Business adoption of artificial intelligence is advancing faster than the infrastructure needed to support it. A new study commissioned by Cloudera concludes that 95% of surveyed organizations delayed or canceled an AI project in the past year due to issues related to data governance, compliance, or regulation. The results point to a less visible shift than the GPU race: many companies are having to redesign the architecture for storing, governing, and moving their data.
The key points of enterprise AI restructuring in 30 seconds
- 95% delayed or canceled AI initiatives due to governance, compliance, or regulatory issues.
- A 72% believe their data architecture needs significant changes to meet future AI requirements.
- 84% say AI workloads have increased their infrastructure costs.
- 66% moved workloads from public cloud to private cloud or on-premises infrastructure.
- The study surveyed 1,500 professionals in architecture, cloud, and data roles across nine markets, including Spain.
The figures come from The Great AI Re-Architecture, a survey commissioned by Cloudera from Wakefield Research among technical leaders of large organizations. The work was conducted between June 5 and 22, 2026, in the U.S., Canada, Brazil, South Africa, Spain, the UK, Singapore, India, and Japan.
The sample requires careful interpretation of percentages. It does not mean that 95% of all companies worldwide have halted AI projects, but that this proportion of the 1,500 professionals surveyed reported experiencing delays or cancellations for those reasons.
Even with that caution, the results align with another study published by Cloudera itself in April. That survey, involving 1,270 IT leaders, found that nearly 80% acknowledged their data and AI initiatives were limited by difficulties in accessing needed information across different environments.
The problem with enterprise AI is no longer just about acquiring GPUs
Over the last two years, much of the conversation about AI infrastructure has focused on accelerators, data centers, and power availability. But a company can have enough compute capacity and still face difficulties putting an AI system into production.
The issue lies a few levels deeper.
Enterprise data is often spread across traditional databases, SaaS applications, object storage, data lakes, legacy systems, various cloud providers, on-premises data centers, and edge devices.
A proof of concept may work with a limited subset of this information. Deploying the same system into production requires determining who can access each data source, where it can be processed, how long it should be kept, and what information a model is truly allowed to use.
This is where limitations of architectures originally built for business intelligence, conventional analytics, or transactional applications begin to appear.
75% of participants in the new study say that AI integration has already altered their organization’s data storage and architecture practices. Additionally, 84% report higher infrastructure costs due to these workloads.
It’s not surprising that 72% see the need for a significant overhaul of their architecture to meet future demands.
Cloudera calls this process The Great AI Re-Architecture. The term comes from the company itself and is part of the commercial approach of the study, but it describes a recognizable technological phenomenon: adding a model on top of an existing data infrastructure does not always suffice to build an enterprise AI platform.
Cloudera warned months ago that initial AI experiments were often deployed in fragmented data environments without preparing governance and security for large-scale operations.
Hybrid data is at the center of architecture again
One of the most interesting results appears when asking where these workloads are run.
66% of respondents say they moved AI workloads from public clouds to private clouds or on-premises infrastructure in the past year.
This does not imply a wholesale retreat from public clouds.
In fact, companies continue using multiple environments and moving data between them. 97% of participants say they transfer data across different infrastructures at least once a month.
What emerges is a more distributed architecture.
Some workloads make sense on hyperscalers due to their elasticity and immediate availability of accelerators. Others may be better suited for dedicated infrastructure when they require consistent utilization, proximity to large data volumes, or are driven by sovereignty, security, and regulation requirements.
The choice can also change during a project’s lifecycle. Training or experimenting in one environment and then performing inference in another is becoming technically feasible if data and model platforms maintain sufficient interoperability.
That’s why hybrid shouldn’t simply mean maintaining some on-prem servers and simultaneously contracting AWS, Azure, or Google Cloud.
The real challenge is establishing consistent policies for identity, security, cataloging, governance, and data access, regardless of where each workload runs.
