37signals turned its exit from the public cloud into one of the industry’s most cited cases, estimating that the savings will exceed $10 million over five years. However, their experience does not suggest that all companies should revert to the traditional in-house data center model. Three years after launching the project, global cloud spending continues to grow, and much of the “repatriated” workloads end up in dedicated infrastructure or colocation services, not within technical rooms inside offices.
The key points of cloud repatriation in 30 seconds
- 37signals expects to save over $10 million in five years after abandoning much of AWS.
- Business spending on cloud infrastructure hit $129 billion in Q1 2026, a 35% increase.
- The rise in DRAM raises the costs of new in-house infrastructure projects and extends their amortization period.
- Many repatriations involve moving to private cloud, bare metal, or colocation, rather than to corporate data centers.
- The decision depends on workload type, scale, control, and operational capacity, not just price per server.
The case of 37signals remains real and relevant. The company behind Basecamp and HEY initially announced that leaving Amazon Web Services (AWS) could save around $7 million over five years. Later, they raised their estimate to over $10 million after realizing they could install new hardware within existing racks and electrical limits.
They started from challenging conditions—experienced technical team, proprietary applications, predictable workloads, and prior data center presence. They didn’t need to buy a building, set up a new server room from scratch, or hire a full department to operate the infrastructure.
Their exit was also not immediate. In January 2026, 37signals was still migrating billions of objects out of Amazon S3, one of the final and most delicate phases. The repatriation required custom tools, planning, bandwidth, and a migration designed to avoid service interruption.
The lesson isn’t that public cloud is always more expensive. It’s that some stable, high-volume workloads may end up paying a premium over years for elasticity, managed services, and on-demand consumption they rarely utilize.
Public cloud continues to grow despite debate
The 37signals case prompted thoughts that the industry was preparing for a mass retreat from hyperscalers. Market data shows a more complex situation.
According to Synergy Research Group, global business spending on cloud infrastructure services reached approximately $129 billion during Q1 2026. This represents a 35% year-over-year increase and suggests an annual rate surpassing half a trillion dollars.
Growth accelerated for nine consecutive quarters. These figures are hard to reconcile with a widespread exit from public cloud, although some companies are shifting specific workloads.
Both movements can occur simultaneously. Organizations are retiring stable systems, massive storage, or predictable databases, while launching new AI, analytics, SaaS, and international expansion projects in the cloud.
Repatriation doesn’t necessarily mean returning to an on-premises data center. It often involves moving from a public instance to dedicated servers, private cloud platforms, or colocation racks.
This nuance explains why data center operators also benefit. Digital Realty, for instance, raised its annual forecast in July, citing sustained demand linked to cloud and AI. The market isn’t replacing one model with another in an orderly way; it’s redistributing workloads among public, private, and dedicated infrastructures.
The key question is no longer whether a company should be “in the cloud” or “off,” but rather where each workload performs best and what level of control, flexibility, and operational responsibility the organization is willing to assume.
Server buying has also become more expensive
Economic comparisons have become more complex in 2026. Repatriation often starts with a simple premise: purchasing infrastructure and amortizing over several years should be cheaper than continuous leasing.
But the cost of ownership has changed.
TrendForce revised its forecast for contractual prices of conventional DRAM in Q1 2026, projecting increases of between 90% and 95% compared to the previous quarter. Demand from AI and data centers has strengthened manufacturers’ bargaining power, leading them to allocate capacity to higher-margin products.
HBM memory, used alongside AI accelerators, competes with standard DRAM for investments, wafers, packaging, and industrial capacity. Meanwhile, manufacturers are avoiding the overproduction cycles that historically drove prices down.
Samsung, SK hynix, and Micron’s announced expansions will take years to deliver significant volumes. Some new capacity won’t come online until 2027 or later, even as business demand continues to grow.
This directly impacts server refreshes. A project budgeted at 2024 or 2025 prices might now require more initial investment, especially when large quantities of DDR5, flash storage, and latest-generation processors are needed.
The opportunity cost of capital has also increased. The budget dedicated to servers, networking, storage, licenses, and staff can’t be simultaneously allocated to AI projects, product development, or market expansion.
This doesn’t make public cloud the cheapest option. Its effective rates can also rise via additional services, data transfers, consumption commitments, or oversized architectures. What changes is that owning infrastructure no longer compares favorably based on old hardware cycle low prices.
