The Great Repatriation
Everyone is telling you to leave the cloud. Most of them are selling private infrastructure. Here is what is actually happening, and which workloads genuinely belong back home.
There is a number doing the rounds in every boardroom and every vendor deck this year. It says that 83 percent of enterprises are moving workloads out of the public cloud and back to private or on-premises infrastructure. It comes from a Barclays CIO survey, and it has been repeated by Michael Dell, by Broadcom's leadership at VMware Explore, and by what feels like every infrastructure provider with a sales target (Channelnomics, 2024; WebProNews, 2024).
The number is real. It is also, on its own, close to meaningless. And understanding why is the most useful thing a business leader can do before making a single decision about their infrastructure in 2026.
What the number actually says
The 83 percent figure measures the percentage of enterprises repatriating at least one workload. As the analyst David Linthicum pointed out when the figure first circulated, that is a very different claim from the one most people hear. If 83 of 100 businesses move a single isolated application out of the public cloud, the statistic is satisfied. It tells you nothing about how many workloads moved, how large they were, or whether they were mission-critical or trivial (Channelnomics, 2024).
The contradictory data confirms the point. IDC has reported for several years that somewhere north of 70 to 80 percent of enterprises are repatriating some workloads, yet the same firm found that only around 8 percent are moving entire workloads off the cloud (mrc, 2026). Flexera's 2025 State of the Cloud Report found that roughly 21 percent of workloads have actually been pulled back, even as net new cloud migration continues (Integrate.io, 2026). Public cloud spending, meanwhile, keeps climbing: Gartner forecasts worldwide public cloud expenditure of over $700 billion for 2025, up roughly 21 percent year on year (HyScaler, 2026).
So both stories are true at once. Almost everyone is repatriating something. Almost no one is leaving. This is not a cloud exodus. It is a correction.
The correction is real, and AI is driving it
Strip away the vendor hype and a genuine shift remains, with genuine causes. Three forces are converging in 2026, and the newest of them is artificial intelligence.
The cost mathematics has flipped for steady-state workloads. The lift-and-shift cloud-first wave of 2018 to 2022 left many organisations with compounding, unpredictable bills: egress fees, cross-region transfer charges, and storage costs that grow with every dataset. IDC found that 59 percent of organisations spent more than budgeted on cloud in 2024, and Flexera reports that 84 percent of organisations cite managing cloud spend as their single biggest cloud challenge (mrc, 2026). For workloads with steady, predictable demand, owning capacity is once again cheaper than renting it. Broadcom's internal analysis claims modern private cloud delivers 40 to 50 percent lower total cost of ownership for steady-state workloads than public cloud (mrc, 2026). The figure comes from a private-cloud vendor and should be read with that in mind, but the underlying logic is sound and widely corroborated.
Data sovereignty is tightening. According to the Nutanix Enterprise Cloud Index 2026, 57 percent of IT leaders feel a need to run infrastructure within a single country, as regulatory pressure in financial services, healthcare, and government intensifies (mrc, 2026). For regulated industries, keeping certain data on home soil is shifting from preference to requirement.
And then there is AI. This is the driver that did not exist during the last repatriation debate, and it cuts both ways. Training and inference on GPU clusters in the public cloud is expensive, and the "data gravity" problem, where moving terabytes of training data in and out of cloud regions becomes a bottleneck, pushes organisations toward co-located or on-premises GPU infrastructure (Tasrie, 2026). Forrester predicts that at least 15 percent of enterprises will shift toward private AI deployments built atop private clouds in 2026, in direct response to rising AI costs, data lock-in, and operational risk (HyScaler, 2026). Gartner goes further, predicting that 40 percent of enterprises will adopt hybrid compute architectures for mission-critical workflows by the end of 2026, up from a far smaller base in prior years (Tasrie, 2026).
The AI repatriation argument is the strongest of the three, because it is driven by economics and physics rather than by sentiment. A model that needs constant access to a large, sensitive, proprietary dataset is often genuinely better off running where that data lives.
The resilience argument nobody priced in
There is a fourth driver that rarely makes the vendor decks but that hardened considerably in 2025: operational risk.
AWS, Azure, and Cloudflare all suffered high-profile outages during the year (TechRadar, 2026). Forrester now predicts that AI data-centre upgrades will trigger at least two major multi-day cloud outages in 2026, as hyperscalers divert investment away from legacy infrastructure toward GPU-centric facilities and aging systems falter under growing complexity (Forrester, 2025).
