From On-Premises to Cloud and Back: Migration Case Studies

Author: Ivan Bogdanov, Technical Writer, HOSTKEY

We've journeyed from affordable cloud VPS solutions to dedicated servers, dissected hybrid architectures with autoscaling, and now, with a solid theoretical understanding and configured setups demonstrating promising results, the crucial question remains: what do companies that have already navigated this path actually do in practice? Are they migrating from the cloud, or vice versa, or are they embracing a combination of technologies? And, most importantly, what are the real-world financial implications?

Migrations are happening in both directions, and on a large scale. Dropbox offloaded petabytes of data from AWS to its own infrastructure, resulting in a $74.6 million cost saving, while Stack Overflow, in 2025, transitioned its operations from a private data center to the cloud (part 2 of the journey can be found here). 37signals made a bold announcement: "We've exited the cloud," and other companies are actively migrating to Google Cloud and Azure. Who's "right"? The answer is, everyone is, given their specific needs and the circumstances in which they aim to achieve their goals.

This article isn't about demonizing the cloud or declaring dedicated servers obsolete. Instead, it focuses on real-world case studies with verifiable data, where companies made migration decisions, analyzed the economics, and publicly shared their outcomes. Let's examine what they've learned from their experiences.

Industry Trends

Before diving into specific case studies, let's consider the broader picture. According to a Gartner forecast from November 2024, worldwide public cloud spending is projected to reach $723.4 billion in 2025. This represents a substantial investment, and the trend of increasing expenditures continues. IDC's forecasts indicate a 33.3% growth in cloud infrastructure spending in 2025 compared to 2024.

Interestingly, Gartner also predicts that by 2027, approximately 90% of organizations will adopt a hybrid cloud model. Neither purely cloud-based nor solely on-premise, but a combination of both. This suggests that the industry recognizes that a one-size-fits-all solution doesn't exist, and different aspects of the infrastructure require tailored approaches.

Why are companies migrating? The reasons are multifaceted, but here are a few key drivers:

• Economics: For stable, predictable workloads, dedicated servers often prove more cost-effective.
• Performance: Dedicated hardware offers predictable latency.
• Regulatory Compliance: Meeting requirements like Europe's GDPR or California Consumer Privacy Act (CCPA).
• Control: Full access to and control over the underlying hardware.

Migration is not a simple "push-button" process but a multi-month project that demands careful planning, rigorous testing, data synchronization, and proactive risk management. However, the data indicates that for many companies, these efforts yield a positive return on investment, although it's important to acknowledge that some migrations can turn into costly disasters.

From Cloud to On-Premise

Perhaps the most well-known case of returning from the cloud is Dropbox. In its S-1 prospectus filed before its IPO, Dropbox publicly stated that it had reduced operational expenses by $74.6 million over two years after deciding to transition from AWS to its own infrastructure.

What happened? In 2015, Dropbox began offloading petabytes of user data from AWS S3 to its own infrastructure. The project concluded in the fourth quarter of 2016, taking approximately two years. In 2016, AWS expenses decreased by $92.5 million, while investments in its own data centers totaled $53 million. The net savings in 2016 amounted to $39.5 million, and an additional $35.1 million in 2017. Notably, Dropbox retained approximately 10% of its data in AWS, primarily to serve users in Europe and regions where it didn't have its own data centers. This is a classic example of a hybrid approach, rather than a complete exit from the cloud.

In 2023, 37signals (Basecamp/HEY) published a prominent post and podcast titled "We Have Left the Cloud," announcing their decision to move Basecamp, the HEY email service, and five other applications from AWS to their own infrastructure. According to their official statement, this migration was achieved "without hiring any additional staff."

The migration didn't necessitate expanding the team because the company already possessed the necessary infrastructure management expertise. The company's founder, David Heinemeier Hansson, provided a concrete example: the databases (RDS) and search functionality (Elasticsearch) for HEY alone were costing over $500,000 per year. Furthermore, 37signals experiences stable and predictable workloads; they don't pursue explosive growth, and their traffic doesn't fluctuate dramatically from day to day.

While the exact figures for the cost savings are not disclosed, the company claims "millions of dollars in the long run." The very fact that this decision was made public speaks to its economic viability, given their scale and workload profile.

