Public cloud vs. on-prem: which workloads belong where?

Key Takeaways

• 42% of organizations have already moved some workloads back from public cloud to on-premises infrastructure
• Predictable, steady-state workloads often cost less on-prem, while bursty or experimental workloads favor the cloud
• The hybrid approach wins for most enterprises, but drawing the line requires honest workload analysis

Cloud repatriation is no longer a contrarian position. A growing number of companies are pulling workloads back from AWS, Azure, and Google Cloud, not because the cloud failed them, but because they finally did the math. The real question has shifted: not whether to use the cloud, but which workloads actually belong there.

According to IDC, 42% of organizations have already repatriated at least some workloads from public cloud. That is not a rounding error. It is a pattern. And it is forcing CTOs to rethink infrastructure decisions made during the rush to cloud adoption in the 2010s.

Why are companies leaving the cloud?

The short answer is money. The longer answer is that cloud pricing works well for some workloads and terribly for others. When your compute needs spike unpredictably, or when you are still figuring out product-market fit, paying only for what you use makes sense. When you run the same workload 24/7 at predictable scale, you are essentially paying a premium for flexibility you do not need.

37signals, the company behind Basecamp and HEY, became the loudest voice in the repatriation debate when CTO David Heinemeier Hansson published detailed financials on their cloud exit. The company projects $7 million in savings over five years by leaving AWS.

That number got attention. But the more important point was what DHH said about which workloads drove those costs: steady, predictable database and compute workloads that ran constantly. These are exactly the workloads where cloud economics break down.

Which workloads make sense on-premises?

The pattern is consistent across organizations that have done this analysis. Workloads with these characteristics tend to cost less on your own hardware:

• Predictable, steady-state compute with minimal scaling requirements
• Large databases with consistent read/write patterns
• Data-intensive applications where egress fees add up
• Long-running batch processing jobs
• Workloads with strict data residency or compliance requirements

Corey Quinn, Chief Cloud Economist at The Duckbill Group, puts it bluntly: "The cloud is not for everything. Predictable workloads with consistent compute needs often make more sense on-premises."

The math is straightforward. A reserved instance on AWS might cost $30,000 per year for compute that you could run on a $10,000 server with a three-year lifespan. Add the operational cost of running your own infrastructure, and on-prem still wins for that specific workload. The tricky part is being honest about those operational costs.

When does public cloud still win?

Cloud providers did not build hundred-billion-dollar businesses by offering a bad deal. For certain workloads, the cloud remains the better choice:

• Bursty, unpredictable traffic patterns where you need to scale quickly
• Experimental projects where you do not know future requirements
• Globally distributed applications requiring edge presence
• Workloads that depend heavily on managed services (ML inference, real-time analytics)
• Small teams without infrastructure expertise

The cloud's value proposition is not compute. It is speed and optionality. You can spin up a Kubernetes cluster in minutes, deploy globally without building data centers, and shut it all down if the project fails. That flexibility has real value, especially for startups and enterprises running many concurrent experiments.

The hybrid reality most companies land on

Gartner analyst Lydia Leong captures the nuance: "Cloud repatriation isn't about rejecting the cloud. It's about optimizing placement based on workload characteristics." Flexera's 2023 data shows 72% of enterprises now pursue hybrid cloud strategies. They are not choosing sides. They are choosing per-workload.

The wasted spend stat is telling. It suggests that many companies are not overpaying because of cloud pricing itself, but because they have not done workload placement analysis. They lifted and shifted everything during migration, and they have not revisited those decisions since.

The smart approach is not "cloud-first" or "on-prem-first." It is workload-first. Map your applications. Understand their scaling patterns, data flows, and operational requirements. Then place each one where the economics and operational model fit best.

What it takes to repatriate

Moving off the cloud is not a weekend project. 37signals spent months executing their migration, and they had an unusually experienced operations team. For most organizations, repatriation requires:

1. Hardware procurement and data center capacity (or colocation contracts)
2. Operational staff who can manage physical infrastructure
3. Rearchitecting applications that depend on cloud-native services
4. Planning for disaster recovery without the cloud's built-in redundancy
5. Accepting longer lead times for capacity changes

The operational complexity is the hidden cost. Cloud bills are visible. The cost of hiring infrastructure engineers, responding to 3am hardware failures, and managing vendor relationships is diffuse and easy to underestimate. Companies that lack deep operations expertise should be skeptical of rosy repatriation projections.

The real conversation CTOs should be having

The public cloud vs. on-prem debate often gets framed as an ideological choice. It should be an analytical one. The question is not whether you believe in the cloud. The question is whether you have actually audited your workloads and matched them to the right infrastructure model.

Kelsey Hightower, formerly of Google Cloud, summed it up: "The right answer is almost always hybrid. The question is where to draw the line." That line will be different for every organization. It depends on your team's skills, your workload characteristics, your compliance requirements, and your appetite for operational complexity.

Global cloud spending hit $600 billion in 2023. That number will keep growing. But the composition of that spend is changing. More companies are running hybrid strategies, and more are willing to move workloads in either direction based on evidence rather than inertia.

Frequently Asked Questions

What is cloud repatriation?

Cloud repatriation is the process of moving workloads from public cloud providers like AWS, Azure, or Google Cloud back to on-premises infrastructure or private data centers. Companies typically repatriate to reduce costs for predictable, steady-state workloads.

How much can companies save by leaving the cloud?

Savings vary widely depending on workload characteristics. 37signals projects $7 million in savings over five years. However, companies must account for hardware costs, operational staff, and reduced flexibility when calculating true savings.

Which workloads should stay in the public cloud?

Workloads with unpredictable scaling needs, experimental projects, globally distributed applications, and workloads heavily dependent on managed services typically benefit from staying in the cloud. Small teams without infrastructure expertise also often benefit from cloud.

Is hybrid cloud the right approach for most companies?

According to Flexera, 72% of enterprises pursue hybrid strategies. Hybrid makes sense when different workloads have different characteristics. The key is placing each workload where its economics and operational requirements fit best.

What are the hidden costs of cloud repatriation?

Hidden costs include hiring and retaining infrastructure engineers, hardware maintenance, disaster recovery planning, longer lead times for capacity changes, and the opportunity cost of engineering time spent on infrastructure rather than product.

Source: The New Stack / Alex Wilhelm