I'm leaving the cloud! (...and why you probably should too)

Moving most of his SaaS stack off AWS onto dedicated bare-metal servers cut monthly infrastructure costs by more than half and also sped up deployments—while reducing the risk of vendor lock-in that can force expensive, disruptive migrations. After shifting over 80% of his systems to Hzner, his average AWS bill fell from about $7,800 per month to under $2,000, with expectations that it could drop below $1,000 as the remaining migrations finish. The change matters because it directly affects both unit economics (costs that otherwise rise with usage) and operational resilience (the ability to keep running if a hyperscaler changes pricing or terms).

The new setup replaces AWS-managed serverless components with a more self-managed infrastructure layer. Code now runs in Docker containers orchestrated by Kubernetes on three dedicated Hzner servers. Data storage moved to PostgreSQL, caching to Redis, and background jobs/message queues to Bullmq. Monitoring and access logging are handled with Prometheus and Grafana. The cost model is also fundamentally different: Kubernetes, PostgreSQL, Redis, Bullmq, and Prometheus are open-source and free, with Grafana relying on its community edition. That leaves the dedicated servers as the primary recurring expense—around $200 per month for Hzner—plus whatever portion of infrastructure remains on AWS.

The trade-off is time and expertise. Migration required learning Kubernetes, Docker, and basic system administration, and infrastructure work spiked early in 2025. Over time, the workload returned to roughly prior levels, helped by better tooling and AI assistance. The financial logic hinges on that shift: the upfront migration effort is largely fixed, while the savings keep compounding month after month with no built-in usage ceiling like AWS’s per-request/per-resource pricing.

A second, arguably bigger reason centers on portability and continuity. AWS services such as DynamoDB, Lambda, and Fargate are robust, but they don’t run outside AWS. If AWS raises prices or stops serving a product configuration, the business can be “cooked,” because migration can’t be done quickly without careful planning. By moving to a stack that can run on any VPS or dedicated server, he aims to avoid being trapped by a single vendor’s infrastructure.

He still keeps some dependencies on AWS and other third parties. Stripe and OpenAI remain in the mix, and he’s experimenting with alternatives like Coinbase for payments and independent vendors for GPU-based AI inference. For the remaining AWS services, he’s tried to replace S3 with S3-compatible options such as Hzner storage buckets and Cloudflare R2, but vendor lock-in still persists. MinIO looked promising for self-hosted S3 compatibility, yet backup complexity and stability issues pushed him back to S3. Authentication also stayed on AWS for now: he tested Keycloak as an alternative to Cognito but returned to Cognito due to poor user/developer experience.

Finally, the move supports his business model. He sells both subscriptions and lifetime deals. Rising operational costs tied to user growth make lifetime pricing riskier; shifting to fixed-cost infrastructure makes financial planning easier and helps address customer questions about what happens if the company can’t host “lifetime” access. His goal is to make all products fully self-hostable so lifetime customers can run the software themselves if needed.

For deciding whether to do the same, he draws a line by stage: hobby projects should stay on cloud one-click platforms, mission-critical or highly scaling businesses may still benefit from managed hyperscalers, and the middle ground is where dedicated servers can be a practical, money-saving option—especially when teams want the freedom to use open-source tools internally.