Ahrefs presents a slightly different case study. They didn't migrate from the cloud; they simply never adopted it. In their technical blog, the company published a cost analysis demonstrating that over six years, they invested $122 million in their own infrastructure. For comparison, the equivalent AWS infrastructure would have cost approximately $448 million over a 2.5-year period. Their infrastructure, comprising 3,300 servers, is managed by a team of just 11 people. In June 2023, this team handled 94 hardware incidents, averaging slightly more than four per day.

Sometimes, the optimal decision is to avoid the cloud altogether, especially if your workload is predictable and you're willing to invest in infrastructure.

Migrations from On-Premise to Cloud

An interesting example of the reverse trend is Stack Overflow's 2025 publication of a series of posts titled "Moving the Public Stack Overflow Sites to the Cloud," detailing their transition from a private data center to the cloud (part1 and part2).

The reason for the migration was straightforward: their data center contract was expiring on July 31, 2025. This wasn't a proactive optimization effort but a necessity. The team had to choose between finding a new data center to house their equipment or migrating to the cloud. They opted for the cloud.

Prior to this, Stack Overflow had already migrated its Teams product to Azure between 2021 and 2023. This migration involved three attempts over three years and encountered numerous setbacks. The team had to disentangle the Teams codebase from the rest of the codebase, containerize the applications, and transition from virtual machines to Kubernetes. This experience informed their planning for the migration of the public platform; they immediately established a dedicated team focused solely on this project, set a clear deadline, and adopted a phased approach rather than a "big bang" migration.

Another example is Baselime, which represents an optimization within the cloud ecosystem, but it's still illustrative. The company migrated its data ingestion points from AWS Lambda and Kinesis to the Cloudflare platform (Workers and Durable Objects). As a result, AWS Lambda costs were reduced by more than 85%.

This highlights that sometimes, you don't need to leave the cloud entirely; instead, you can find a more efficient provider for a specific workload. Serverless models can be expensive with one provider but cost-effective with another, depending on the pricing model, architecture, and other factors.

Examples of Hybrid and Specific Use Cases

Discord is an example of a company that evolved its infrastructure alongside its growth and made decisions about the appropriate deployment model as needed. In its official blog, Discord published a post titled "How Discord Stores Trillions of Messages," providing a detailed account of the evolution of its storage system. In 2017, the system ran on 12 Cassandra nodes. By early 2022, it had grown to 177 nodes and stored trillions of messages. This isn't a migration in the traditional sense but a gradual evolution to accommodate increasing workloads.

At this scale and with the demands for low latency and cost efficiency, companies often have to build custom solutions. It's no longer about "cloud versus dedicated servers" but rather "what works for our trillions of messages." Discord utilizes a hybrid architecture with managed clusters and custom optimizations.

In technical blogs, discussions on Hacker News, and DevOps communities, European startups and small to medium-sized businesses frequently share their experiences of migrating from AWS to European providers like Hetzner or OVH, particularly for stable, predictable workloads. While the actual figures vary depending on the workload profile and include support costs, the trend is clear: for certain scenarios, European providers offer significantly lower costs than AWS, Azure, or GCP.

It's important to understand that these providers don't offer the same range of managed services as AWS. This is closer to physical servers or basic virtual machines with minimal enhancements. In other words, cost savings are achieved by taking on more operational responsibilities.

Managed cloud provider solutions are another topic. They're convenient but have limitations. They typically support only the latest 2-3 software versions, system tuning is restricted to the provider's configurations, and low-level metrics may not be available. If your application uses older versions or requires specific kernel customizations, managed solutions may not be suitable. This issue often surfaces during the traffic transfer phase, when it's too late to make changes.

The European and US markets have increasingly embraced sovereign cloud strategies, driven by data residency laws, national security concerns, and the desire for digital autonomy. In Europe, this has led to a wave of migrations from global hyperscalers to locally controlled or EU-based cloud infrastructures. For instance, in 2023, Capgemini and Orange finalized the setup of Bleu, a joint venture designed to migrate sensitive French government workloads to a sovereign cloud environment that complies with the strict SecNumCloud 3.0 certification. Simultaneously, Atos secured a major multi-year framework agreement with UGAP to accelerate the digital transformation and cloud migration of over 300 French public sector entities. Similarly, a German healthcare insurer completed a 15-month migration of its entire IT infrastructure to a sovereign cloud environment operated under EU jurisdiction.

