The Cloud Repatriation Movement: Why Companies Are Leaving the Cloud and Saving Millions

A growing wave of companies is moving workloads out of the public cloud and back onto owned or colocated infrastructure — a trend known as cloud repatriation. While the cloud computing market continues to grow (approaching $700 billion in 2026), the narrative that every workload belongs in the cloud is being challenged by organizations that have done the math and concluded that, for their specific workloads at their specific scale, owning infrastructure is significantly cheaper. The most prominent example is 37signals (the company behind Basecamp and HEY), which publicly documented saving over $7 million over five years by leaving the cloud. But they’re not alone: a 2025 survey by Andreessen Horowitz found that 72% of large enterprises have repatriated at least some workloads from public cloud to on-premises or colocation environments.

The Economic Inflection Point

Cloud computing’s economic proposition is strongest at small to medium scale and for variable workloads. When you’re a startup burning through investment capital and don’t know what your infrastructure needs will look like in six months, the cloud’s pay-as-you-go model is transformative. You can spin up servers in minutes, scale globally without building data centers, and avoid the capital expenditure of hardware purchases. The cloud effectively converts capital expenditure (buying servers) into operational expenditure (monthly cloud bills), which is favorable for companies in growth mode.

The economics shift at steady-state scale. When workloads are predictable — running 24/7 with consistent resource requirements — the cloud’s flexibility premium becomes waste. Cloud providers price their services to cover infrastructure costs, operational costs, and profit margins (estimated at 30-50% for major providers). An organization running the same servers continuously is paying that margin forever. At a certain scale, it’s cheaper to buy the servers, colocate them in a data center, and run them for their 3-5 year useful life.

David Heinemeier Hansson of 37signals documented this analysis in granular detail. Basecamp and HEY were running on AWS with an annual cloud bill of approximately $3.2 million. The equivalent owned infrastructure (Dell servers, colocated at Deft data centers) cost approximately $600,000 in annual depreciation and colocation fees — a savings of roughly $2.6 million per year. Even after accounting for the salary of three operations engineers to manage the infrastructure, the savings exceeded $1.5 million annually.

The key variable in this calculation is utilization. Cloud instances that run steadily at 60-80% utilization are significantly more expensive than owned hardware running at the same utilization. Cloud instances that spike to 1000% of baseline at unpredictable times and sit idle 90% of the day may be cheaper in the cloud because owned infrastructure would need to be sized for peak demand and sit mostly idle. The breakeven point varies by workload type, but as a rough guideline: workloads that run consistently above 40-50% utilization for more than three years are usually cheaper to run on owned infrastructure.

What’s Being Repatriated (and What’s Staying)

Cloud repatriation is selective — organizations aren’t ripping everything out of the cloud. The pattern is to repatriate “boring” workloads with predictable resource requirements while keeping dynamic, variable, and managed services in the cloud.

Compute-intensive workloads with steady demand are the prime repatriation candidates. Application servers running web backends, API servers, and business logic at consistent scale are straightforward to run on owned hardware. These workloads don’t benefit much from cloud elasticity because they don’t experience significant traffic variability, and the compute costs at cloud pricing accumulate to substantial sums.

Data storage and databases with large, growing datasets are frequently repatriated. Cloud storage costs (while individually cheap per GB) compound at scale, and egress charges (the cost of transferring data out of the cloud) can be punishing for data-intensive applications. Companies with petabytes of data in S3 or Azure Blob Storage may find that equivalent storage on-premises costs a fraction of the cloud price, especially for data that’s accessed frequently enough that archive-tier pricing doesn’t apply.

AI training workloads are a special case. GPU cloud instances are expensive ($2-$8 per hour for a single high-end GPU, $12,000-$32,000 per month for a dedicated GPU server), and training runs consume them continuously for weeks or months. Companies training AI models regularly find that purchasing GPUs (even at NVIDIA’s premium pricing) pays for itself within months compared to cloud GPU rental. CoreWeave, Lambda (the GPU cloud company), and other GPU-focused providers offer significantly cheaper GPU cloud options than AWS/Azure/Google, but owned hardware is often cheaper still for organizations with consistent GPU demand.

What stays in the cloud includes: bursty workloads that need elastic scaling (marketing campaign backends, event-driven processing, seasonal retail traffic), managed services that would be expensive to replicate on-premises (managed Kubernetes, managed databases with automatic failover, serverless functions), global edge services (CDN, edge compute, global load balancing), and innovation workloads where rapid provisioning of diverse hardware is valuable (ML experiments using different GPU types, testing with exotic database engines, proof-of-concept deployments).

