Share with your CTO
Ninety-five percent of enterprises have delayed or cancelled AI projects in the past year because their existing infrastructure wasn’t built to handle AI’s demands for scale, governance, and flexibility, according to a Cloudera survey of 1,500 enterprise and cloud architects. Nearly three-quarters say full infrastructure revamps are required. Two-thirds have already moved AI workloads back from public cloud to on-premises or private cloud, and 84% report that AI workloads have driven infrastructure costs higher. Within two years, one in four enterprises plans to commit to hybrid-first architecture as a default posture.
What this means for your business
The companies most exposed here aren’t the laggards who haven’t started AI projects. They’re the ones who did start, hit a wall, and are now carrying both the sunk cost of paused initiatives and the mounting pressure to re-architect before they can restart. If your data estate was designed for batch analytics on a single cloud, you’re not looking at an upgrade cycle, you’re looking at a structural rebuild. The deciding variable isn’t AI ambition; it’s whether your data governance model was designed for workload portability or bolted on after the fact.
Cloudera, whose commercial interest sits squarely in hybrid data management, frames the solution predictably as hybrid-first infrastructure, and that framing does subtly flatten the range of remedies available. But the underlying problem the survey surfaces is real and vendor-neutral. The regulatory and compliance drag that legacy systems create isn’t a configuration problem. Enterprises running sensitive data through architectures that weren’t built to log, audit, and segment AI inference at scale are accumulating compliance debt faster than they’re shipping AI value. Liberty Mutual’s approach, building a model-agnostic abstraction layer on top of mainframe data rather than ripping the mainframe out, is the more instructive case study here than any infrastructure vendor’s roadmap.
The repatriation signal deserves more weight than it’s getting. Two-thirds of enterprises moving AI workloads off public cloud isn’t a tactical cost-trim, it’s a reassessment of the assumption that cloud-native infrastructure is the natural home for enterprise AI. If that trend holds, the CTOs who locked in multi-year public cloud commitments during 2022 and 2023 are now defending contracts against workloads that have drifted back on-premises. That’s the renewal to weigh differently, not as a cloud-versus-on-prem ideology debate, but as a question of whether your current contract structure gives you the workload placement flexibility the next two years will actually require.
Concept deep-dive: Model-agnostic abstraction layer
A model-agnostic abstraction layer sits between your underlying data systems and whichever AI models consume that data, acting like a universal adapter so you can swap or combine AI models without rewiring your data pipelines each time. It exists because AI model selection is still volatile, the best model for a task today may not be the best one in eighteen months. For enterprises with legacy data on mainframes or siloed systems, this layer is often what makes AI access possible without a full re-architecture first.
Based on reporting from Legacy IT forces enterprises to delay AI projects, originally published 2026-08-12 16:01:00.
