- AWS wins on breadth and hiring pool, Azure wins when your identity and Microsoft estate are already there, and Google Cloud wins on data and Kubernetes ergonomics.
- The bill is decided by architecture, not by the provider's price list: egress, managed database tiers, and support plans move the number more than compute rates.
- Managed Kubernetes is mature on all three, so pick based on IAM, networking, and observability fit rather than on which control plane is 'best'.
- Lock in is mostly a data gravity problem: Postgres, S3-compatible storage, and containerized workloads port; proprietary queues, auth, and AI endpoints do not.
- Run a two-week spike on your real workload before committing; a spreadsheet comparison hides the operational cost that shows up in month three.
Your team just finished a migration to AWS, and the first bill after the free tier expired is double what the spreadsheet predicted. Nothing on the price list changed. What changed is that you started moving data between availability zones, running a managed Postgres instance with multi-AZ failover, and paying for a support plan that finance approved without reading the tiers. Devs & Logics is a US-focused software development agency that builds SaaS MVPs, web platforms, mobile apps, AI products, and cloud setups for startups and enterprises, and we have watched mid-market teams make this exact mistake on all three hyperscalers. The provider you pick matters less than the architecture you put on top of it. Here is the comparison we run with clients before they sign a commit, along with the tradeoffs that show up in month three rather than month one. If you want the operational patterns behind this, our DevOps best practices for startups guide covers the CI/CD and environment side.
How Do AWS, Azure, and Google Cloud Actually Differ for a Mid-Sized Business?
They differ less on raw compute and far more on identity, networking defaults, data services, and the commercial agreements your finance team can sign. For a mid-market company, the provider that matches your existing identity system and compliance posture usually beats the one with the lower list price.
Identity is the first fork. If your company runs Microsoft 365 and Entra ID (formerly Azure AD), Azure lets you wire single sign-on, conditional access, and role based access into your internal tools with almost no glue code. On AWS you will use IAM Identity Center and federate to Entra ID or Okta; on Google Cloud you will use Cloud Identity and Workforce Identity Federation. All three work, but the one that matches your directory is the one your security team will actually approve without a six-week review.
Networking defaults are the second fork. AWS gives you VPCs, subnets, security groups, and NACLs with a lot of knobs and a lot of ways to lock yourself out. Azure gives you virtual networks and network security groups with a flatter mental model that maps to how Windows admins already think. Google Cloud gives you a global VPC with firewall rules that apply across regions, which is simpler until you need per-region isolation and have to build it yourself.
Commercial terms are the third fork. Mid-market companies often qualify for enterprise discount programs or committed-use discounts on all three, but the paperwork, minimums, and renewal cycles differ. Azure agreements often ride on an existing Microsoft Enterprise Agreement; AWS and Google Cloud typically require a separate private pricing negotiation. Ask your finance team which vendor is already in the procurement system before you fall in love with a console.
None of these differences show up in a compute price per hour. They show up in how many engineer-weeks it takes to get your first production workload through security review.
Pricing Models Compared: Where the Bill Actually Comes From
Compute list prices are close enough to be noise; the bill is driven by egress, managed database tiers, support plans, and committed-use discounts. A mid-market workload that looks 15% cheaper on paper often lands within 5% of the others once data transfer and support are included.
The pattern we see repeatedly: a team compares on-demand EC2, VM, and Compute Engine rates, picks the cheapest, and then gets surprised by the line items that are not compute. Egress to the public internet is the classic one. AWS charges for data out; Azure and Google Cloud charge too, but the free tiers and per-GB rates differ, and the direction your traffic flows (in vs out, cross-region vs cross-zone) changes the math. If your product serves large media files or streams API responses to mobile clients, egress can rival compute.
Managed databases are the second hidden cost. A multi-AZ Postgres instance with automated backups, read replicas, and point-in-time recovery is priced differently on each provider, and the smallest instance size that meets your IOPS needs is often larger than you expect. On AWS, RDS and Aurora have different cost shapes; on Azure, Flexible Server vs Single Server matters; on Google Cloud, Cloud SQL vs AlloyDB matters. Budget for the tier you will actually run in production, not the dev tier.
Support plans are the third. Business-level support on all three is a percentage of your monthly spend with a minimum, and the minimum is not trivial for a mid-market company. Read the response-time SLAs carefully: the difference between a one-hour and a four-hour response on a production incident is a real business cost.
