Summary

• Private AI adoption varies sharply by sector, with healthcare, financial services, manufacturing, and public services applying AI to different workflows, risks, and data types.
• The shared infrastructure issue is the full data pipeline, including ingestion, storage, governance, inference, protection, recovery, and long-term reuse.
• As AI moves from pilots into production, storage performance, object storage, sovereignty, and lifecycle control are becoming central to enterprise AI strategy.

By Paul Speciale, CMO at Scality

Enterprise AI is moving from isolated experimentation into production workflows, but the shape of that shift is not uniform. A hospital using AI to support diagnostic imaging carries a different risk profile from a bank using AI for fraud detection, while a manufacturer analysing sensor data from factory equipment has different latency and data movement requirements from a public sector agency building AI into citizen services.

Although the use cases vary by industry, the infrastructure requirement is becoming more consistent: AI success depends on the data layer. As organisations move AI from pilots into operational systems, scaling the technology becomes less a question of models alone and more a question of how data is stored, moved, governed, protected, and made available across the full AI lifecycle.

That is one of the clearest findings from independent research by Freeform Dynamics, based on input from 504 senior IT and data professionals at medium and large enterprises active with private AI. The study, Storage Infrastructure for Enterprise AI: Lessons from Seasoned Adopters on Building Scalable Sovereign Environments, focused on storage infrastructure requirements for private AI implementations running in datacentres or hosted environments where organisations control the infrastructure stack.

The findings show that private AI is becoming a production infrastructure decision, not just a model strategy. Across the survey, 81% of respondents said private AI based on infrastructure they control is critical to success, including 43% who said this fully applies and 38% who said it somewhat applies.

That does not mean the public cloud disappears from enterprise AI. Hybrid models remain common, and cloud services will continue to play an important role. However, as organisations scale AI into business-critical environments, they are becoming more deliberate about where data and workloads should live, and which systems need direct control over location, access, governance, performance, resilience, cost, and compliance.

Those requirements become sharper in regulated and data-intensive sectors, where AI cannot be separated from the infrastructure that feeds, protects, and governs it. Looking at the findings by industry makes the point clearer, because each sector is adopting AI through the lens of its own workflows, risk profile, and data types. Across those sectors, the same infrastructure questions keep reappearing: where the data lives, how it is protected, how quickly it can be accessed, how it moves across the pipeline, and how it remains useful over time.

Enterprise AI is not one workload

Much of the public conversation around AI still centres on generative AI and large language models. Inside enterprises, the picture is broader. The Freeform Dynamics research found organisations active across traditional machine learning, RAG-enhanced foundation LLMs, fine-tuned or customised models, computer vision, time-series and IoT analytics, recommendation and personalisation technologies, and edge or distributed AI.

That diversity changes the infrastructure conversation. In the survey, 68% of respondents were active with at least three different AI genres, while 29% were active with at least five. Training may require high-throughput access to massive datasets, inference may depend on low-latency access to models, vector stores, and reference data, and computer vision requires efficient handling of large image repositories.

At the same time, time-series and IoT analytics depend on continuous streams of operational data, while RAG-enhanced LLMs need trusted access to enterprise content and knowledge sources. Enterprises are therefore not adopting AI in the abstract. They are applying it to specific workflows, data types, and business risks, and those choices shape the infrastructure beneath them.

Healthcare and life sciences need governed AI

Healthcare and life sciences organisations are moving forward with AI, although their adoption pattern is more measured than in some other sectors. The survey shows strong use of established AI approaches in healthcare and life sciences, with traditional machine learning and computer vision leading the sector’s production and active development workloads.

AI workload adoption among healthcare and life sciences respondents. Source: Freeform Dynamics.

The report notes that healthcare and life sciences organisations show notably lower LLM adoption, likely reflecting concerns about output variability, hallucinations, and regulatory constraints. At the same time, computer vision is already well established in diagnostic imaging, where AI can be applied to constrained, high-value tasks with more clearly defined validation requirements.

Rather than indicating that the sector is behind, the figures point to an environment where AI must be deployed in ways that can be validated, governed, and trusted. Healthcare and life sciences organisations manage some of the most sensitive and data-rich environments in the enterprise world, from patient records and clinical notes to imaging data, research datasets, and genomic information.

As AI expands across diagnostics, operations, research, and patient engagement, infrastructure has to support more than raw performance. It must also support access control, lifecycle management, data integrity, retention, recovery, and auditability. In these settings, trusted AI starts with trusted data infrastructure.

Financial services is applying AI at data-driven scale

Financial services organisations have spent decades building around data, analytics, and automation, and that history shows up clearly in the survey. Among financial services respondents, RAG-enhanced foundation LLMs and traditional machine learning were the most commonly cited AI workloads in production or active development.

AI workload adoption among financial services respondents. Source: Freeform Dynamics.

This points to a mature and balanced AI strategy. Financial institutions are adopting newer LLM-based approaches for knowledge management and decision support, while continuing to rely on machine learning for long-established use cases such as fraud detection, risk modelling, algorithmic trading, customer segmentation, marketing, and retention.

The infrastructure stakes are high because financial services organisations need to move quickly without loosening control. AI systems often depend on large volumes of transaction data, market data, customer data, and reference data. Latency can affect decision-making, data quality can affect risk, and security, resilience, auditability, and regulatory compliance remain central to deployment.

As a result, private AI in financial services is not only about adopting advanced models. It is about operationalising them inside data environments that can support real-time insight while preserving control over sensitive information, regulated processes, and business-critical systems.

