Enterprises Are Pulling AI Workloads Back From the Cloud, and Latency Is the Real Reason
Digital Transformation

Enterprises Are Pulling AI Workloads Back From the Cloud, and Latency Is the Real Reason

A new Cloudian survey of 212 senior IT decision-makers finds 89 percent planning to expand on-premises infrastructure, and the honest answer for why has less to do with cost than with AI workloads that public cloud genuinely struggles to serve.

PublishedNovember 18, 2026
Read time6 min read
Share

The headline numbers

A survey of 212 senior IT decision-makers, commissioned by storage vendor Cloudian and fielded in February 2026, found that 89 percent plan to expand on-premises infrastructure over the next two years. Three quarters of respondents say they have already moved at least some workloads back from public cloud in the past 24 months. Ninety six percent of the people surveyed were either primary decision-makers or key influencers on infrastructure strategy, which gives the findings more weight than a typical vendor-commissioned poll aimed at junior buyers.

Jon Toor, Cloudian's chief marketing officer, pushed back on the obvious framing before anyone else could make it: "This isn't a story about enterprises souring on cloud." His point is supported by the data itself. About 30 percent of respondents are still actively expanding their cloud presence even as the broader group plans more on-premises capacity, and 76 percent currently keep more than half their workloads in public cloud. This is a rebalancing of workload placement, not a wholesale reversal of the last decade's cloud migration.

AI changed the calculation, not just the budget

The most striking number in the survey is that 85 percent of respondents say AI requirements are actively shaping their move toward on-premises infrastructure, ahead of cost or any single compliance factor on its own. Fifty five percent say public cloud cannot consistently meet AI inference latency requirements, a specific and technical complaint rather than a general unease about cloud economics. Toor frames AI as the forcing function behind the shift, arguing that enterprises tolerated workarounds for conventional workloads that they simply cannot accept for AI.

That distinction matters because it separates two different IT decisions that often get conflated in budget conversations. A company can remain fully committed to public cloud for its general application portfolio while concluding that specific AI inference and training workloads need different infrastructure because of latency, data gravity, or security requirements unique to how those workloads actually run. Fifty two percent of respondents specifically cite the need to keep AI training data on-premises for security or compliance reasons, which is a workload specific decision, not a cloud-wide one.

The cost story is still real underneath

AI may be the headline driver, but cost pressure has not gone away. Eighty four percent of surveyed organizations exceed their cloud storage budgets, and nearly one in five are over budget by more than 30 percent. Only half a percent of respondents say they are under budget, which is about as close to a universal finding as a survey like this produces. The main cost drivers respondents named are data egress fees at 46 percent, costs that rise with data volume at 45 percent, and premium pricing for data residency-compliant regions at 43 percent.

Those specific cost drivers point to a pattern CIOs should recognize from their own cloud bills: the pain is concentrated in data movement and data residency charges rather than in raw compute pricing, which public cloud vendors compete on aggressively. That distinction explains why repatriation conversations tend to start with storage and data-heavy workloads rather than with compute-bound applications, where cloud elasticity still delivers a clear advantage that is harder to replicate on-premises.

Data sovereignty is the slower moving driver

Nearly every respondent, 99 percent, rates data sovereignty as at least a moderate factor in their infrastructure decisions, and 82 percent call it a primary or significant driver. Fifty three percent say they operate under customer or partner contracts that restrict where their data can reside, and 45 percent have faced new cross-border data restrictions in just the past two years. Fifty nine percent say they worry specifically about cloud providers accessing their data for analytics or model training, a concern that has sharpened considerably as AI vendors look for training data.

This driver moves on a different timeline than cost or latency. Cost pressure can ease with a renegotiated contract, and latency can sometimes be solved with better architecture inside the same public cloud. Sovereignty and contractual data residency requirements are set by regulators and customers, not by the enterprise's own technical choices, which makes this the most durable of the drivers the survey identifies. A CIO can out-negotiate a cloud bill. A CIO cannot out-negotiate a new cross-border data law.

Why this looks like rebalancing, not retreat

Toor's own description of what good practice looks like reinforces the rebalancing framing over a retreat narrative: "The enterprises doing this well have gotten very precise about what goes where and why." That precision is the actual skill this moment demands from infrastructure teams, replacing the earlier era's simpler default of moving everything to public cloud and optimizing later. The survey's own priority ranking supports this: AI and ML infrastructure placement ranks as the top priority for the coming year at 57 percent, ahead of sovereignty and security at 56 percent and cost predictability at 54 percent, three closely clustered priorities rather than one dominant concern.

It is worth being direct about the source here. Cloudian sells on-premises object storage, so a survey commissioned by Cloudian and framed by Cloudian's own CMO naturally favors a narrative that supports on-premises investment. The underlying statistics still reflect real respondent answers from 212 senior IT buyers, but the framing and the choice of which numbers to lead with both come from a vendor with a direct financial interest in this exact shift, and that context belongs in how much weight a CIO gives the headline percentages.

What this means for your infrastructure roadmap

Treat workload placement as a per-workload decision rather than a single cloud-or-not policy. The survey's own data supports exactly that approach: respondents are simultaneously expanding on-premises capacity and continuing to grow cloud usage, which only makes sense if different workloads are being routed to whichever environment actually fits their latency, data residency, and cost profile. A blanket repatriation policy driven by headline anxiety would ignore the nuance the survey itself reveals.

Start the placement review with your AI training and inference workloads specifically, since that is where the survey's respondents report the clearest technical mismatch with public cloud, rather than with your general application portfolio, where cloud elasticity still works in your favor. Build any repatriation business case around actual egress and data volume costs you can audit in your own cloud bill, not around a vendor survey's aggregate percentages, however directionally useful those percentages are for spotting the trend early.

Tagged#news#digital-transformation#enterprise#cio#erp#strategy#governance#cloud-repatriation#data-sovereignty#on-premises-infrastructure#ai-infrastructure#hybrid-cloud