37signals, the company that owns HEY and Basecamp, shifted to the public cloud in 2006 following the launch of Amazon S3. However, it began pulling back AWS workloads in 2022 after facing annual bills in the millions during peak cloud usage. By 2023, it had fully replaced its AWS and Google Cloud servers with company-owned hardware running on open-source platforms like KVM, Docker, and Kamal.
The company said it spent half a million dollars in one-time hardware costs to replace over $3.5 million in annual cloud bills. Instead of acquiring physical data centers, it rented colocation spaces to store servers without investing in new real estate.
These cost concerns aren’t unique to 37signals. Rebecca Weekly, a VP of infrastructure at Geico, said the company’s cloud bills went up two and a half times over a ten-year migration to public servers, including Azure. SEO software company Ahrefs likewise claimed that it had never relied on public cloud providers like AWS in the first place and estimated it had saved over $400 million since 2020 as a result.
Large-scale enterprise cloud surveys by Barclays, IDC, and Cast AI have confirmed that these are not isolated examples. While a full exit from the public cloud remains rare, many enterprises are shifting at least some of their workloads to private infrastructure. ZDNET spoke to consultants, experts, and business leaders in the industry to determine how these exits typically go, who makes the decision, and when they actually work out.
Sticker shock and unused resources
Typically, it takes more than just one shocking cloud bill to start this transition.
“The most common trigger is a renewal,” said Spencer Kimball, CEO and co-founder of Cockroach Labs. According to Kimball, it’s usually flagged by someone on the finance team when they run multiyear cost projections to see where the bills would land as workloads continue to scale.
In 2025, Cast AI benchmarked the cloud computing expenses of 2,100 organizations against their actual CPU and RAM usage. They found that Kubernetes clusters on AWS, Google Cloud, and Azure used, on average, 10% of provisioned CPU resources and 23% of memory. Kimble confirmed that most of the cloud resources enterprises pay for offer idle headroom that’s never fully utilized.
Also: Who owns AI risk at work? Business and tech leaders can’t agree, PwC survey finds
Kunal Agarwal, CEO of Unravel Data, told us that this unexplained gap between the resources that the company paid for versus what they actually used is the first thing that finance teams flag on invoices.
“When the team digs in, they find the bill grew much faster than the workloads did,” he said, adding that, when comparing these against hardware costs, they often find “the same jobs and the same data created a much bigger invoice.”
Teams may also uncover line items that make no sense, like cloud egress and cross-region transfer charges or platform licensing costs that don’t map to the business’s actual use. Relatedly, the UK Competition and Markets Authority (CMA) is actively investigating Microsoft, citing unfair cost barriers that businesses face when attempting to switch to rival cloud providers.
Sirish Raghuram, chief growth officer at Platform9, corroborated that cloud egress cost disparity. His own company has moved about 40% of its workloads on-premises following a spike in cloud resource costs in December 2025. He also found that the least SLA-sensitive workloads, like developer, test, and lab environments, were 70% cheaper to run locally.
Still, regardless of cost concerns, not all workloads can be moved safely out of public cloud infrastructure. Experts agree that predictability is the primary metric that determines whether repatriation makes sense, which is why full exits are so rare. Almost every company has at least some workloads that rely on the flexibility to scale with demand, which is what public cloud deployments are best suited for.
“If you can forecast a workload’s demand a year out and run it at high utilization, owning the infrastructure can be cheaper,” Kimball explained.
But the reverse is also true. Andreessen Horowitz released a paper in 2021 that put an actual number on the cost versus usage gap. Across the 50 leading public software companies it analyzed, approximately $100 billion in market value was being wasted on cloud bills between all these companies combined.
Extending this math to the broader public software market, Andreessen Horowitz estimated that the actual dollar amount of waste was at least $500 billion, compared to how much it would cost to run the same workloads on local hardware.
While public cloud offers a great starting point for digital transformation, OpenMetal president Todd Robinson said that the math starts to favor private cloud earlier than most people think.
“1,000 VMs (virtual machines) with 150TB of monthly bandwidth runs roughly $60,000 a month on a major public cloud. That same deployment on a managed private cloud is closer to $16,000 a month,” he explained. “That’s over half a million dollars a year.”
Audits reveal where moving makes sense
All that said, Agarwal pointed out that a lot of wasted resources in the public cloud are just that — a waste. Simply deprovisioning unused cloud servers can cut down on costs, which can be enough for many companies.
“A lot of these conversations die the moment an organization actually monetizes its cloud bill, because the savings from cleaning up waste often match what the move would have saved, without the migration risk, ” he said.
