Cloud repatriation: when moving workloads back from public cloud makes sense, and for which ones

Cloud repatriation: what it is and when moving back makes sense

Cloud repatriation means moving workloads or data from a public cloud back to your own data centre, to rented space in a colocation facility, or to a private cloud hosted for your organisation alone. It can make sense for workloads with a steady load, large volumes of data leaving the cloud, or firm requirements on where data is stored and who operates the platform. It rarely makes sense for services built on a provider’s managed platform, because moving them means rewriting them. A move to a regional provider’s IaaS is strictly a change of provider, but the decision follows the same steps.

An IDC blog post of 28 October 2024 called repatriation “not a wholesale migration” and, citing IDC’s Server and Storage Workloads Survey, put the share of companies that plan full workload repatriation at 8 to 9%. Moving selected workloads leaves a hybrid estate, and the decision is made workload by workload, on measured utilisation and invoice lines.

Why companies move workloads back from public cloud

Public cloud capacity is provisioned and released in line with demand and metered as it is used; NIST’s definition of cloud computing (SP 800-145) lists rapid elasticity and measured service among its essential characteristics. A workload that runs at the same level all month makes no use of that elasticity, so the comparison becomes metered use at a constant level against fixed capacity sized to its measured peak.

Data transfer adds a cost line that grows with the data leaving the cloud: replies to users, copies to on-premise systems, backups and restores. The Data Act, Regulation (EU) 2023/2854, defines data egress charges as fees for extracting data through the network to another provider’s system or to on-premises infrastructure (Article 2(35)), and for the move itself egress counts among the switching charges it limits.

Licence terms can change the result in either direction. Software counted per physical core, per host or per processor may have separate rules for shared cloud hardware, and licences supplied with cloud instances are used only through the provider’s contract, so they do not move with the workload.

Contracts, customers and sector rules may require data to stay in the EU, or to be held by an operator under EU jurisdiction. Location and jurisdiction are separate questions, and our article on the CLOUD Act and EU data residency explains what each of them changes.

AI inference with a high, steady token volume keeps dedicated GPUs busy, and the prompts and documents it processes may have to stay on systems you control. Use limited to office hours leaves GPUs idle for most of the week, so whether dedicated GPUs cost less than a cloud API depends on token volume and utilisation; our comparison of a private LLM and a cloud API shows how to calculate the break-even with your own numbers.

Cloud repatriation surveys and forecasts: Flexera, IDC and Gartner

The surveys and forecasts below rest on different questions and samples, so their figures cannot be added up. Flexera’s 2026 State of the Cloud Report, published on 18 March 2026, is based on a survey of more than 750 cloud decision-makers and users. It found 73% of organisations running hybrid environments, and Flexera’s German release of the same day adds that only 14% rely on public cloud alone. Managing cloud spend remained a top challenge for 85% of respondents. In Flexera’s figures, wasted cloud spend rose to 29%, the first increase in five years, which it attributes to surging cloud-based AI workloads.

IDC’s blog post of 28 October 2024 lists cost management, performance optimisation, security and compliance, and operational control as drivers of repatriation. An IDC Survey Spotlight of March 2026 examines how far AI integration drives repatriation; its abstract says certain workloads, particularly those involving high-value models and sensitive data, are “being drawn back on-prem”.

Gartner uses the term geopatriation for moving company data and applications out of global public clouds into local options, such as sovereign clouds, regional cloud providers or an organisation’s own data centres, “due to perceived geopolitical risk”. In a press release of 20 October 2025 it predicted that more than 75% of European and Middle Eastern enterprises will geopatriate their virtual workloads by 2030. It put the share in 2025 at less than 5%.

Which workloads to move back and which to keep in public cloud

Our reading; the backup and AI inference rows also rest on IDC’s blog post of 28 October 2024 and IDC Survey Spotlight US54381526 (March 2026), the serverless and managed service rows on the cases our EU Cloud page names as not a fit.

A virtual machine moves with its operating system and applications, after a disk conversion if it ran as a native cloud instance. Code written for a provider’s functions, queues or proprietary databases has to be rewritten first, and that effort can outweigh the saving. Our EU Cloud platform is built for classic workloads and is not the right fit where a company needs many of a global hyperscaler’s managed services, such as serverless, ML platforms or exotic databases.

On the first call we work through your workloads and requirements, and you leave with two or three configuration options and an indicative monthly invoice. Send us the workloads you would move first, with their CPU, RAM and storage use.

How to compare the cost of a workload before moving it back

The comparison holds only if both sides cover the same period and the same scope. Rightsize the cloud side before you compare, because an instance larger than its load needs makes the cloud look dearer, and a smaller one may close part of the gap without any move. For steady load, also price the capacity commitments your provider offers.