Here is the comparison we use with clients. The numbers are directional ranges, not quotes, because your actual bill depends on architecture and negotiated discounts.
| Cost driver | AWS | Azure | Google Cloud |
|---|---|---|---|
| Compute on-demand | Broad instance families, savings plans and reserved instances | Broad VM families, reserved instances and savings plans | Broad machine families, committed-use discounts and sustained-use discounts |
| Egress to internet | Charged per GB after free tier; cross-AZ traffic also billed | Charged per GB after free tier; zone-to-zone traffic billed | Charged per GB after free tier; premium vs standard network tier changes rate |
| Managed Postgres | RDS and Aurora, tiered by instance, storage, IOPS, backups | Flexible Server, tiered by vCores, storage, backups | Cloud SQL and AlloyDB, tiered by vCPU, memory, storage |
| Support plan | Business support as % of spend with monthly minimum | Standard and Professional Direct tiers with monthly minimum | Standard, Enhanced, Premium tiers with monthly minimum |
| Committed-use discount | Savings Plans (1 or 3 year), Reserved Instances | Reserved Instances, savings plans | Committed Use Discounts (1 or 3 year), sustained use |
If you are still scoping the product itself, our SaaS MVP development cost breakdown pairs well with this table, because cloud spend is only one line in the build budget.
Managed Kubernetes: EKS vs AKS vs GKE in Practice
All three control planes are production grade in 2026, so the decision comes down to IAM integration, node autoscaling behavior, and how much of the platform you want to own. GKE has the least operational surface, AKS has the tightest Entra ID story, and EKS has the widest ecosystem of add-ons and hiring.
On EKS, the control plane is managed but the surrounding platform is largely yours: you pick the CNI, the ingress controller, the autoscaler (Karpenter has become the default for many teams), and the observability stack. That flexibility is great if you have a platform engineer and painful if you do not. The IAM story uses IRSA (IAM Roles for Service Accounts) to give pods AWS permissions, which works well but requires understanding the OIDC provider setup.
On AKS, the tightest integration is with Entra ID and Azure RBAC. If your team already lives in the Azure portal, workload identity and managed identities feel native. Node autoscaling is handled by the cluster autoscaler and, increasingly, Karpenter on Azure. The tradeoff is that some AKS features and add-ons are Azure-specific, so portability takes more discipline.
On GKE, Autopilot mode removes most node management entirely: you define pod resource requests and Google provisions and bills nodes for you. For a mid-market team without a dedicated platform group, Autopilot is the lowest-operational-surface option we have shipped. The tradeoff is less control over node-level tuning and a billing model that can surprise teams used to paying for idle nodes.
Our rule: if you have fewer than two engineers who can debug a CNI issue at 2 a.m. pick the managed option with the least surface (GKE Autopilot or AKS with managed node pools), not the one with the most knobs. Kubernetes is not the goal; shipping is.
Database and Data Ecosystem Fit
Pick the provider whose managed Postgres, queueing, and analytics services match the shape of your data, because migrating a live database is the most expensive move you can make. AWS has the broadest menu, Azure has the strongest Microsoft data stack integration, and Google Cloud has the best serverless analytics story.
For a typical mid-market SaaS, the core is a relational database. All three offer managed Postgres: RDS and Aurora on AWS, Flexible Server on Azure, Cloud SQL and AlloyDB on Google Cloud. If you use PostgreSQL row level security for multi tenancy, verify that your managed offering supports the extensions and connection pooling you need. PgBouncer in transaction mode, for example, can break session-level features, so plan your pooling layer before you pick the tier.
Queues and eventing are where lock in creeps in. AWS has SQS, SNS, and EventBridge; Azure has Service Bus and Event Grid; Google Cloud has Pub/Sub. These are excellent services, and they are also proprietary. If portability matters, wrap them behind an interface in your code so you can swap the implementation later. We have seen teams spend weeks untangling direct SDK calls scattered across a codebase.
Analytics is the tiebreaker. If your product needs serverless SQL over large datasets, BigQuery on Google Cloud is the most ergonomic option we have used, with Redshift and Synapse as the AWS and Azure counterparts. If your data already lives in Microsoft Fabric or Power BI, Azure keeps the pipeline short. If you are building an AI feature with embeddings, all three now offer vector support in their Postgres and search services, so the decision comes back to where your operational data already lives.
Developer Velocity and Lock-In Risk
Developer velocity is mostly a function of how fast your team can provision, deploy, and debug, and lock in risk is mostly a function of how much proprietary surface you adopt. Teams that keep workloads containerized and data in open formats can move providers in weeks; teams that lean on proprietary auth, queues, and AI endpoints cannot.
Velocity starts with the inner loop: how fast a developer can get a working environment, run tests, and see a change in a preview environment. All three providers support this, but the ergonomics differ. AWS has the widest set of CI/CD integrations and the largest hiring pool, which matters when you are recruiting. Azure integrates tightly with GitHub Actions and Azure DevOps, which is a real advantage if your team already uses them. Google Cloud pairs well with Cloud Build and GitHub, and its Cloud Run service is the fastest path from container to URL we have used on any provider.
