Back to blog
awsfinopsarchitecturecost-optimization

When your AWS bill stops being a cost problem and becomes an architecture problem

How to distinguish one-off spending optimization from architectural decisions that create recurring costs in AWS.

When your AWS bill stops being a cost problem and becomes an architecture problem

A high AWS bill does not always point to a problem with the architecture. Sometimes it says something healthy: there are more customers, more traffic and more actual product usage. Costs rise because the business is consuming more infrastructure too. There may be room to optimize, but not necessarily to redesign.

Other times, the problem is more basic: forgotten resources, environments nobody shuts down, logs retained without clear criteria, oversized instances or accumulated snapshots. That can usually be resolved with a lightweight cost review, better visibility and clear owners.

A resource looking forgotten does not give you permission to delete it: before shutting it down or changing its retention, confirm its owner, dependencies, and recovery and retention requirements. An old snapshot may still be necessary.

There is also a third situation: the cost is neither an incident nor overdue housekeeping, but a direct consequence of how the platform is designed and operated.

That is where the question changes. It is no longer “what do we cut this month?” but “which technical decisions make this cost return every month?”

One-off optimization is not the same as structural cost

It is worth separating three levels that are often mixed together:

  • Basic cleanup: resources that are no longer used, temporary environments left running, sizes chosen out of habit, old disks, forgotten snapshots or log retention settings nobody reviewed. These are real problems, but they do not normally require rethinking the architecture.
  • Lightweight FinOps optimization: understanding which services account for the bill, reviewing trends, assigning owners, detecting anomalies and deciding whether certain spending commitments make sense. For many small teams, this is enough to stop making decisions blindly without building a full FinOps program.
  • Recurring architectural cost: the current design naturally generates spending: through how traffic flows, how services scale, how environments are duplicated, how data is stored, how systems are observed or how responsibilities are distributed.

The important signal is not that the bill goes up. It is that the same pattern returns even after the team performs one-off cleanups.

Signs that cost points to architecture

There are several useful clues:

  • Nobody can explain which business flow justifies a significant portion of the spending. The cost exists and returns every month, but it is unclear whether it reflects more users, more useful data or simply infrastructure that grows by design.
  • The bill increases because of foundational components rather than clear functional usage: data transfer, networking, observability, clusters, databases, storage or duplicated environments. These are not problems by definition, but they do deserve an architectural reading, not just an accounting one.
  • Every new customer, feature or environment replicates expensive resources. If adding a small part of the product forces you to duplicate too much infrastructure, adjusting sizes alone will not fix the problem.
  • The team tweaks limits, sizes and settings but never questions the pattern producing the cost. An instance is downsized, a retention period changed, an autoscaling setting adjusted, but the design driving the spending remains the same.

Where these patterns tend to appear in AWS

In AWS, many costs have a reasonable technical explanation. The goal is not to label services as bad by default, but to understand which decisions make them expensive, opaque or hard to control.

Networking and data transfer

Networking is one of the places where architecture turns into a bill. NAT Gateway, traffic between availability zones, traffic between regions, access to public endpoints or routes that force traffic through intermediate components can have a financial impact.

NAT Gateway is not a mistake. But its cost model depends on how long it is available and the volume of data processed, as well as possible transfer charges. That is why it matters to review which traffic passes through it, whether it needed to, and whether resource placement is creating unnecessary traffic.

The question is not “do we remove NAT Gateway?” The question is which traffic pattern we are paying for every month and whether it reflects a conscious decision.

Compute and containers

Something similar happens with containers. EKS, ECS, Fargate or EC2 are not good or bad decisions in the abstract. They depend on the team, the workload, the operating model and the complexity the team can take on.

With EKS, you need to consider the cost of the cluster and the resources associated with its workloads. With ECS, the execution model — Fargate or EC2 — changes how you interpret the cost. In both cases, it is worth looking at what stays on, how the workload scales, and which operational complexity you are paying for.

An underused cluster is not always a mistake: it may reflect availability, isolation or anticipated growth requirements. But if nobody can defend that decision using technical and business criteria, the bill is telling you something.

