cd /news/artificial-intelligence/youve-decided-to-repatriate-cloud-wo… · home › topics › artificial-intelligence › article
[ARTICLE · art-170904] src=networkworld.com ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

You’ve decided to repatriate cloud workloads. Now what?

Two-thirds (66%) of 527 IT decision makers surveyed in Foundry's 2026 Cloud Computing Study say their organizations have moved or plan to move applications or workloads out of the public cloud to on-premises or colocation, with respondents reporting an average of about 22% of public cloud workloads already repatriated and about 24% more expected to move over the next 12 months. Security concerns drive the shift for 60% of respondents, ahead of budget and cost control (42%), performance and reliability (42%), data sovereignty and privacy (40%) and compliance (31%), while 69% still expect cloud spending to increase and 74% default to the cloud for new capabilities. AI and machine learning is the most-cited planned move: 34% have already moved AI/ML workloads out of the public cloud and another 38% plan to, and only 13% plan to run AI primarily in the public cloud versus 38% hybrid and 28% primarily private cloud or data center.

read6 min views1 publishedNov 17, 2026
For 15 years, enterprise IT moved in one direction: out of the data center and into the public cloud. Hardware teams were trimmed, leases weren’t renewed, and “cloud first” became “cloud only.” That era is ending, not because the cloud failed, but because many organizations never made a placement decision. They made a migration decision and assumed it would fit everything. AI, global macro issues, data ownership, and other trends are prompting IT leadership to rethink their cloud strategy.

New data from Foundry’s [2026 Cloud Computing Study](https://foundryco.com/tools-for-marketers/executive-summary-cloud-computing/), which surveyed 527 IT decision makers, shows how far the pendulum has swung. Two-thirds (66%) of respondents say their organizations have moved or plan to move applications or workloads out of the public cloud to on-premises or colocation locations. On average, respondents report that about 22% of their public cloud workloads have already been moved back and expect to move about 24% over the next 12 months.

What gets lost in the “cloud is dead” headlines is that the same study shows 69% of organizations expect cloud spending to increase, and 74% default to the cloud when buying new capabilities. Repatriation isn’t a retreat from the cloud. It’s a rebalancing, and a hybrid estate is fundamentally harder to run than either an all-cloud or all-on-prem estate.

The strategic “why?” is one important question (See our related CIO feature: Cloud repatriation on the rise but tricky questions remain). For the people who rack the gear, light the circuits and keep things running, the question is “how?”

The top driver may surprise those who remember that security was the primary objection to moving to the cloud in the first place. Sixty percent of respondents cite security concerns as a reason for moving workloads out of the public cloud, followed by budget and cost control (42%), performance and reliability (42%), and data sovereignty and privacy (40%). Compliance comes in at 31%.

That’s a control story more than a cost story. After a decade of shared-responsibility models, many security teams have concluded that “shared” often means “unclear.” When regulators and boards ask exactly where data lives and who can access it, owned infrastructure provides a clearer answer. Sovereignty will only grow in importance, particularly for the half of the sample outside North America. On cost, network pros should pinpoint where the pain is. Data transfer (31%) and networking costs (29%) rank near the top of respondents’ cloud cost-management challenges. Egress has been the cloud’s quiet tax for years, and data-heavy workloads expose it most quickly.

The early movers are what you’d expect: storage and backup (39% have already moved), databases (38%), web apps (38%), and data integration (36%). These are data-intensive workloads, where moving data to and from a hyperscaler no longer makes sense because of ingress/egress costs.

The more interesting workload is AI. Thirty-four percent have already moved AI or machine learning workloads out of the public cloud, and another 38% plan to do so, tying for the highest “will move” figure in the study. That aligns with how companies say they plan to run AI overall: Only 13% plan to run AI workloads primarily in the public cloud, while 38% plan a hybrid approach and 28% plan to run primarily in a private cloud or data center.

So, is AI better suited to owned infrastructure? Increasingly, yes, at least for steady-state inference and fine-tuning on proprietary data. Renting GPUs by the hour works for bursty experimentation. It works far less well once a model runs 24×7 on data that already sits on-prem and that security doesn’t want to leave the building. Large-scale training and burst capacity will often stay with a hyperscaler or neocloud, so the answer is a mix, which puts the network in the middle.

This is where repatriation gets real, and where I see the most underestimation.

Space and power aren’t sitting around waiting for you. Companies that shut down data centers often assume they’ll just lease colocation space. Good luck. CBRE reports that primary-market vacancy in North America fell to a record low of 1.4% in the first half of 2026, with Northern Virginia at 0.2%. JLL says 95% of the 66 GW currently under construction in North America is already pre-committed. Enterprises are competing with hyperscalers and AI companies that have far deeper pockets. GPU racks can draw several times the power of a traditional enterprise rack and may require liquid cooling that older facilities can’t support.

The network is the migration. Repatriation means connecting more places, not fewer: high-capacity interconnects between on-prem, colo, and multiple clouds; cloud on-ramps; consistent segmentation and policy; and a fresh look at data center fabric design if AI clusters are coming in-house. Already, 52% of respondents say network latency hurts application performance, and 56% struggle to integrate cloud with on-premises systems. Repatriation amplifies both unless the network is designed for hybrid from the start.

Cloud-native doesn’t come home easily. Lifting a VM back is easy. Moving an app built on managed databases, serverless functions, and proprietary services is not. Respondents cite data portability (49%) and environment incompatibilities (42%) as migration obstacles, and those don’t disappear in reverse. Expect re-architecting and budget for the egress bill of the move itself.

The skills gap runs in both directions. Many organizations let deep data center expertise walk out the door. The study shows companies adding AI engineers, platform engineers, and cloud architects, while just 15% added cloud network engineers. Now they need people who understand power density, fabric design, and the hardware lifecycle, plus the cloud operating model, because users won’t go back to ticket-driven provisioning. The goal isn’t to rebuild the 2010 data center; it’s a cloud-like experience on infrastructure you own.

Operations are getting more complex, not less. Hybrid means multiple control planes, telemetry sources, and cost models. Without unified observability and automation, ops teams will spend their time stitching together dashboards.

Repatriation is the market correcting a decade of poor placement decisions. That’s healthy, but the burden of getting it right falls on network, I&O, and data center teams. If you’re facing a repatriation project, here’s where I would focus.

Cloud repatriation is often framed as a referendum on the public cloud, but that misses the point. The Foundry data shows companies are still increasing cloud spending even as they bring a meaningful share of workloads back on-premises. What’s changing is the assumption that every workload belongs in one place. Security, sovereignty, cost predictability, and AI are forcing organizations to be deliberate about where applications and data reside.

For IT pros, that’s both a challenge and an opportunity. The challenge is that hybrid is harder than all-cloud or all-on-prem. It requires scarce data center capacity, a network built to span multiple environments, operations tools that span them all, and skills many organizations let lapse. The opportunity is that infrastructure teams are back at the center of the strategy conversation. For years, the network and data center were viewed as plumbing to be minimized. Repatriation, and AI in particular, puts them back in the role of a competitive differentiator. The companies that win won’t be the ones that pick cloud or on-prem. They’ll be the ones that build an infrastructure team capable of running both as a single environment and placing each workload where it delivers the best mix of performance, cost, and control.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @foundry 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/youve-decided-to-rep…] indexed:0 read:6min 2026-11-17 · —