You do not need to build a data center to own the computers running your software.
The 37signals case
37signals published $3.2 million in cloud spending for 2022 after active optimization and long-term commitments. It later moved major workloads to owned hardware across two professional colocation facilities.
The boundary
The company retained an experienced operations team and had stable, highly utilized workloads. Its savings are evidence for calculation—not a universal promise.
Good candidates
- Steady demand and high utilization
- Large persistent storage
- Predictable growth
- Existing infrastructure expertise
- Control or jurisdiction requirements
Bad candidates
- Short-lived experiments
- Extreme or unpredictable bursts
- Small teams relying heavily on managed services
- Workloads whose cloud premium buys scarce capability
The point: “cloud first” and “on-prem first” are both shortcuts. Place each workload where its economics and control requirements make sense.