Lock in is a data gravity problem more than a compute problem. Containerized workloads on Kubernetes port between providers with modest effort. S3-compatible object storage ports if you avoid proprietary features. Managed Postgres ports with a dump and restore. What does not port cheaply: proprietary identity (Cognito, Entra External ID), proprietary queues, proprietary AI endpoints, and serverless function runtimes with vendor-specific triggers. Adopt those deliberately, not by accident.
Testing discipline is part of velocity. If your test suite runs in CI on every PR with Playwright for end to end and Vitest for units, you can refactor infrastructure with confidence. Our software testing and QA services exist because we have watched teams skip this and pay for it during a migration.
One honest limitation: if your team is under five engineers and your workload is a single Next.js app with a Postgres database, a managed platform like Vercel plus Supabase or Neon will beat all three hyperscalers on operational cost and time to first deploy. The hyperscalers win when you need private networking, compliance boundaries, or a data platform that has to live next to your application; they lose when you are paying for a control plane you do not have the headcount to operate.
How to Choose: A Decision Framework and Migration Path
Choose based on identity, data gravity, and hiring, then validate with a two-week spike on your real workload before signing a commit. The provider decision is reversible if you keep the architecture portable, and expensive if you do not.
Start with identity. If your company runs Microsoft 365 and Entra ID, Azure removes friction. If you are a Google Workspace shop, Google Cloud aligns. If you are a startup with no strong directory, AWS has the widest default hiring pool and the most third party integrations. This one input often decides the answer before cost enters the room.
Next, map data gravity. Where does your production data live today, and what would it cost to move it? If you have years of data in S3 and Redshift, moving to another provider is a project, not a sprint. If you are greenfield, you have more freedom, so optimize for the team you can hire rather than the service catalog.
Then run a two-week spike. Deploy a representative slice of your real workload: one service, one managed database, one queue, and your CI/CD pipeline. Measure time to first deploy, time to debug a failed deploy, and the bill for that slice at production-like scale. This is where spreadsheet comparisons fall apart, because the operational cost shows up in the second week, not the first.
Finally, design for portability where it is cheap. Keep workloads containerized, use Postgres, use S3-compatible storage where possible, and wrap proprietary services behind interfaces. If you need a worked example of what this looks like under a real compliance constraint, our GDPR-ready B2B SaaS case study shows the architecture decisions we made and why. For teams ready to move, our DevOps and cloud engineering services cover the migration, CI/CD, and cost tuning work.
Is AWS or Azure cheaper for a mid-sized business?
Neither is reliably cheaper; the bill depends on architecture and negotiated discounts more than list price. Azure can be cheaper if you already have Microsoft enterprise agreements and Entra ID, because you avoid duplicate identity and support spend. AWS can be cheaper if your workload is containerized and you use savings plans aggressively. Run the two-week spike on your real workload before you trust any comparison.
Can we run a mid-market SaaS on Google Cloud without a dedicated platform team?
Yes, especially with GKE Autopilot or Cloud Run, which reduce the platform surface you have to own. The tradeoff is that some operational knowledge is Google-specific, so hiring is narrower than AWS. If your team is small and your workload is containerized, Google Cloud is a credible first choice, not just an also-ran.
How hard is it to migrate from AWS to Azure or Google Cloud later?
It ranges from weeks to quarters depending on how much proprietary surface you adopted. Containerized workloads, managed Postgres, and S3-compatible storage move with moderate effort. Proprietary queues, identity, and AI endpoints require rewrites. The cheapest migration insurance is architectural discipline applied early, not a migration project later.
Which cloud provider is best for a HIPAA-aware or GDPR-ready workload?
All three offer the compliance tooling and regions you need; the difference is in how much work it takes to configure and document. Azure often has the shortest path if your compliance team already works in Microsoft tooling, and AWS has the broadest set of pre-built compliance artifacts. Google Cloud is competitive, particularly for data workloads. Note that a HIPAA-aware build describes the software and its controls, not a certification of your agency or provider; you still need a business associate agreement and a real security review.
Three takeaways to leave with. First, pick the provider that matches your identity system and hiring reality, because those are the costs that persist. Second, design for portability where it is cheap: containers, Postgres, and open formats. Third, validate with a two-week spike before you sign a commit, because the bill and the operational burden show up after the spreadsheet is closed.
If you want a second set of eyes on the decision, our team at Devs & Logics can run the spike with you and hand back a costed architecture. Start with DevOps and cloud engineering services.
