Companies are increasingly reassessing their reliance on public cloud infrastructure, with many reporting cost savings by disengaging from or avoiding public cloud services altogether. Data repatriation from the cloud is a response to rising costs and concerns about data processing in the context of AI. ZDNET spoke with executives to find out how - and if - this "repatriation" works in practice. The company 37signals moved to public cloud in 2006 after the launch of Amazon S3. It began reducing its workloads on AWS in 2022 after having to deal with annual bills amounting to several million dollars during cloud usage peaks. In 2023, it had completely replaced its AWS and Google Cloud servers with its own hardware, running on open source platforms such as KVM, Docker and Kamal. The company stated it had spent half a million dollars on server purchases to replace more than 3.5 million dollars in annual cloud-related bills. Instead of acquiring physical data centers, it rented colocation spaces to host its servers. These cost concerns are not unique to 37signals. Rebecca Weekly, vice president of infrastructure at Geico, said the company's cloud bills had been multiplied by 2.5 after a migration to public cloud servers, notably Azure. While a complete exit from public cloud remains rare, many companies are transferring at least part of their workloads to a private infrastructure. The Ahrefs software company also claims to have never used public cloud providers such as AWS and estimates having saved over 400 million dollars since 2020. Large-scale surveys conducted by Barclays, IDC and Cast AI among cloud users confirm that these are not isolated examples. In general, it takes more than a single shocking cloud bill to trigger this transition. "The most common trigger is a contract renewal," explains Spencer Kimball, CEO and co-founder of Cockroach Labs. According to Kimball, it is usually a member of the finance team who reports the problem when they make multi-year cost projections to evaluate the amount of bills as workloads continue to increase. In 2025, Cast AI compared the cloud computing expenses of 2,100 organizations to their actual CPU and RAM usage. The company found that Kubernetes clusters on AWS, Google Cloud and Azure used, on average, 10% of the provisioned CPU resources and 23% of the memory. Kunal Agarwal, CEO of Unravel Data, explained that this unexplained gap between the resources paid for by the company and those actually used is the first thing that finance teams notice on the invoices. "When the team looks at the situation more closely, they find that the invoice has increased much more rapidly than the workloads," he said, adding that when comparing these figures to material costs, they often find that "the same tasks and the same data have generated a much higher invoice." Teams can also discover invoice items that make no sense, such as cloud egress fees and inter-regional transfer fees, or platform licensing costs that do not correspond to the company's actual usage. The UK Competition and Markets Authority (CMA) is currently investigating Microsoft, citing unfair financial barriers that companies face when trying to switch to competing cloud providers. Sirish Raghuram, director of growth at Platform9, confirmed this disparity in cloud exit costs. His own company transferred about 40% of its workloads on-site following a spike in cloud resource costs in December 2025. He also found that workloads least sensitive to service level agreements (SLA), such as development, testing and lab environments, cost 70% less when executed locally. However, regardless of cost concerns, not all workloads can be safely transferred out of the public cloud infrastructure. Experts agree that predictability is the main criterion that determines if a repatriation is prudent, which explains why complete exits are so rare. Almost all companies have at least some workloads that depend on the flexibility needed to adapt to demand, to which public cloud deployments are particularly well suited. "If you can predict a workload's demand a year in advance and exploit it at a high utilization rate, it can be more economical to own your infrastructure," explained Kimball. "1,000 virtual machines (VM) with 150 terabytes of monthly bandwidth cost $60,000 per month on a large public cloud." Andreessen Horowitz published a report in 2021 that quantified the gap between costs and usage. Among the 50 largest publicly traded software companies it analyzed, about $100 billion in market capitalization was wasted on cloud bills across these companies. Extrapolating these figures to the entire public software market, Andreessen Horowitz estimated the actual amount of waste to be at least $500 billion, compared to the cost that would have been represented by running the same workloads on on-site hardware. Todd Robinson, president of OpenMetal, said the numbers are beginning to favor private cloud earlier than most people think. "1,000 virtual machines (VM) with 150 terabytes of monthly bandwidth cost about $60,000 per month on a large public cloud. The same deployment on a managed private cloud costs rather $16,000 per month," he explained. "That represents more than half a million dollars per year." Audits reveal the cases where migration is prudent. However, Agarwal points out that many resources are wasted in the public cloud. Simply turning off unused cloud servers can reduce costs, which may be sufficient for many companies. "Many of these discussions stop as soon as an organization calculates the concrete cost of its cloud bill, because the savings achieved by eliminating waste often correspond to what the migration would have allowed