7 There’s a conversation happening in enterprise IT right now that goes something like this: the AI tools are selected, the use cases are approved, the GPU budget is signed off — and then someone asks where the data actually lives. That’s when things get complicated. Most enterprises aren’t held back by AI capability. They’re held back by a data foundation that was never designed to support it. Fragmented across on-premises infrastructure, public clouds, and everything in between, enterprise data is often inconsistent, ungoverned, and frankly untrusted. According to Enterprise Strategy Group research, only 48% of organizations trust their data enough to confidently apply it to AI systems. That’s not a model problem. That’s a data strategy problem. A strong hybrid cloud data strategy doesn’t start with which cloud to use. It starts with understanding where your data lives, how it moves, and whether it can be trusted at scale. For AI workloads specifically, getting that foundation right is the difference between AI that works in production and AI that stays stuck in a proof of concept. Why Hybrid Cloud Is the Default for Enterprise AI Despite the growth of public cloud, two-thirds of enterprise data still resides on-premises. This isn’t inertia — it’s deliberate. Regulatory requirements, latency constraints, data sovereignty obligations, and sheer cost keep significant workloads anchored on-premises, while cloud environments handle burst compute, analytics pipelines, and model inference at scale. Gartner found that two out of three enterprises use hybrid cloud to power their AI initiatives. The reasoning is clear: AI doesn’t respect where your data happens to be stored. A fraud detection model needs transaction data from core banking systems. A genomics AI pipeline needs petabytes of sequencing data from lab infrastructure. A customer intelligence application needs behavioral data from multiple cloud-resident platforms. The enterprises getting AI right aren’t centralizing all their data before running models. They’re building a hybrid cloud data strategy that brings consistent access, governance, and performance to data wherever it lives. That’s a fundamentally different approach — and it changes what your storage infrastructure needs to do. The Hidden Cost of Data Fragmentation Ask most enterprise IT teams about their AI readiness and they’ll talk about GPUs, frameworks, and model selection. Ask them where their training data lives and the answer gets complicated fast. Data fragmentation is the quiet killer of AI initiatives. It develops gradually — a new cloud workload here, a departmental data store there — until you have multiple versions of the same datasets, inconsistent retention policies, access controls that don’t span environments, and no reliable way to know whether the data feeding your model is current or complete. For AI, this is particularly damaging. Model quality is a direct function of data quality. If your training data is fragmented across environments with inconsistent governance, you don’t have an AI problem — you have a data problem wearing an AI problem’s clothes. The cost isn’t just in model accuracy. Fragmented data means your teams spend time reconciling differences between systems instead of doing the work that matters. It means compliance gaps as policies fail to apply uniformly across environments. And in AI specifically, it means GPU clusters sitting idle because the storage layer can’t feed them fast enough. Understanding what an AI data center actually requires makes it clear how quickly storage becomes the limiting factor — not compute. What a Hybrid Cloud Data Strategy for AI Actually Requires A hybrid cloud data strategy that can support AI workloads at scale needs to address four things that traditional infrastructure was never designed to handle together. 1. A pipeline architecture built for AI from the start AI doesn’t consume data the way enterprise applications do. The volumes are larger, the access patterns are less predictable, and the consequences of pipeline failures are expensive. Understanding how AI data pipelines are architected — from raw ingest through preparation, training, checkpointing, inference, and archival — is the prerequisite for building storage infrastructure that won’t become the bottleneck. Each stage has different requirements. Ingest needs high-throughput sequential writes. Training needs low-latency random reads at high concurrency. Checkpointing needs durable, immutable writes. Long-term archival needs cost-efficient capacity with retrieval speed appropriate for RAG corpora. A hybrid cloud data strategy that tries to serve all of these stages with a single undifferentiated storage tier will fail at several of them. 2. Tiered storage that matches performance to cost Not all AI data is equal, and treating it as if it were is one of the most expensive mistakes in AI infrastructure. Hot training sets and active model artifacts need flash-speed access. Cleaned and labeled datasets awaiting their next training run can tolerate slightly higher latency at significantly lower cost. Cold archives and compliance copies need long-term durability at the lowest possible price per gigabyte. Tiered storage for AI is how enterprises reconcile the performance demands of GPU workloads with the economic reality of petabyte-scale data. Getting the tiering policy right — automated, policy-driven, and consistent across hybrid environments — is where most organizations find significant cost reduction without sacrificing the performance that actually matters. 