Key takeaways
- A single VPC with public, private, and isolated database subnets across two Availability Zones is sufficient for most 5-50 person companies.
- Choose Lambda for spiky event-driven work, Fargate for steady-state services, and EC2 only when you need instance control or have committed spend.
- Set S3 lifecycle rules to Standard-IA after 30 days and Glacier after 90, then confirm the drop in Cost Explorer.
- Use IAM roles with an external ID and least-privilege scoping for all compute, CI/CD, and third-party access.
- A 30-minute monthly Cost Explorer review grouped by service and tag catches billing drift before it becomes a surprise.
A well-structured aws cloud architecture for a 5-50 person SaaS business comes down to five decisions: one VPC split across three subnets, containers or serverless for compute, S3 plus managed databases for storage, a read-only IAM boundary for security, and Cost Explorer reviews each month. This article walks one coherent reference architecture through each layer, stating the trade-off at every step so you can compare your own account against a single design instead of stitching together ten blog posts.
What does a well-structured AWS cloud architecture actually look like for a small business?
The reference architecture at a glance
A well-structured AWS cloud architecture for a 5-50 person company has five layers: a segmented VPC, a compute tier, a storage and data tier, a security perimeter, and a cost control loop. Each layer maps to a decision you can check in your own account today. The AWS Well-Architected Framework defines the pillars behind these layers: operational excellence, security, reliability, performance efficiency, cost optimization, and sustainability.
The layers exist because they fail independently. A flat VPC means a compromised subnet reaches your database. No tagging convention means nobody can attribute a surprise bill. Untested backups mean an RDS snapshot is a hope, not a recovery plan. Benchmarking against a reference architecture turns vague unease about your account into a checklist of specific gaps, which is the same structure we use in our AWS solutions architecture work.
What changes at your scale versus an enterprise design
Enterprise designs add multi-account landing zones, transit gateways, and dedicated network teams. At your scale, one production account, one staging account, and a single VPC with public and private subnets across two availability zones covers the requirements. Skip anything that needs a person whose only job is that thing.
How should you segment your VPC and networking layers?
For most UK SMBs, a single VPC with three tiers of subnets is the right starting point: public subnets for load balancers, private subnets for application workloads, and isolated private subnets for databases. Spread subnets across at least two Availability Zones so a single AZ failure doesn't take your service down.
Subnet layout and availability zones
Use a /16 CIDR block for the VPC and carve out /24s per subnet per AZ. That leaves room to grow without re-architecting later, which is painful once peering and endpoints are in place. Keep NAT gateways in the public subnets and route private subnets through them.
The NAT gateway cost trap
Each NAT gateway costs roughly $32/month plus data processing, and running one per AZ triples that. For a low-traffic account, one NAT gateway (accepting the cross-AZ exposure) is a reasonable trade-off. A VPC single-NAT setup with S3 and DynamoDB gateway endpoints avoids NAT charges for the heaviest traffic paths.
When one VPC is enough
One VPC with clear subnet separation is enough until you have compliance boundaries or separate teams needing isolation. Network security within one VPC comes from security groups and NACLs, not from more VPCs. If you want a second opinion on how your current setup is segmented, KloudCare's free Health Check reviews it under a read-only IAM role with an external ID you control (https://www.kloudcare.co.uk/security.html).
When should you choose serverless versus containers versus EC2?
Three criteria decide this for most small teams: traffic pattern, ops overhead, and what idle time costs you.
Lambda for spiky and event-driven work
If a workload sits idle most of the day and fires in bursts (webhooks, image processing, scheduled jobs), Lambda is hard to beat. You pay per invocation, so idle cost is zero, and there are no servers to patch.
Fargate for steady-state services
Long-running APIs and background workers with predictable traffic fit Fargate. You get container portability without managing instances, and costs track the vCPU and memory you reserve rather than whole machines.
EC2 when you need control or have committed spend
If you have a Savings Plan or Reserved Instances to burn, or you need specific instance types, GPUs, or OS-level access, EC2 wins on unit economics. The trade-off is AMI patching, scaling groups, and capacity planning you own.
Lambda Fargate EC2 Traffic pattern Spiky, event-driven Steady-state Steady or predictable
Idle cost Zero You pay for reserved capacity You pay for running instances
Ops overhead Lowest Low (no host patching) Highest (you manage the OS)
One hidden cost either way: any private subnet needing outbound internet access requires a NAT gateway, billed hourly plus per GB processed per AWS VPC documentation. If you want a second opinion on which workloads belong where, KloudCare's free Health Check reviews your architecture read-only and flags mismatches with an estimated savings figure.
How do you design storage and data layers for cost and resilience?
Choosing a database you can actually operate
RDS Postgres is the default for most SMBs: one instance, standard backups, skills you already have. Aurora costs more per instance but handles failover and read replicas with less babysitting. DynamoDB is cheapest at scale and has no server to patch, but its query model forces you to design access patterns up front. Pick the one your team can operate at 2am, not the one with the best benchmark.
S3 lifecycle rules that pay for themselves
Most S3 spend is data nobody touches. Set lifecycle rules to move objects to Standard-IA after 30 days and Glacier after 90, and check Cost Explorer to confirm the drop. One rule per bucket, reviewed quarterly, usually covers it.
Backups and restore testing
Common backup gaps: relying only on RDS automated backups with no cross-region copy, EBS snapshots missing entirely, and S3 versioning off on the one bucket that matters. The bigger gap is never testing a restore. Schedule a quarterly restore into a scratch environment and time it. If a documented, tested backup story is missing from your setup, a Health Check from KloudCare will flag it alongside estimated savings.
