Cloud Repatriation in 2026: When Moving Off the Public Cloud Pays

Every few months a company announces it has left the cloud and saved millions, and every board asks its CTO whether it should do the same. The honest answer depends on the workload. The public cloud is still growing quickly, the companies that left publish their own numbers, and AI has added a new kind of workload with its own economics. This analysis reconciles the headline statistics, looks at what the named exits actually moved and saved, explains how GPUs and power change the decision, and sets out the rules and license terms that now shape any move.

• 8 to 9%of companies plan a full repatriation of their workloads; most move selected data and services²
• 43%growth in cloud infrastructure revenue in the second quarter of 2026, the fastest in eight years³
• 29%of cloud spend is wasted by respondents' own estimate, the first rise in five years¹²

What the repatriation statistics actually count

The most quoted number is Barclays' finding that 83% of enterprise CIOs plan to move workloads back from public cloud. The survey counts companies. As Channelnomics points out, if 83% of enterprises move a single workload back on-premises, the figure is still true.¹ IDC's survey of server and storage workloads gives the better measure of leaving the cloud: only 8 to 9% of companies plan full repatriation, and most move specific pieces such as production data, backup and disaster recovery, compute or development environments.²

The market moves the other way at the same time. Synergy Research puts cloud infrastructure revenue at $143.4 billion in the second quarter of 2026, up 43% on a year earlier, its fastest growth in eight years.³ Uptime Institute's 2026 survey found colocation and cloud providers holding 46% of enterprise IT workloads against 44% in enterprise-owned data centers, the first time third-party venues have led, with the gap expected to widen to 48% against 42% by 2028.⁴ Repatriation is a counter-current inside a larger move off-premises, and placement is decided workload by workload.

Waste muddies every comparison. Flexera's 2026 respondents estimate 29% of their cloud spend is wasted, up from 27% after five years of decline.¹² IDC found close to half of cloud buyers spent more than they expected in 2023.² A wasteful estate moved to owned hardware takes its waste along, so the first comparison worth making is against a well-run cloud bill. FinOps practice has widened to match: 57% of FinOps teams now cover private cloud and 48% cover data centers, which puts every venue on one ledger.¹³

What the named exits moved and saved

The public cases are few, mostly self-reported and unusually detailed. Only Dropbox's figures come from an audited filing. The table lists what each company moved and what it says it gained.

The pattern is consistent. Storage and data transfer are the recurring breaking point: 37signals' S3 bill, Dropbox's user files, Railway's egress charges, and GEICO's engineering lead's verdict that storage is one of the most expensive things you can do in the cloud.⁷ The leavers ran steady loads, and 37signals says plainly that a business facing regular spikes of five to ten times its baseline may be a good candidate for the cloud.⁶ Most exits were partial or kept a way back: Dropbox kept AWS for part of its storage, and OneUptime keeps a standby cluster that can start in under ten minutes.¹⁴¹⁷

The counter-cases make the same point from the other side. Netflix spent seven years rebuilding its systems for the cloud before closing its last data center in January 2016, and found its cost per streaming start ended up a fraction of the data center's.²⁰ Capital One closed eight data centers and rebuilt thousands of applications, valuing the ability to scale up and down as demand moved.²¹ Stack Overflow moved to Google Cloud in 2025 because its data center contract ended with no option to renew.²² GEICO's lesson runs the other way: ten years into a cloud migration that moved legacy systems largely as they were, its bills had gone up 2.5 times.⁷ Architecture decides more than venue.

Andreessen Horowitz's 2021 essay remains the reference for the scale argument. It reported committed cloud spend averaging about half of cost of revenue across the public software companies it studied, and practitioners converging on repatriated workloads costing a third to a half of their cloud equivalent.²³ That applies to software companies whose cloud bill is a large share of what they sell. Few enterprises look like that, and none of the leavers publishes the salary cost of its own operations team beside its savings.

AI and GPUs change the question

AI workloads are moving too. In a Cloudera-sponsored survey of 1,500 infrastructure and data architects in June 2026, 66% said their organizations had moved AI workloads from public cloud back to private cloud or on-premises in the past year, and 84% reported higher infrastructure costs from AI.⁸ Cloudera sells hybrid data platforms, and the question counts any single workload, so the direction matters more than the size. The reasons matter most.