to save, without carrying its risks," he said. Martin Summerhayes, head of technical services at Northdoor PLC, a computer consulting firm, is even more direct. He believes that many cloud repatriation projects are actually disguised FinOps projects, because companies are simply correcting course from previous large-scale digital transformation initiatives that were not cost-optimized. "The cheapest repatriation is the one you never need," he told ZDNET. The cost of migration may outweigh the savings you could realize. In other words, unless you can identify a specific problem that the cloud exit would allow you to solve, the cost of the migration may outweigh the savings you could realize. "The cloud bill disappears, and in its place appear a salary mass, a colocation contract and a hardware renewal scheduled for 2031," adds Summerhayes. He recommends doing a three-year cost analysis of your public cloud infrastructure, taking into account licenses, number of users and resources, to determine if a transfer would actually be beneficial. Known examples seem to corroborate these observations. The Zynga gaming company spent $100 million on on-premise to exit AWS starting in 2011. A few years later, the company announced the closure of its data centers and its return to AWS, because it depended on a scalability that led to high hardware acquisition costs during periods of high demand. If data loss could have serious operational or legal consequences, keeping data in the cloud is a much safer financial option. "Unused resources, limited optimization and difficult-to-calculate egress fees are some of the reasons why companies may not benefit from the cloud," says Uwe Geuss, DT of Retarus. He also explains how some workloads tend to show mediocre performance when they are migrated out of the cloud. "Despite the renewed interest in repatriation, the cloud remains the preferred deployment model for many critical communication workloads," he says. In essence, wherever data loss could have serious operational or legal consequences, keeping data in the cloud is a much safer financial option. Summerhayes recounted his own experience within a company that insisted on moving from its own data centers to a private cloud provider, then returning to on-site hardware, and finally opting for public deployments using Microsoft Azure. Each migration was triggered by internal frustrations without tangible evidence to support them, which explains why they always ended up costing the company more than expected. Beyond the increase in cloud-related bills, data sovereignty is quickly becoming another major factor in decision-making. As many cloud providers are now pushing for AI integration at the platform level, companies are beginning to realize what they are consenting to in exchange for the convenience of the cloud. Summerhayes refers to Microsoft's own statement before the French Senate that the company cannot prevent American authorities from accessing its customers' data due to the CLOUD Act, a topic that comes up often during his meetings with his clients. In the UK, the British Treasury has already designated AWS, Google Cloud and Microsoft as critical third parties as of July 13, 2026. This means that these organizations will now be subject to stricter compliance requirements in terms of cybersecurity and data resilience. Nevertheless, AI training and security concerns have been at the heart of the concerns of enterprise security officers tasked with evaluating cloud-related risks. "Cost remains an important factor, but governance, regulatory requirements and data sovereignty are increasingly part of infrastructure decisions," said Michael Lau, senior product manager at IBM Apptio. He told ZDNET that he recommended considering governance as a design criterion from the start, then relying on economic criteria to determine the best option among competing platforms. Agarwal and Kimball also explained how the issue of sovereignty has evolved beyond the simple physical storage location of data to include: What these AI models can access, How these data are used, How long they are retained by the cloud provider. AI-related workloads are a major driver of current cloud repatriation projects. AI-related workloads, in particular, constitute a major driver of current enterprise data repatriation projects to the cloud. Cloudera surveyed 1,500 data and enterprise infrastructure managers at the beginning of the year and found that 66% of them had repatriated at least part of their AI workloads from public cloud to self-hosted or on-site solutions. "Many companies are becoming increasingly reluctant to permanently transfer a large part of their enterprise data to external environments for inference and retrieval workflows," says Craig Cook, VP in charge of cloud and data protection at Flexential. Cook also added that feeding AI models with large volumes of enterprise data can become very token-intensive in the long run, which often explains a large part of the increase in cloud bills. Robinson notes that the rise in material costs has affected the information economy, both in the cloud and locally. The price of NVMe disks has increased by 223% over the past three years, while the price of 64 GB DIMM DDR5 memory sticks has risen by 474% compared to two years ago. When asked what a CIO or technical director should focus on first before transferring a workload out of the cloud, the experts agreed that flexibility requirements remained the main determining factor. "The cloud transforms infrastructure into a variable expense," said Lau. "Ownership becomes easier to justify when demand is sufficiently predictable to support a long-term commitment."