Major cloud providers have responded by launching dedicated sovereign offerings: AWS launched its European Sovereign Cloud with operations restricted to EU personnel, while Microsoft and Google have partnered with local entities like T-Systems to deliver compliant solutions. These efforts reflect a broader trend—by 2024, 40% of large enterprises were expected to migrate at least 10% of their workloads to sovereign clouds, with 37% of European companies already investing in such infrastructures. This shift is no longer just about regulatory compliance—it’s becoming a strategic lever for control, resilience, and long-term cost efficiency.

When Migration Goes Wrong

It's also worth noting that there are plenty of documented cases of failed migrations: British bank TSB lost £300 million due to inadequate testing before the switchover, and the Warsaw University spent years failing to effectively utilize its HPC system, with a hybrid solution attempt ultimately failing. These cases are well-known, but it's interesting to examine other examples, as they reveal that problems can arise even for those who ultimately achieve the desired results.

GEICO, the third-largest auto insurer in the US with revenue of $39.6 billion and profit of $3.6 billion in 2023, is part of Warren Buffett's Berkshire Hathaway empire. In 2013, the company began migrating over 600 applications to the cloud. Ten years later, the outcome proved to be the opposite of expectations. Cloud costs doubled, and system reliability declined significantly. Rebecca Weekly, GEICO's vice president of infrastructure, stated in October 2024: "Over the past decade, GEICO has not fully transitioned to the cloud, costs have doubled, and reliability issues have increased significantly."

What went wrong? GEICO has massive amounts of data, as is typical for most insurers, which rely on predictive risk analytics and must retain data for regulatory purposes. It turned out that storing data in the cloud is one of the most expensive things to do. Furthermore, the company adopted a lift-and-shift approach, simply moving legacy applications to the cloud without refactoring. Weekly concluded: "Simply running legacy applications in the cloud is prohibitively expensive. Our case clearly demonstrates this."

GEICO is now returning to on-premise infrastructure with OpenStack, Kubernetes, and its own infrastructure. Berkshire Insurance's management publicly acknowledged in 2023: "GEICO's technology requires much more work than we anticipated. The company has over 600 legacy systems that do not interact with each other." The goal is to reduce these to 15-16 systems.

SilkRoad Technology decided to migrate to the cloud to reduce costs and improve efficiency. This seemed logical: less hassle with hardware, more focus on the product. In practice, after the migration, the opposite happened: costs increased, infrastructure performance was suboptimal, and performance suffered.

They had to revert and move the systems back to their own infrastructure. The company spent additional time and money, and its reputation suffered as customers experienced service issues and asked difficult questions.

In 2019, IHS Markit published a study for Fortinet showing an interesting phenomenon of dynamic multi-cloud. 74% of the 350 organizations surveyed had moved at least one application back from the cloud to their own infrastructure. The reasons were typical: performance issues (52%), security concerns (52%), and planned temporary deployments (40%).

Looking at the experiences of real migrations, several common problems emerge:

• Total cost of ownership dictates the choice, not just the price of a server. Direct infrastructure costs are only part of the picture. You need to consider the costs of the team, hidden costs for training and migration, and the opportunity cost of the team being focused on infrastructure rather than the product. Companies that only consider the cost of virtual machines make incorrect decisions.
• Team expertise is more critical than the size of the budget. Cheap hardware without skilled personnel ends up costing more than the cloud. If you don't have a team to support your own infrastructure, the cloud remains a reasonable choice, even if it appears more expensive on paper.
• Hybrid architecture outperforms pure solutions. The industry is moving towards combining technologies rather than choosing between cloud and on-premise. Different types of workloads require different approaches: storage on one platform, compute on another, peak loads on a third.
• Context and specific needs are more important than trends. What worked for one company may fail for another. Migration should solve a specific business problem with a measurable outcome, rather than following hype or other companies' case studies.
• Migration is a full-fledged project, not a simple technical operation. Months of planning, a dedicated team, dozens of stages, load testing. Companies that treat migration as a "quick move" encounter performance problems, increased costs, and the need to roll back.
• Failures happen even with large budgets. Underestimating complexity, skipping testing stages, lift-and-shift without refactoring lead to increased costs instead of savings. 74% of companies have moved at least one application back after migration.

The main takeaway from all the case studies described above is that there is no "right" infrastructure; there are only solutions that are appropriate for your needs right now.