The budget numbers behind the repatriation trend
Cloudian commissioned Centiment to survey 212 senior IT decision-makers at organizations with significant public cloud adoption, fielding the survey in February 2026. Ninety-six percent of respondents were primary decision-makers or key influencers on infrastructure spending, and every response passed an embedded attention-check question, a methodology detail worth noting given how often cloud cost surveys get dismissed as self-serving vendor research. The headline number is stark: 84% of respondents are over their cloud storage budgets, and nearly one in five are over by more than 30%. Only 0.5% report being under budget.
The three cost drivers respondents named most often map directly onto structural features of public cloud pricing rather than one-off overruns. Data egress fees led at 46%, followed by costs that rise with data volume at 45%, and premium pricing for residency-compliant regions at 43%. None of these are usage spikes a FinOps team can simply optimize away. They are pricing mechanics built into how hyperscalers charge for data movement and geographic placement, which explains why 76% of respondents still keep more than half their workloads in public cloud despite the budget pressure.
Repatriation is already happening, not just planned
Seventy-five percent of respondents have already moved at least some workloads back from public cloud within the past 24 months, and 89% plan to expand on-premises infrastructure over the next two years. That is a meaningfully different signal than a survey measuring intent alone, since three-quarters of this sample are describing something they have already executed rather than something still under consideration. Roughly 30% say they are still expanding their cloud footprint at the same time, which means growth and repatriation are happening inside the same organizations simultaneously, segmented carefully by workload rather than applied as a single blanket strategy across the board.
Cloudian CMO Jon Toor frames this as workload-specific rather than wholesale, telling reporters it reflects organizations getting smarter about workload placement decisions generally. That reading fits the data better than a simple cloud-exodus narrative would suggest. Enterprises are keeping substantial workloads in public cloud while routing specific categories, storage-heavy, latency-sensitive, or residency-constrained workloads in particular, back to infrastructure they fully control, leaving elastic, bursty, or genuinely cloud-native workloads exactly where they already are and performing well.
Sovereignty mandates are now a daily operating constraint
Ninety-nine percent of respondents say data sovereignty is at least a moderate factor in their infrastructure decisions, and 82% call it a primary or significant driver, figures too high to treat as a niche compliance concern limited to a handful of regulated industries. Fifty-three percent already operate under customer or partner contracts restricting where data can reside, and 45% have faced new cross-border data restrictions in just the past two years, a pace of new restriction that outstrips most organizations' ability to redesign architecture in response. For infrastructure teams, that means data residency has moved decisively from a legal review checkbox to a recurring architectural constraint that actively shapes vendor selection from the outset.
Fifty-nine percent say they are specifically concerned about cloud providers accessing their data for analytics or model training purposes, a worry that has sharpened noticeably as hyperscalers bundle AI services more tightly into their core infrastructure offerings. That concern sits alongside, and actively compounds, the cost and performance drivers described above, since avoiding a provider's AI training pipeline sometimes means walking away from the exact features that would otherwise justify staying on that platform for a given workload.
AI workloads are pulling infrastructure in two directions
Eighty-five percent of respondents say AI requirements are influencing their shift toward on-premises infrastructure, which runs counter to the assumption that AI workloads require cloud elasticity above all else. Fifty-five percent say cloud cannot consistently meet AI inference latency requirements, and 52% say they need to keep AI training data on-premises for security or compliance reasons. Both figures point to a specific, falsifiable complaint: cloud infrastructure is not meeting real-time performance needs for AI inference at the edge of where that inference actually needs to happen.
This matters because it decouples the repatriation trend from the older story about compute cost. A team can be perfectly happy with cloud GPU pricing and still pull inference workloads on-premises because latency requirements cannot be met from a regional cloud data center. Infrastructure leaders should treat AI workload placement as a separate decision from AI cost management, since the two are being driven by different, independent constraints in this data.
What CIOs should take from a hybrid-by-default posture
Toor describes hybrid infrastructure today as a deliberate strategy that organizations choose on purpose, a framing that captures what the numbers actually show: 57% name AI/ML infrastructure as a top priority for the coming year, 56% name data sovereignty and security, and 54% name cost predictability, three distinct priorities that all point toward hybrid architecture as the rational default rather than a fallback position reached only after a cloud-first plan fails. Organizations still planning infrastructure around a single deployment model, all-cloud or all-on-prem, are optimizing for a world this survey suggests most of their peers have already left behind in practice.
The practical move is to build workload placement criteria explicitly around the three drivers this data surfaces: egress and volume-based cost exposure, contractual or regulatory residency requirements, and latency sensitivity for AI inference specifically. A workload that triggers any one of those should default to an on-premises or sovereign evaluation rather than an automatic cloud-first decision, and that evaluation should happen before procurement, not after the first unexpectedly large bill.
