on steady compute with Savings Plans or reserved capacity
- Up to 90% - discount with Spot capacity for interruptible workloads
Key takeaways
- Measure cost per customer or per transaction, not just the monthly total. A rising bill is fine if unit cost is falling.
- Visibility comes first: aim for over 90% of spend tagged to a team, product or customer before you optimise.
- Quick wins (idle resources, non-production schedules, storage clean-up) typically return 10–15% in the first month.
- Commit to your steady baseline only after rightsizing, or you lock in waste at a discount.
- The largest long-term savings come from architecture: data transfer, managed service choices, autoscaling and AI inference design.
- Make cost a shared, weekly engineering metric with an owner, not a quarterly finance surprise.
Why Cloud Bills Grow Faster Than Revenue
The cloud makes it easy to start and easy to forget. A developer launches a large database for a load test and never shrinks it. A staging environment runs all weekend. Logs are kept forever because nobody decided otherwise. Each choice is small and reasonable on the day. Together they become a bill that grows faster than customers do.
Across reviews, the same six causes explain most overspend:
📈 Overprovisioned compute
- Servers, containers and databases sized for a peak that rarely comes. Many workloads use less than half of what they reserve.
🌙 Idle non-production
- Development, test and staging environments running 24/7 when people use them about 50 hours a week.
💸 Paying on-demand prices
- A steady baseline billed at full price because nobody owns commitment purchases.
🔁 Data transfer and egress
- Traffic between zones, regions and the internet, often invisible until it is a top-five line item.
🗄️ Forgotten storage
- Old snapshots, unattached volumes, unused buckets and logs kept forever on premium tiers.
🧩 Service sprawl
- Overlapping monitoring, managed services and SaaS tools each billed separately, with no single owner.
None of these are engineering failures. They are ownership failures. When cost is nobody’s metric, it drifts. That is why the first fix is not a tool or a discount, it is visibility and accountability, a theme covered in more depth in Cloud Maturity Is the New Digital Transformation.
The Cloud Cost Metrics Decision Makers Should Track
A monthly total tells you what happened, not whether it was good. These eight metrics tell leadership whether cloud spend is under control and whether it scales with the business. Put them on one page and review them monthly.
MetricWhat it tells youHealthy targetUnit cost (cost per customer, per 1,000 transactions or per AI request)Whether the platform gets cheaper to run as you growFlat or falling while usage growsCloud cost as a share of revenueImpact on gross margin, especially for SaaSTypically 5–12% of revenue for mature SaaS; trending downCost allocation coverageHow much spend has a clear owner (tagged to team, product or customer)Above 90%Commitment coverageHow much steady usage is on discounted Savings Plans or reservations70–90% of the steady baselineCommitment utilisationWhether you are using what you committed toAbove 95%Idle and waste rateSpend on resources doing nothing usefulBelow 5–10%Forecast accuracyHow well teams predict next month’s spendWithin ±5–10%Anomaly response timeHow quickly an unexpected spike is noticed and fixedUnder 24–48 hours
The one number for the board
If you only report one metric, report unit cost. “Cloud cost per active customer fell from $14.20 to $10.90 while customers grew 40%” tells a far better story than “the bill went up 8%”.
The 10 Cloud Cost Optimization Levers, Ranked
Not every saving is worth the effort. The table ranks the ten most effective levers by typical savings on the affected spend, the effort involved and how quickly they pay back. Work from the top down.
LeverTypical saving on affected spendEffortTime to value1. Delete idle and orphaned resources100% of that wasteLowDays2. Schedule non-production environments60–70%Low1–2 weeks3. Rightsize compute and databases20–40%Medium2–6 weeks4. Storage lifecycle and tiering30–70% on cold dataLow1–3 weeks5. Savings Plans and reserved capacityUp to 72%Low (after rightsizing)Immediate6. Spot or preemptible capacityUp to 90%Medium2–4 weeks7. Modern processors (e.g. AWS Graviton)Up to 40% better price-performanceMedium1–3 months8. Autoscaling and Kubernetes bin-packing20–50% of cluster costMedium–high1–3 months9. Reduce data transfer and egress20–60% of transfer costHigh1–6 months10. Re-architect cost hotspotsVaries; often the largest long-term gainHigh3–12 months
Notice the order. Commitments (lever 5) save the most per hour of effort, but only after rightsizing (lever 3). Commit first and you lock in today’s waste for one to three years.
Cloud Cost Optimization Use Cases by Company Stage
The right priorities depend on where you are. A seed-stage startup and a public enterprise both overspend, but for different reasons.
🚀 Early-stage startup
- Spend: usually under $20k a month.
- Focus: free tiers and credits, managed services, scheduling, a monthly budget alert.
- Avoid: building a FinOps team. One owner and an hour a week is enough.
📊 Scaling SaaS
- Spend: $20k–$500k a month.
- Focus: tagging, cost per customer, rightsizing, Savings Plans, Kubernetes efficiency.
- Win: gross margin improves before the next funding round or audit.
🏢 Enterprise
- Spend: $500k+ a month across many teams and accounts.
- Focus: a FinOps operating model, showback or chargeback, commitment portfolio, governance.
- Win: predictable budgets and cost decisions made inside each team.
🤖 AI and GPU-heavy products
- Spend: dominated by model inference, GPUs and vector databases.
- Focus: cost per AI request, model routing, caching, batching, right-sized models.
- Win: AI features that stay profitable as usage grows.