The Modern On-Premises Stack

Cloud repatriation in 2026 doesn’t mean returning to the pre-cloud era of racking servers in closets and managing operating system installations manually. The modern on-premises stack uses the same cloud-native technologies (containers, Kubernetes, Terraform, GitOps) that organizations use in the cloud, running on owned or colocated hardware.

Server hardware has never been more capable or more cost-effective. A single Dell PowerEdge R760 server with two Intel Xeon processors, 512GB of RAM, and 30TB of NVMe storage costs approximately $20,000 — less than two months of renting an equivalent cloud configuration. Modern servers include built-in hardware management (iDRAC, iLO) that enables remote management, monitoring, and troubleshooting without physical access. Redundant power supplies, hot-swap drives, and ECC memory provide hardware-level reliability.

Colocation data centers provide the physical infrastructure (power, cooling, network connectivity, physical security) that organizations don’t want to manage. A colocation rack in a Tier 3 data center costs $1,000-$3,000 per month depending on location and power density, and can house 10-40 servers. The data center provides redundant power with UPS and generators, redundant cooling, multiple network carriers, 24/7 physical security, and SLA-backed uptime guarantees. Colocation effectively outsources the parts of infrastructure management that aren’t core competency while retaining control of the hardware and software.

Infrastructure-as-code tools (Terraform, Ansible, Puppet) automate the provisioning and configuration of on-premises hardware with the same declarative approach used for cloud resources. MAAS (Metal as a Service) from Canonical and Tinkerbell from Equinix enable automated bare-metal provisioning that boots servers from PXE, installs operating systems, and configures software without manual intervention. Combined with container orchestration (Kubernetes or simpler alternatives) and GitOps workflows (Flux or ArgoCD), modern on-premises infrastructure can achieve the same operational speed and reliability as cloud infrastructure — it just requires more upfront expertise to set up.

The Hybrid Compromise

Most organizations pursuing cloud repatriation end up with a hybrid architecture: a stable base of owned or colocated infrastructure handling predictable workloads, with cloud resources providing elastic overflow for traffic spikes and hosting managed services that are impractical to replicate on-premises.

This hybrid model requires robust connectivity between on-premises and cloud environments. Direct connect services (AWS Direct Connect, Azure ExpressRoute, Google Cloud Interconnect) provide dedicated, high-bandwidth connections between colocation facilities and cloud providers, typically at 1-100 Gbps with consistent sub-10ms latency. These connections eliminate the public internet’s variability and provide the performance needed for hybrid architectures where data and traffic flow between environments.

Kubernetes running in a hybrid configuration — some nodes on-premises, some in the cloud — is technically possible but operationally complex. Most organizations run separate clusters in each environment and use service mesh or API gateway technology to route traffic between them. The complexity of hybrid Kubernetes is one of the reasons some organizations use simpler orchestration tools (Nomad, Docker Swarm) for their on-premises workloads while keeping cloud workloads on managed Kubernetes.

Who Should Consider Repatriation

Cloud repatriation makes economic sense for organizations that meet several criteria: predictable, steady-state workload profiles; significant monthly cloud spend (generally $50,000+ per month before repatriation becomes worth the effort); sufficient engineering talent to manage infrastructure (or the willingness to hire); and workloads that don’t heavily depend on cloud-specific managed services that would be difficult to replicate.

Organizations that should stay in the cloud include: early-stage startups that need to minimize operational overhead and maximize development speed; companies with highly variable workloads that benefit from elastic scaling; companies without operations engineering expertise and no intention to build it; and organizations that depend heavily on managed cloud services (machine learning platforms, managed databases, serverless functions) that provide functionality too complex to replicate on-premises.

The cloud computing industry’s marketing has long promoted the narrative that the cloud is universally the best choice for every workload. The repatriation trend reveals a more nuanced reality: the cloud is the best choice for some workloads, at some scales, for some organizations. For others, the economics of owned infrastructure are compelling, and the operational tools needed to manage on-premises deployments have matured to the point where the cloud’s convenience advantage has significantly narrowed. Running your own servers in 2026 is nothing like running your own servers in 2006 — and the economic argument for doing so is strong enough that even cloud-native organizations are reconsidering the assumption that everything belongs in the cloud.