Cloudera has a direct commercial interest in this trend because its platform is designed precisely to combine data centers and different cloud environments. In April, the company announced new features aimed at extending BYO (bring your own) capacity to the cloud and sharing Apache Iceberg tables without data duplication.
The study should be interpreted with this provider stance in mind.
But the issues they identify are not eliminated by this approach.
Data governance is becoming an AI infrastructure concern
Perhaps the most striking statistic in the report is not about servers.
73% believe AI has made data governance more complex, while 55% state they have delayed or canceled more than six projects in the past year due to governance, compliance, or regulatory concerns.
This significantly changes the meaning of being “AI-ready.”
Having a corporate account with a model provider, GPU access, or a team capable of developing applications no longer guarantees that a project can deploy on actual enterprise data.
An organization must understand what data the system uses, its origin and permissions, what information the model can access, and what happens to its outputs.
Generative AI complicates this further because it often works with unstructured data: documents, emails, conversations, images, recordings, PDFs, or code repositories.
The challenge increases with architectures like RAG (Retrieval-Augmented Generation), agents, and tools capable of querying enterprise systems. Instead of working solely with pre-prepared datasets, models demand real-time information from multiple sources during execution.
The more tools an agent can use, the more critical it becomes to determine which identity it uses and the extent of its permissions.
Cloudera’s earlier Data Readiness Index results already showed this contradiction: although many companies claimed confidence in their data, fewer than 20% reported having full governance over it.
Continuously moving data isn’t free either
The study also raises an economic consideration.
If 97% transfer data between environments at least monthly, AI architecture cannot be analyzed solely based on GPU costs.
Moving large amounts of data incurs costs for network, storage, replication, and operations. It can also introduce latency and create new copies that require their own security and governance policies.
In certain scenarios, it may be better to bring the model to the data rather than move data continually to the model.
This is especially true for sectors handling massive data volumes or with strict regulatory constraints.
For example, in telecommunications, another Cloudera report from June found that 90% of respondents experienced operational limitations related to infrastructure performance, despite most claiming to know where their data resides.
The issue isn’t just knowing the data location; it’s being able to access it with sufficient performance, maintaining applicable policies, and doing so in a cost-effective manner.
The next phase of enterprise AI will be less visible
The first stage of generative AI was driven by models. Later came GPUs and the massive data centers required to train and run them.
Now, companies are entering a less flashy phase: restructuring years of technological infrastructure so that these models can work with real corporate data.
This includes modernizing data lakes, reducing silos, establishing catalogs, managing identities, implementing observability, updating networks, deciding where to run inference, and determining what data can leave each infrastructure.
It also requires revisiting a common misconception from early generative AI: that simply connecting a powerful model to company data is enough to start deriving value.
In production, there are requirements that a demonstration often overlooks.
Cloudera’s study reflects that gap: 77% of surveyed organizations are already using AI, but a majority also believe their current architecture needs significant changes.
So, the bottleneck isn’t necessarily adopting a more advanced model.
For many large companies, the real challenge is making their data accessible, governable, and usable across different infrastructures without losing control.
Frequently Asked Questions
Is it really true that 95% of companies have halted AI projects?
The 95% figure from the Cloudera study refers to the professionals who reported that their organization delayed or canceled initiatives over the past year due to data governance, compliance, or regulation issues. This figure is based on the sample studied and should not be automatically extrapolated to all companies worldwide.
Why are workloads moving back from public cloud?
66% of respondents say they transferred some workloads from public cloud to private cloud or on-premises infrastructure. Factors influencing this include costs, performance, governance, data control, and regulatory requirements.
Does this mean companies are abandoning the cloud?
Not necessarily. The results point to hybrid environments: public cloud, private cloud, data centers, and edge devices used together depending on each workload’s characteristics.
How many companies participated in the study?
Wakefield Research conducted the survey for Cloudera among 1,500 enterprise architects, cloud infrastructure managers, and data architects across nine markets, including Spain, from June 5 to June 22, 2026.
via: cloudera