Example: a SaaS company preparing for due diligence
A B2B SaaS company with steady growth wanted to show investors a healthier gross margin. The bill was not out of control, but nobody could explain cost per customer. Tagging and a simple unit-cost dashboard took three weeks. That alone revealed that two large enterprise customers on a legacy deployment model cost four times more to serve than everyone else, which turned a cost problem into a pricing and architecture decision.
Example: an AI feature that got popular
A product team launched an AI assistant using the largest available model for every request. Usage tripled in a quarter and inference became the biggest line on the bill. Routing simple requests to a smaller model, caching repeated answers and trimming prompts cut cost per request by more than half without users noticing a quality drop. The lesson: AI cost is an architecture decision, made in the first design, not a billing problem fixed later. See how to move AI pilots into production for the wider production checklist.
Managing AI and GPU Costs in 2026
AI workloads behave differently from web applications. Costs scale with every request, GPUs are expensive to leave idle, and model choice can change cost per request by ten times or more. The levers that matter most:
- 🧭 Route by task: small models for simple requests
- ♻️ Cache repeated prompts and answers
- ✂️ Trim prompts and retrieved context
- 📦 Batch non-urgent jobs
- ⏱️ Scale GPUs to zero when idle
- 📏 Track cost per AI request and per feature
Before funding a large AI programme, model the cost per request at expected volume alongside the benefit. The same discipline is covered in how to evaluate the ROI of AI automation before investing.
A 90-Day Cloud Cost Optimization Plan
A cost programme fails when it becomes a one-off clean-up. This plan delivers savings early, then builds the habits that keep the bill under control.
Worked Example: Cutting a $90,000 Monthly Cloud Bill
Here is how the plan typically adds up for a mid-sized SaaS platform spending $90,000 a month on AWS, with 1,200 customers. The figures are illustrative, but each step reflects savings commonly seen in practice.
That is $360,000 a year back in the business, roughly the cost of two senior engineers, and gross margin improves by several points. The next round of savings comes from the roadmap: reducing cross-zone data transfer and moving a batch workload to Spot capacity.
Cloud cost review
Want to know where your cloud spend is going?
An independent architecture and cost review maps your spend to teams and features, ranks the savings by effort and gives you a 90-day plan your engineers can deliver without slowing the roadmap.
Cloud Cost Management Tools: What You Actually Need
Tools help, but they do not replace ownership. Start with what your cloud provider already gives you and add a specialist tool when multiple clouds, Kubernetes or unit economics outgrow it.
CategoryExamplesBest forNative cloud toolsAWS Cost Explorer, AWS Budgets, AWS Compute Optimizer, Azure Cost Management, Google Cloud Billing reportsEvery company. Free, accurate and the right place to start.Cost management platformsCloudZero, Vantage, Finout, Apptio Cloudability, Flexera OneMulti-cloud spend, unit economics, showback and allocation across many teams.Kubernetes cost toolsOpenCost, KubecostCost per namespace, workload and team inside shared clusters.Automation and commitmentsAWS Instance Scheduler, ProsperOps, Spot by NetAppAutomating schedules, commitment purchases and Spot capacity.Billing data standardFOCUS (FinOps Open Cost and Usage Specification)Combining bills from several clouds and SaaS vendors in one format.
Most of these follow the FinOps Framework from the FinOps Foundation, which organises the work into three repeating phases: Inform, Optimize and Operate.
Who Owns Cloud Cost: The FinOps Operating Model
FinOps works when engineering, finance and product share one set of numbers and each has a clear role. It does not require a large team. In most companies under $1M a month, one part-time FinOps lead and a monthly review are enough.
RoleResponsibilityExecutive sponsor (CTO or CFO)Sets targets such as unit cost and cost-to-revenue, and removes blockersFinOps leadRuns reporting, commitments, the tagging standard and the monthly reviewEngineering teamsOwn their budgets, act on recommendations and design for costFinanceForecasting, budgeting and allocating cost to products and customersProductWeighs cost per feature or customer against its value and pricing
Signs it is working
- Teams can answer “what does this feature cost to run?” in minutes
- Unit cost falls quarter after quarter
- Spikes are caught within a day, not on the invoice
- Cost is discussed in design reviews, not only in finance meetings
Signs it is not
- Savings come from one big clean-up, then the bill creeps back
- Engineers learn about cost from finance emails
- Large commitments bought without rightsizing first
- Nobody can say which customers are profitable to serve
Common Cloud Cost Optimization Mistakes
- **Cutting cost at the expense of reliability. **Removing redundancy or backups to save money is a false economy. Optimise waste, not resilience.
- **Buying three-year commitments too early. **Lock-in feels like savings until the architecture changes. Start with one-year terms and flexible plans.
- **Optimising the total instead of unit cost. **A falling bill during a growth phase can mean lost customers, not efficiency.
- **Treating it as a project. **Without weekly habits and owners, waste returns within two or three quarters.
- **Ignoring data transfer. **It rarely shows up in early reviews and often becomes one of the biggest lines as traffic grows.
- **Leaving it all to tools. **Recommendations are only useful if a team owns and acts on them.
When Cost Is Really an Architecture Problem
If you have done the quick wins, rightsized and committed, and unit cost still rises with every customer, the problem is no longer housekeeping. It is design: a chatty service architecture that moves data across zones, a database model that cannot scale horizontally, or a legacy system that needs dedicated infrastructure per customer. At that point an independent software architecture review is usually the fastest way to find the few design decisions that drive most of the cost, and application modernization turns them into a staged plan.
The companies that win treat cloud cost the way they treat security and reliability: as a quality of the system, designed in from the start and measured every week. Do that, and the question in the finance meeting changes from “why is the bill so high?” to “where should we invest the savings?”.
Originally published at phpscientist.com.