1. Collect twelve months of the workload’s invoice lines for compute, storage, data transfer, licences, support and capacity commitments.
2. Export utilisation for the same months: CPU, RAM, storage growth, IOPS and outbound traffic, with peaks as well as averages.
3. Size the destination on the measured peaks plus headroom for growth and maintenance, not on the instance sizes in use.
4. Cost the destination over the same term: hardware and its depreciation or a monthly hosting invoice, licences to buy, power, space, backup, a second site if needed, and staff time or a support contract.
5. Add the one-off costs of migration work, the parallel run, rebuilding what depended on managed services and transfer charges for the move.
6. Decide per workload, and move first those where the difference is clear and the effort small.

Where repatriated workloads go: on-premise, colocation or hosted cloud

NIST SP 800-145 (September 2011) for what the IaaS consumer controls; our EU Cloud page for the private cloud, IaaS and hybrid rows; the first two rows are our reading.

With your own data centre or colocation, the capital cost, the hardware refresh and the people to run it come back to you, and in your own building the space, power and cooling as well. The hosted options keep the hardware with the provider, and our comparison of IaaS, private cloud and hybrid sets out the criteria between them.

Leaving public cloud under the EU Data Act: the January 2027 date

The Data Act’s definition of switching in Article 2(34) covers a move from one provider’s data processing service to another provider and a move to “on-premises ICT infrastructure”. Since 12 September 2025 the cloud contract must set out the switching terms in writing, with a notice period of at most two months and a transitional period of at most 30 calendar days for the move, followed by at least 30 calendar days for data retrieval (Article 25). From 11 January 2024 to 12 January 2027 providers may impose reduced switching charges, data egress included, capped at their costs directly linked to the switching process; from 12 January 2027 they may impose none (Article 29). As of 6 October 2026 these rules apply as published in 2023; the Digital Omnibus proposal to amend parts of them has not been adopted.

Standard service fees and early termination penalties are not switching charges under Article 2(36), so check minimum terms and commitments before you fix a date; for a large data move, the date also decides whether reduced switching charges can apply at all. Article 25(2)(a) requires a contract clause that lets you port “all exportable data and digital assets” to on-premises infrastructure; for a partial repatriation, which keeps other workloads with the provider, ask in writing which charges apply. Whether moving only some workloads counts as switching, and whether an exception such as Article 31 for custom-built services applies, is a legal assessment for your legal department. Our guide to Data Act cloud switching covers the periods, the exceptions and the cloud exit strategy.

Migration method and the costs that stay after moving back

VMs that already run on a VMware-based service can use the methods in our guide to migrating VMs to a VMware cloud, where the provider gives the access they need. Native instances of a public cloud use the provider’s virtualisation and image formats, so they need an export and conversion of their disks, where the provider supports it, and drivers for the new virtual hardware in the guest, or a rebuild with the data copied across. Either way, an application moves in the same maintenance window as its data.

Several costs remain after the cloud invoice shrinks. Services that depended on a provider’s managed platform are replaced with software you run, such as a database cluster with its own backups and failover, message queues and monitoring. Your staff take on hypervisors, storage, networks, patching, restore tests and on-call duty, unless a provider does this under contract. Bought hardware needs a refresh cycle and spare capacity for maintenance, and a workload that ran across several zones of a provider needs a second site to keep that protection. During the move both environments are paid for, and a hybrid estate keeps a cloud invoice for the services that stay and the data flowing between the two.

Our technical assessment reviews the systems and the target configuration and delivers a step-by-step migration plan, at a price fixed before work begins. Describe the systems you would bring back in the form below, with their dependencies.

What we do

Our EU Cloud service hosts IaaS on the VMware vSphere platform, with CPU and RAM guaranteed and not shared with other clients, or a dedicated private cloud, in Baltneta’s Tier-3 data centres in Lithuania (ISO 27001, PCI DSS), with data and backups that stay in the EU and a transparent monthly invoice in euros. Hybrid, with some systems staying with you and replication between the sites, is a standard scenario. Our engineering partner Vixen.UNO handles migration and support, moving systems step by step in agreed maintenance windows with a rollback plan; where systems are deeply tied to one global cloud’s proprietary services, we first calculate whether moving them makes sense. For AI inference, our Private AI/ML service runs models on-premise or on dedicated hardware in a Tier-3 data centre in Lithuania and includes a TCO calculation against cloud GPUs before the purchase. The first call is free of charge, and the price of the technical assessment is fixed before work begins.

FAQ