Ownership remains optimal when workloads are stable, utilization is high, and operational capacity exists. It loses appeal when demand is uncertain, hardware can become obsolete quickly, or the organization needs to grow without capital lock-in.
Stackscale’s experience: Repatriation doesn’t usually mean returning to the office
Stackscale’s accumulated experience in private cloud projects and migrations shows that few companies want to regain full control of a traditional data center.
Clients considering partial or full exit from public cloud typically seek three things: more predictable costs, greater control over infrastructure, and a clearer link between resources contracted and performance achieved.
But they don’t want to handle power supply, cooling, connectivity, component replacement, physical security, or late-night interventions themselves.
In many cases, the alternative ends up being dedicated infrastructure hosted in a professional data center. The company retains reserved resources and a designed architecture, while delegating physical operation and some management tasks.
This pattern repeats especially for databases, virtualization platforms, persistent storage, and enterprise applications with stable loads. In these scenarios, the nearly unlimited elasticity of hyperscalers may offer less value than the predictable monthly cost of a private cloud.
There are also situations where public cloud remains the rational choice—such as temporary environments, highly variable peaks, global deployments, managed services hard to replicate, or projects with unknown demand.
Therefore, responsible migration shouldn’t start from ideological positions. It should begin with an inventory: workloads, dependencies, historical consumption, traffic, regulatory requirements, and total operating costs.
Total cost includes more than CPU, memory, and storage. It must also account for licenses, connectivity, backups, disaster recovery, monitoring, support, technical team time, and hardware refreshes.
Repatriation can reduce a high bill, but poorly planned architecture merely shifts visible costs into harder-to-measure expenses.
Scarcity favors those already operating at scale
Rising hardware costs impact buyers differently. Hyperscalers, colocation providers, and private cloud vendors can negotiate volume discounts, plan purchases, and distribute infrastructure costs across multiple projects.
A company that replaces servers every four or five years has less flexibility, potentially renewing during a price peak and bearing the full investment in a single cycle.
This reinforces an intermediate option between public cloud and in-house data center: dedicated infrastructure managed by specialized providers.
It avoids purely variable consumption of hyperscalers but doesn’t require building and operating a complete physical setup. Additionally, it allows combining dedicated resources for stable workloads with public cloud services for elasticity needs.
AI adds another layer to this decision. In 2026, executive focus is heavily on projects integrating AI into products and processes. Traditional server upgrades compete for budget with initiatives backed by stronger internal support, though their returns may still be unproven.
This can delay economically justified repatriations—not because the calculations are wrong, but because companies prioritize other projects and have limited capacity for complex changes.
37signals succeeded because it turned infrastructure into a product decision and handled the technical work themselves. Not all organizations want or can do the same.
Their case demonstrates that public cloud isn’t the only destination for every workload. It also shows that leaving the cloud requires knowledge, time, discipline, and a prepared alternative platform.
Repatriation is never a one-way destination. It’s a negotiation and design tool. A company can use it to cut costs, regain control, or pressure its provider without building a new data center. Often, the final outcome will be hybrid: public cloud for variable workloads, dedicated infrastructure for stable ones, and colocation for property without physical complexity.
Frequently Asked Questions
How much has 37signals saved by leaving the cloud?
37signals states that their accumulated savings will surpass $10 million over five years. This forecast is based on avoided costs and results observed since the migration began, not a universal figure applicable to all companies.
Is public cloud usage decreasing?
Not worldwide. Cloud infrastructure spending grew 35% year-over-year in Q1 2026, although some organizations are shifting specific workloads to private cloud, dedicated servers, or colocation.
What workloads are better candidates for repatriation?
Stable, predictable workloads, storage-intensive applications, or those with high utilization over long periods are good candidates. Temporary or highly variable environments may still benefit from flexible public cloud consumption.
Does repatriation mean installing servers on-site?
Not necessarily. Many organizations transfer workloads to dedicated infrastructure or private cloud hosted in professional data centers, maintaining control and predictability without operating the physical installation.
Sources:
- 37signals, publications, and podcasts discussing their cloud exit and projected savings exceeding $10 million.
- 37signals, Moving Mountains of Data off S3, 08/01/2026.
- Synergy Research Group, global cloud infrastructure service expenditure during Q1 2026.
- TrendForce, DRAM and NAND Flash price forecasts for Q1 2026.
- Digital Realty, Q2 results and forecast review for 2026.
- Stackscale, experience with private cloud projects, dedicated infrastructure, and cloud migrations.