Martin Summerhayes, who leads technical services at Northdoor PLC, an IT consultancy firm based in London, put it even more bluntly. He believes that a lot of cloud repatriation projects are actually FinOps projects in disguise, because companies are merely course-correcting from earlier wholesale digital transformation initiatives that weren’t optimized for cost. “More often than not, the cheapest repatriation is the one you never need,” he told ZDNET.
In other words, unless you can point to a specific problem that leaving the cloud will solve, the cost of migration may outweigh any savings you could achieve.
“The cloud bill disappears, and a salary bill, a colocation contract, and a 2031 hardware refresh arrive in its place,” Summerhayes added. He recommends performing a three-year cost review of your public cloud infrastructure by licensing, user volume, and resources to see if a move would actually be beneficial.
Popular examples seem to back up these observations. Social gaming company Zynga spent $100 million on on-premises hardware to pull back from AWS starting in 2011. A few years later, the company announced that it was closing its data centers and returning to AWS, because the company relied on scalability that incurred high hardware acquisition costs during peak demand.
“Unused resources, limited optimization, and hard-to-calculate exit fees are a few reasons organizations may not see a payoff from the cloud,” explained Uwe Geuss, CTO of Retarus. He also explained how certain workloads tend to perform poorly when migrated off-cloud. “Despite the renewed focus on repatriation, the cloud remains the preferred deployment model for many mission-critical communication workloads,” he said.
Basically, anywhere that data loss could have serious operational or legal consequences, keeping things on the cloud is a much safer financial option.
Summerhayes recounted his own experience working with a company that insisted on moving from its own data centers to a private cloud provider, then returning to on-premise hardware, only to eventually land on public deployments using Microsoft Azure. Each migration was triggered by internal frustrations without proper evidence to back it up, which is why they always ended up costing the company more than they had planned.
Data sovereignty and AI concerns
Beyond rising cloud bills, data sovereignty has quickly become another major factor driving decisions. With many cloud providers now pushing for platform-wide AI integration, enterprises are beginning to realize what they’re signing off on in return for the convenience of the cloud.
In 2025, a US federal court ordered OpenAI to preserve every ChatGPT output log, even when regional privacy laws or customer agreements make deletion mandatory. Summerhayes also referenced Microsoft’s own admission to the French Senate that it can’t protect customer data from being handled by United States authorities because of the US CLOUD Act, which often comes up during his client meetings.
In the UK, HM Treasury has already designated AWS, Google Cloud, and Microsoft as critical third parties as of July 13, 2026 — signaling that these organizations will now be held to stricter compliance requirements for cybersecurity and data resilience. Nevertheless, AI model training and security concerns have been top of mind for enterprise security leaders assessing cloud risk.
“Cost remains important, but governance, regulatory requirements, and data sovereignty are increasingly part of infrastructure decisions,” said Michael Lau, a senior outbound product manager at IBM Apptio. He told ZDNET that he recommends treating governance as a design consideration from the very onset, and then using economics to determine the best option between rival platforms.
Agarwal and Kimball also explained how the sovereignty question has evolved beyond where the data is stored physically to cover what these AI models can touch, how that data is used, and for how long it is retained by the cloud vendor.
AI workloads specifically are a major driver of the enterprise cloud repatriation projects happening today. Cloudera surveyed 1,500 enterprise data and infrastructure leaders earlier this year to find that 66% had repatriated at least some of their AI workloads from the public cloud to self-hosted or on-premise solutions.
“Many companies are becoming uncomfortable with the idea of continuously moving large portions of their enterprise data into external environments for inference and retrieval workflows,” said Craig Cook, SVP of Cloud and Data Protection at Flexential. Cook also added that continuously feeding large volumes of enterprise data to frontier AI models can become very token-intensive over time, which often accounts for a large portion of those rising cloud bills.
Robinson also noted that rising hardware costs have affected the information economy, both in the cloud and on-premises. NVMe drives have gone up by 223% in the last three years, while 64GB DDR5 DIMMs are up 474% from two years ago.
That one factor
When asked what a CIO or CTO should focus on most before moving a workload off cloud, the experts agreed that flexibility requirements remain the primary driver. “Cloud turns infrastructure into a variable expense,” Lau said. “Ownership becomes easier to justify when demand is predictable enough to support a long-term commitment.”
That one factor
When asked what a CIO or CTO should focus on most before moving a workload off cloud, the experts agreed that flexibility requirements remain the primary driver. “Cloud turns infrastructure into a variable expense,” Lau said. “Ownership becomes easier to justify when demand is predictable enough to support a long-term commitment.”