The lesson most organisations took from the 2025 outages was the wrong one: switch provider, or wait for the cloud to get more reliable. The right lesson is architectural. The exposure was never simply that a provider went down. It was that the business had built single-provider dependency with no tested failover, and "the cloud went down" became the explanation that let everyone avoid the harder conversation about their own design choices.
Which workloads actually belong home
For a business leader, the honest framing is not "should we repatriate" but "which workloads have outgrown the public cloud, and which have not." The evidence supports a clear test.
Workloads with strong repatriation logic: steady-state, predictable-demand systems where you are paying rental prices for capacity you use constantly; workloads bound by data-sovereignty or sector regulation; and AI training or inference workloads with heavy, persistent data-gravity that makes cloud egress and GPU rental punishingly expensive.
Workloads that almost always belong in the public cloud: anything with spiky, unpredictable, or seasonal demand, where elasticity is the entire point; early-stage products still finding their shape, where speed of iteration matters more than unit cost; and global-facing services that benefit from a hyperscaler's geographic reach.
The organisations getting this right are not the ones reacting to a headline statistic. They are the ones placing each workload deliberately, on the infrastructure that suits its actual demand profile, cost behaviour, and regulatory weight. That is hybrid architecture done properly, and it is the opposite of both the 2018 "cloud-first" reflex and the 2026 "bring it all home" overcorrection.
Where Neurotic comes in
The repatriation decision is being made badly across the mid-market right now, in both directions. Some businesses are staying fully in the public cloud and bleeding money on steady-state workloads they could own. Others are being sold an expensive private-infrastructure project on the strength of a misleading statistic, and will move the wrong workloads at the wrong time.
The work that gets skipped is the unglamorous part: a clear-eyed, vendor-independent assessment of which workloads actually belong where, what the real total cost of ownership is in each case, and how to move the ones that should move without breaking the ones that should not.
Neurotic works with growth companies, scale-ups, and corporates to make exactly those decisions. Pragmatic, engineering-led, and independent of any infrastructure provider, which means the recommendation is driven by your demand profile and your economics, not by what someone is trying to sell you. If your cloud bill has outgrown your cloud value, or if someone is pitching you a repatriation project on the back of a statistic, that is the moment to get an independent read.
Talk to us โ neurotic.co
References
Channelnomics (2024) Breaking Down the 83% Public Cloud Repatriation Number. [online] Available at: https://channelnomics.com/breaking-down-the-83-public-cloud-repatriation-number/ [Accessed 26 May 2026].
Forrester (2025) Predictions 2026: Cloud Outages, Private AI On Private Clouds, And The Rise Of The Neoclouds. [online] Available at: https://www.forrester.com/blogs/predictions-2026-cloud-outages-private-ai-on-private-clouds-and-the-rise-of-the-neoclouds [Accessed 26 May 2026].
HyScaler (2026) Cloud Repatriation in 2026: Why Enterprises Are Moving Back from the Cloud. [online] Available at: https://hyscaler.com/insights/cloud-repatriation-the-strategic-shift-in-it/ [Accessed 26 May 2026].
Integrate.io (2026) Data Integration Adoption Rates in Enterprises: 45 Statistics Every IT Leader Should Know in 2026. [online] Available at: https://www.integrate.io/blog/data-integration-adoption-rates-enterprises/ [Accessed 26 May 2026].
mrc (2026) What's Driving Cloud Repatriation in 2026? [online] Available at: https://www.mrc-productivity.com/blog/2026/04/whats-driving-cloud-repatriation-in-2026/ [Accessed 26 May 2026].
Tasrie IT Services (2026) Cloud Repatriation 2026: Why 86% of CIOs Are Moving Workloads Back. [online] Available at: https://tasrieit.com/blog/cloud-repatriation-guide-2026 [Accessed 26 May 2026].
TechRadar (2026) The year of the AI agents? More outages? Here's what lies ahead for IT teams in 2026. [online] Available at: https://techradar.com/pro/the-year-of-the-ai-agents-more-outages-heres-what-lies-ahead-for-it-teams-in-2026 [Accessed 26 May 2026].
WebProNews (2024) 83% Public Cloud Repatriation Stat is Misleading. [online] Available at: https://www.webpronews.com/83-public-cloud-repatriation-stat-is-misleading/ [Accessed 26 May 2026].