Utilization is the hidden variable. Cast AI, which sells optimization software, measured average GPU utilization of 5% across tens of thousands of cloud Kubernetes clusters, against 8% for CPU.⁹ An idle rented GPU costs money by the hour, and an idle owned one costs capital, so the first saving comes from scheduling and right-sizing in whichever venue the GPUs sit. Andreessen Horowitz's 2023 rule of thumb still frames the ownership case: the cloud is right in most cases, and building your own data center typically takes infrastructure spend above $50 million a year, hardware the clouds lack, or geopolitical constraints.²⁴

Prices no longer move in one direction. AWS cut on-demand prices for its H100 instances by 44% in June 2025, then raised prices for reserved GPU capacity blocks by about 15% in January 2026.²⁵²⁶ Nebius reports that roughly 70% of its deals now include prepayment, which it calls the market standard for securing capacity.²⁷ Space and power are scarcer still. CBRE puts data center vacancy across North America's primary markets at 1.4%, with more than 80% of capacity under construction already leased, and the IEA expects data center electricity use to more than double to about 945 TWh by 2030, with around a fifth of planned projects at risk of delay from grid constraints.²⁸²⁹ Owning AI capacity now means booking power and space well ahead, which favors steady, well-forecast inference over bursty experiments.

Exit fees are falling, and lock-in moved to licenses

Leaving has become cheaper on paper. Google, AWS and Microsoft each waive data transfer fees for customers who leave, under conditions the UK competition regulator has documented: apply to support before moving, finish within a fixed window, and delete the data or close the account afterwards, with no credit for ongoing multicloud use.³⁰ AWS extended its window to 90 days in 2025.³¹ 37signals timed its storage exit to its contract end and had about $250,000 of egress fees waived.¹⁹

The Data Act's guarantees are narrower than the headline. Functional equivalence applies to infrastructure services, providers need not build new technology or hand over intellectual property, and egress for workloads running in parallel across providers may still be charged at cost.¹⁰ Google has chosen to make parallel transfers free for EU and UK customers.³⁸ Proprietary managed services remain the real lock-in, because no rule makes them portable.

Licenses now decide more business cases than egress does. Since October 2019, Microsoft licenses bought without Software Assurance cannot be used on dedicated hosts at Microsoft, Alibaba, Amazon or Google, so moving Windows Server or SQL Server between clouds can raise the license bill even when the data moves free.³⁹ On-premises got dearer too. Broadcom raised VMware's minimum subscription from 16 to 72 cores in April 2025, with a 20% penalty for late renewal.¹¹ AT&T told a court it faced a proposed 1,050% annual increase and put the cost of migrating off VMware at $40 to $50 million.⁴⁰ The European cloud lobby CISPE complained to the European Commission in March 2026 that prices have risen tenfold since the acquisition.⁴¹ A repatriation plan built on pre-2024 virtualization prices needs repricing before anyone signs.

Where each kind of workload belongs

How to run the decision

1. 1Optimize in place first Rightsize, commit, schedule and shut down idle resources, GPUs included, so the comparison is against a well-run cloud bill. A wasteful estate moved elsewhere stays wasteful.
2. 2Price each workload in full Hardware, colocation, power, network, people on call, licenses, the months of running both venues during the move, and the managed services you would have to rebuild. Dropbox paid for duplicate storage while it migrated.
3. 3Check the licenses before the hardware Model Microsoft and VMware terms in the destination, including core minimums and renewal penalties, before any plan reaches a board.
4. 4Time it to the contract and the window Exit when a commitment ends, apply for free transfer before copying anything, and in the EU plan around the Data Act's notice and transition periods.
5. 5Book capacity early With data center vacancy near 1% in North America's primary markets and GPU capacity sold on prepayment, power, space and hardware need booking well before the move.
6. 6Keep the exit open both ways Containers, infrastructure as code and open data formats keep the next move possible, and a tested standby in the other venue turns an exit plan into something a regulator can see working.

The cloud debate has settled into something more useful than a verdict. The public cloud keeps growing because elasticity and managed services are worth paying for, and specific workloads keep leaving because steady load, heavy storage and scarce GPUs are cheaper to own when a capable team runs them. The CTOs who come out ahead treat placement as a recurring engineering decision with a full price on each option and an exit that works in both directions.

This is how we run cloud and DevOps work: an audit of what each workload costs and needs, waste removed in place, moves planned and timed workload by workload, and infrastructure as code that keeps the next move open. In one engagement our engineers cut a production datastore's cost by 40% by moving it to a better-suited engine on a platform with millions of users.