Databases

Databases bring together cost, risk and caution. Large instances, storage, IOPS, replicas, Multi-AZ, backups and snapshots may be justified by availability, performance or recovery needs.

The problem appears when those decisions persist through inertia. A database sized for an old peak, replicas without a clear purpose, storage without a lifecycle or backups nobody has tested can turn a prudent decision into a structural cost.

You need to be careful here. Reducing database costs without understanding workload, recovery, availability and dependencies can prove expensive. The review must distinguish actual need from inherited configuration.

Observability

More logs do not always mean more visibility. Sometimes they mean more noise, longer retention and higher costs without better diagnostic capability.

CloudWatch Logs and other observability components are billed by usage; that is not the problem. The problem is producing and storing signals without deciding their operational value. Overly verbose logs, indefinite retention, metrics nobody looks at or traces without a specific question can turn into a recurring bill that is hard to defend.

If observability lacks clear criteria, its costs will too.

Environments and duplication

Development, staging, pre-production and production may all be necessary. What is not always necessary is for all of them to replicate the same shape, size and permanence.

Duplicating resources may look like a quick and safe way to create another environment. It can make sense at first. But without a lifecycle, a clear owner or a basic policy on what must always stay on, environments stop being helpful and become a fixed cost.

The question is not how many environments we have. It is what purpose each one serves, how closely it needs to resemble production, and which part of its cost gives the team real confidence.

Useful questions before redesigning

Before discussing major changes, it is worth getting specific about:

  • Which part of the cost is tied to actual product growth.
  • Which costs appear even when there are no additional users or traffic.
  • Which decisions were made for speed and have now become permanent.
  • Which components scale together even though they should not.
  • Which spending has no clear owner.
  • Which services account for the bill and which technical decisions they correspond to.
  • Which changes would be reversible and which increase operational risk.

These questions prevent two mistakes: treating everything as minor housekeeping, or turning any high bill into an excuse to redesign half the platform.

When you do not need a redesign

Not every high bill calls for architectural work.

If the problem is a lack of basic visibility, you need to establish it first. If basic tags, owners, trend reviews or a clear understanding of Cost Explorer are missing, redesigning may be premature.

If the spending comes from forgotten resources or clearly neglected settings, cleanup, process and accountability are probably enough. You do not need to rethink a platform because someone left a temporary environment running.

If costs rise because the product is growing and the relationship between usage and business value is reasonable, the work may be to optimize without breaking what works.

And if one-off costs have not yet been separated from recurring ones, start with a fact-based review. Architectural changes are better made when you know which problem you are solving.

How to approach an architecture-and-cost review without turning it into a huge project

A useful review does not start with an endless list of best practices. It starts with the services that account for the bill and connects them to technical decisions:

  1. Identify the major spending categories and their trends, not to carry out an accounting audit, but to know where to look.
  2. Map those costs to product flows, technical components and owners. If part of the bill has neither an owner nor an explanation, there is already work to do.
  3. Separate quick improvements from structural changes. Shutting down forgotten resources is not the same as changing a network pattern, reviewing a container strategy or rethinking how logs are generated.
  4. Prioritize by recurrence, risk and reversibility. A low-risk change that addresses a recurring cost is usually a better candidate than an ambitious restructuring with uncertain impact.
  5. Document the decision: why a component exists, which cost is accepted, which alternative was ruled out, and when it should be reviewed.

If costs depend on tribal knowledge, they will return as a surprise.

The bill as a symptom, not an enemy

The AWS bill should not be read only as a financial problem. It is also a signal about architecture, operations and team maturity.

Sometimes it calls for cleanup, more visibility or accepting that the product has grown. Other times, it calls for reviewing technical decisions that were reasonable months ago but now generate recurring costs.

FinOps and architecture do not compete. They complement each other. FinOps helps you see where the spending is and how it evolves. Architecture helps you understand why it exists and what would need to change for it to make sense.

The goal is not to spend as little as possible. It is to give the cost a defensible technical and business explanation.

If your AWS bill starts recurring as a symptom rather than an isolated incident, an architecture-and-cost review can help separate quick improvements from technical decisions worth rethinking.

Sources consulted

AR

Alejandro Rodríguez

Freelance Cloud & DevOps Consultant, Madrid