Where do security and IAM controls fit into the architecture?
Security in the reference architecture is a layer that crosses every other one: IAM at the access boundary, Security Hub as the baseline check, and encryption and logging wired into each tier rather than bolted on afterwards.
Roles, external ID and least privilege
Use IAM roles, not IAM users, for everything: your compute, your CI/CD, and any third party with a reason to look at your account. For external access, a cross-account IAM role with an external ID is the standard pattern: the third party assumes the role only from their account, and you can revoke it in seconds. Keep the role read-only unless a task genuinely requires write, and scope permissions to specific resources rather than *.
This is how we work with customers at KloudCare: a read-only IAM role, customer-controlled, with an external ID you set and can rotate.
A security baseline you can set up in an afternoon
Turn on Security Hub with the AWS Foundational Security Best Practices standard, enable CloudTrail in all regions, and review the first findings batch. Expect noise; triage the high-severity items first.
How do you keep AWS architecture costs under control as you scale?
Cost control is an architecture layer, not a monthly cleanup. Decisions you make in the VPC, compute, and storage layers determine whether your bill grows linearly with usage or quietly outpaces it.
Tagging as an architecture decision
Decide your tag schema before you build: service, environment, team, and cost-centre as a minimum. Enforce tags with AWS Config rules or Terraform checks in CI, because untagged resources are the ones nobody owns and nobody deletes. A tag applied at creation is free; retrofitting one across 200 resources is an afternoon you will not get back.
A monthly cost review that takes 30 minutes
Open Cost Explorer on the first of each month, group by service and by tag, and compare against the previous month. Then check Compute Optimizer for over-provisioned EC2 instances and RDS sizes. Thirty minutes, same slot every month, catches drift before it becomes a surprise bill.
Where waste hides in the reference architecture
In the reference design, the usual suspects are idle dev/staging environments running 24/7, unattached EBS volumes and old snapshots, NAT gateway data processing charges, and log groups with no retention policy. Each is small alone; together they are often 10-20% of a bill (illustrative, not measured). Budget the fixes as architecture changes: scheduled shutdowns for non-prod, lifecycle policies on logs, and a VPC design that keeps cross-AZ chatter down.
How do you know when your current AWS architecture needs redesigning?
Five warning signs
Surprise bills are the loudest signal: if you cannot trace a 40% month-on-month jump to a specific service, nobody owns the cost layer. Manual console changes that never make it into Terraform mean your architecture exists only in someone's memory. A single NAT gateway, database, or EC2 host that takes the whole product down is a single point of failure. Security exposure shows up as broad admin IAM roles and open security groups. The fifth sign is drift: staging and production behaving differently for reasons no one can explain.
One common trap: leaving always-on EC2 instances running for spiky workloads. Fargate bills per vCPU and GB of memory per second, so intermittent jobs can cost less there (AWS Fargate pricing).
Fix-first order for a small team
Fix IAM exposure and backups first, because the downside is unrecoverable. Then Terraform the console changes and tag resources so Cost Explorer can attribute spend. Leave working serverless and container setups alone; redesign only what is actively failing or burning money.
Should you redesign it yourself or get outside help?
Comparing your options honestly
DIY in the AWS console with spreadsheets costs nothing but your evenings, and it works until a surprise AWS bill or a security exposure lands during a release week. Built-in tooling covers real ground: Cost Explorer shows spend, Compute Optimizer flags oversized instances, Security Hub and Trusted Advisor surface misconfigurations. What none of them do is make the changes or own the outcome.
Consultancies like The Scale Factory, Cloudsoft, or AWS Professional Services bring depth, but engagement models and pricing are usually built around larger accounts than one doing $20k a month.
Where KloudCare fits and where it does not
Run KloudCare through the same criteria: it is SMB-focused, works on a fixed price, and takes read-only access through a customer-controlled IAM role with an external ID, so you keep the keys. It starts with a free assessment and a Health Check that produces estimated savings, backup gap findings, and security exposure items. Where it does not fit: if you want a team to implement changes inside your account or you need continuous hands-on engineering, a read-only advisory model is not that. If you enjoy the work and have the hours, DIY plus the built-in tools is a legitimate answer.
FAQ
How much should a small SaaS company spend on AWS per month?
There is no fixed number; the right question is whether you can attribute spend. If Cost Explorer grouped by service and tag explains every line item, your spend is under control at any level. Waste typically hides in idle staging environments, unattached EBS volumes, NAT gateway data processing, and log groups without retention policies, so audit those before cutting capacity.
Do I need multiple AWS accounts or is one VPC enough?
For most 5-50 person companies, one production account, one staging account, and a single VPC with public, private, and isolated database subnets across two Availability Zones is sufficient. Additional VPCs earn their keep only when you have compliance boundaries or separate teams needing isolation. Within one VPC, security comes from security groups and NACLs, not more VPCs.
Is it worth getting a Well-Architected Review for a small team?
A structured review is worth it if it produces a checklist of specific gaps rather than a vague report. You can approximate one yourself by benchmarking your account against five layers: VPC segmentation, compute choice, storage and backups, IAM boundaries, and a monthly cost review. An outside read-only assessment adds value mainly when you lack the hours to do that honestly.
What is the cheapest way to run a production database on AWS?
RDS Postgres on a single instance with standard backups is usually the cheapest option your team can operate, and operating skill matters more than sticker price. Aurora costs more per instance but reduces failover babysitting. DynamoDB is cheapest at scale but forces you to design access patterns up front. Right-size with Compute Optimizer rather than guessing.

