I’ve sat in too many cloud cost reviews where the team is celebrating a 3% saving from rightsizing while a $200K/month workload runs on the wrong service tier. The optimization conversation has matured but the prioritisation is still mostly wrong.

Here’s what I see actually moves the needle, having been through this with several mid-to-large enterprise environments.

Where the real money is

Cloud spend at most enterprises follows a familiar pareto distribution. A small number of workloads account for most of the cost. Optimization effort that doesn’t address those workloads is largely cosmetic.

The biggest line items at most enterprises:

Database services. RDS, Cloud SQL, Cosmos DB, Aurora, etc. These are typically the most expensive services per workload because of the combination of compute, storage, and IO charges. The workloads using them are also typically business-critical, so changes need to be careful.

Container orchestration and compute clusters. EKS, GKE, AKS clusters running production workloads. The cluster nodes themselves often run inefficiently — wrong instance types, poor bin-packing, autoscaling configured pessimistically.

Data transfer and egress. Underestimated by many teams. Cross-AZ traffic, internet egress, and inter-service traffic in microservices architectures can be a major hidden cost.

Storage tiering misconfiguration. Hot storage holding rarely-accessed data is one of the most common waste sources. Lifecycle policies exist but are often not configured aggressively enough.

Reserved capacity vs on-demand mismatch. Enterprises that haven’t actively managed their reserved instance portfolio are usually overpaying. The right reserved capacity strategy can save 25-40% on long-running workloads.

What’s overhyped

A few categories get more optimization attention than they deserve:

Aggressive instance rightsizing. The savings from rightsizing individual instances are real but small relative to the engineering cost. Spending six weeks moving every instance to slightly smaller types might save $5K/month, while a tier change on the database could save $20K/month.

Spot/preemptible obsession. Spot instances are valuable for the right workloads but they’re not universally applicable. Trying to push spot adoption beyond stateless, fault-tolerant workloads usually creates more operational pain than the savings justify.

Single-cloud cost dashboards. The cost dashboard market is crowded. Most enterprises have plenty of cost visibility. The bottleneck isn’t visibility — it’s prioritisation and execution. More dashboards don’t deliver more savings.

Multi-cloud arbitrage. Theoretically interesting, practically rare. The cases where moving workloads between clouds saves real money are usually offset by the operational complexity of managing two cloud environments. The exceptions are specific (specific compute types, specific egress patterns) and don’t justify a multi-cloud strategy.

Where the operational cost actually hides

Beyond the direct cloud bill, several costs ride alongside cloud spend that often aren’t counted:

Engineering time on cost-related changes. Every architectural change made for cost reasons consumes engineering capacity. The opportunity cost of not delivering features needs to be in the equation.

Vendor lock-in technical debt. Heavy use of cloud-specific services makes future moves expensive. This isn’t a current cash cost but it’s a real future cost that should factor into architectural decisions.

Reserved capacity utilisation risk. Reservations save money when fully utilised but waste money when underutilised. Aggressive reservation strategies that don’t account for workload changes can backfire.

Data sovereignty and compliance overhead. Sometimes the cheapest cloud configuration violates data sovereignty or compliance requirements. Discovering this in audit is expensive.

The pattern that works

Enterprises getting cloud cost discipline right share some patterns:

1. Quarterly cost reviews led by engineering, not procurement. Procurement-led reviews focus on contract terms. Engineering-led reviews focus on architectural decisions. The latter generates much bigger savings.

2. Tagging discipline that actually allows attribution. Most enterprises have tagging policies that aren’t enforced. Cost data without reliable tagging is hard to act on. Investing in tagging enforcement pays off.

3. A clear stance on engineering effort vs cost saved. Some teams chase every dollar of savings. Others ignore cost entirely. The right stance is somewhere in between, and it should be explicit.

4. Architecture review for cost early, not late. Cost discussions in design reviews catch expensive choices before they’re built. Cost discussions in production catch the wrong things at the wrong time.

5. Active reservation portfolio management. This is the highest-use activity for stable workloads. It needs ownership and regular attention, not a once-a-year contract negotiation.

The vendors and partners worth working with

The cloud cost management vendor space is mature but uneven. Most enterprises don’t need a tool — they need disciplined process. A few engagements worth considering:

• AWS, Azure, and GCP all offer enterprise cost management programs. These are mostly free and provide reasonable starting frameworks.
• For deeper optimization, working with a partner like AI project delivery on architectural review often delivers more value than tool subscriptions.
• For specific high-cost services (databases, large compute clusters), service-specific specialists can find optimization opportunities that general-purpose tools miss.

The pattern that doesn’t work is buying a cost dashboard tool and assuming it will deliver savings on its own. Tools surface information. Action requires people and process.

What to do this quarter

If you want concrete next steps:

1. List your top 10 cloud cost line items by service, by month
2. For each, identify whether it’s growing, stable, or declining
3. For the growing or stable line items above $20K/month, schedule architectural review
4. Audit your reservation portfolio against the past 90 days of actual usage
5. Pick one quick win (storage tiering, idle resource cleanup, database tier review) and complete it before adding more

Most enterprises see 10-20% savings from disciplined cost management without any major architectural changes. The savings come from picking the right battles, not from spreading optimisation effort thinly across everything.