Manufacturing links AI from edge to core

Manufacturing’s AI story is strongly operational. AI is increasingly tied to physical systems, production environments, plant-floor data, and supply chain execution, and the survey shows broad adoption across manufacturing respondents.

AI workload adoption among manufacturing respondents. Source: Freeform Dynamics.

Those findings reflect the wide range of AI use cases already emerging across manufacturing. The report notes that manufacturers have long used computer vision with trained models for automated quality control, while time-series analytics and machine learning have become integral to production management, supply chain optimisation, and predictive maintenance.

Manufacturing also brings a distinctive infrastructure challenge because AI has to work across edge and core environments. Data may come from sensors, machines, video feeds, inspection systems, logistics platforms, and enterprise applications. Some workloads require fast processing close to equipment or production lines, while others need centralised analysis across plants, suppliers, and regions.

That creates a demanding storage and data management profile. Manufacturing AI needs infrastructure that can handle distributed data, mixed workloads, metadata growth, video and sensor data, and resilience across operational environments. As AI becomes more embedded in production, it becomes part of how operations run rather than a separate analytics layer.

Public sector AI depends on control and continuity

Public sector AI is shaped by a different set of pressures. Use cases may include citizen services, records management, transportation, public safety, benefits administration, research, or mission-oriented analytics. Many of these areas involve sensitive data, long retention requirements, and high expectations for service continuity.

For public sector organisations, the case for private AI often starts with governance and trust. Teams need clarity over where data resides, who can access it, how it is protected, and how services will recover if something goes wrong.

The survey’s broader findings support that view. In the report, private AI is defined as a model where AI workloads run on infrastructure the organisation controls, such as its own datacentre, a co-location facility, or a bare-metal hosting environment. The report also notes that sovereignty, compliance, and data proximity requirements are driving greater use of private deployments.

That does not make public sector AI less ambitious, but it does make the infrastructure requirements more explicit. AI systems in these environments have to be governed, resilient, and explainable enough to support public trust, particularly where automated systems touch services that citizens rely on.

Production AI depends on the full data pipeline

The vertical differences explain why AI infrastructure cannot be designed around a generic view of the enterprise. Yet the shared requirement is just as important: production AI depends on the full data pipeline.

That pipeline includes data preparation, ingestion, training, tuning, inference, model storage, vector stores, reference data, protection, recovery, and lifecycle management. Each stage can place different demands on storage, and the survey found that 86% of respondents recognise that different AI pipeline stages have distinct storage needs.

Storage performance is also emerging as a major concern. In the research, 57% of respondents said they are highly focused on preventing storage performance from becoming a bottleneck as AI activity and data volumes grow. That is higher than the percentages focused on compute and GPU availability or network bandwidth.

The finding is a useful corrective to the GPU-centred AI narrative. Compute remains critical, but production AI also depends on how efficiently organisations can feed models with data, retrieve data at runtime, protect pipelines, recover from incidents, and govern information across its lifecycle.

Object storage is increasingly part of that foundation. Across the survey, 91% of organisations reported meaningful use of object storage to support AI applications and pipelines, with 44% using it extensively and 47% using it quite a bit. That aligns with the realities of enterprise AI, where workloads often involve massive volumes of unstructured data — images, documents, logs, sensor data, and business records — that must remain accessible across tools, teams, and pipeline stages.

For every industry, the storage layer is where performance, resilience, sovereignty, and lifecycle control come together.

Experience changes the infrastructure conversation

The survey also shows that experience changes how organisations think about AI infrastructure. Seasoned adopters tend to define AI storage needs earlier in the project lifecycle, looking across the full AI-related data pipeline rather than treating each stage as a separate infrastructure problem. They also show a stronger preference for flexible storage platforms over individual point solutions, with the goal of preserving longer-term relevance and return on investment.

That shift becomes more important as AI programmes mature, because many infrastructure problems do not surface during early pilots. They appear when AI projects move into production, when more teams depend on the outputs, when data volumes grow, when compliance requirements become more complex, or when several AI workloads begin competing for the same infrastructure.

Healthcare and life sciences organisations need trusted data pipelines for sensitive and regulated data. Financial services firms need speed, scale, resilience, and auditability. Manufacturers need edge-to-core architectures that can support real-time operational insight. Public sector organisations need sovereignty, continuity, and strong governance.

The use cases differ, but the direction is converging. Across industries, organisations seeing the greatest success with private AI are not treating infrastructure as an afterthought. They are designing for the full data pipeline, from ingestion and storage to governance, protection, inference, and long-term reuse.

AI strategy is becoming data infrastructure strategy

As enterprise AI scales, success will not be defined only by who adopts the newest model first. It will also be defined by which organisations can put AI to work safely, reliably, and repeatedly across real business environments.

That requires infrastructure designed around data: where it lives, how it moves, how it is secured, how it performs, how it is governed, and how it is recovered. Private AI is gaining momentum because many organisations need that level of control, especially when AI workloads touch sensitive records, regulated processes, physical operations, or public services.

The industry lens makes the case clearer. Every sector has its own AI adoption pattern, but the infrastructure requirements are converging. Across healthcare, financial services, manufacturing, and the public sector, AI success increasingly depends on the ability to manage data across the full lifecycle, from ingestion and storage to governance, protection, inference, recovery, and long-term reuse.

About the author
Paul Speciale is CMO at Scality.