3. Governance and security that spans every environment AI workloads are high-value targets. The datasets that train your models, the model artifacts themselves, the customer data powering inference — all of it needs protection with the same consistency regardless of where it lives. Applying different security policies across on-premises and cloud environments is how gaps develop. A production-grade hybrid cloud data strategy needs immutability enforced at the storage level, not just at the policy level. It needs encryption, access controls, and audit trails that work uniformly across environments. Multi-cloud storage architectures that lack unified governance create exactly the kind of fragmentation that makes enterprise AI unreliable and audit-unfriendly. 4. The ability to bring AI to your data, not your data to AI This is the principle that changes the economics of hybrid cloud AI. Moving data between environments is expensive in hard costs, latency, and operational complexity. The instinct to centralize all data before running AI on it is understandable — but at enterprise scale, it’s often impractical. Regulatory constraints may prevent it. Volume may make it cost-prohibitive. Latency may make it unacceptable. This is why cloud repatriation is an increasingly common outcome for enterprises that initially pushed everything to public cloud: the economics of keeping petabyte-scale AI datasets in cloud object storage don’t hold up once you factor in egress costs, API call volume, and the latency overhead of feeding GPU clusters from a remote location. The better model is infrastructure that brings AI tools to data wherever it lives — not the other way around. Object Storage as the Practical Standard for Hybrid AI Infrastructure Object storage has emerged as the default foundation for hybrid cloud AI data strategy, and the reasons are architectural. It scales horizontally without silos. It uses standardized S3 APIs that AI tools, frameworks, and analytics platforms already speak natively. It handles the mix of structured and unstructured data that AI pipelines consume. And it manages the full data lifecycle across both data lakes and more structured analytical environments without requiring separate systems for each stage. The performance question is where the conversation gets more nuanced. Standard object storage is well understood. But AI training workloads — simultaneously serving hundreds of compute nodes, handling checkpoint writes, managing concurrent ingest streams — push well beyond what conventional object storage was designed for. Object storage use cases for AI span the full pipeline, but the performance requirements at each stage vary significantly, and the storage architecture needs to reflect that. A major US bank running fraud detection AI on Scality RING achieves 80GB/sec read throughput per site — 160GB/sec aggregate across sites — serving a large Splunk analytics cluster. A global travel services company ingests one petabyte of log data per day into RING to maintain a 14-day rotating AI and analytics data lake. A French national genomics laboratory uses RING as the repository for 90% of its genomics data pipeline — nearly 10 petabytes — accessed from thousands of compute nodes at 3-5x lower cost than all-flash file storage alternatives. These are production numbers, not benchmarks. The Strategic Shift: From Infrastructure Distribution to Data Consistency Hybrid cloud started as a story about workload placement — which applications run on-premises, which run in the cloud. That question is largely settled. The harder question enterprises are now grappling with is how to make data consistent, trusted, and performant across all of those environments simultaneously. For AI, that question is urgent. Models trained on inconsistent data produce inconsistent results. Inference applications powered by fragmented data lakes produce outputs that can’t be trusted. The big data analytics challenges that have always existed at the storage layer are amplified when AI is the consumer — and the stakes for getting it wrong are higher. The enterprises winning at AI aren’t necessarily the ones with the most advanced models. They’re the ones with a hybrid cloud data strategy that can deliver clean, consistent, high-throughput data to whatever AI workload needs it — across on-premises infrastructure, multiple clouds, and everything in between. Infrastructure distribution was the first chapter. Data consistency is the one that determines whether AI actually delivers. Ready to see what this looks like for your environment? Scality RING is purpose-built for every stage of the enterprise AI data pipeline — from petabyte-scale data lake ingest through high-throughput model training, checkpointing, and long-term archival — across hybrid and multi-cloud environments without lock-in. Request a personalized demo with a Scality engineer → Further Reading AI Data Pipelines: Architecture, Stages, and Best Practices — A deep dive into how enterprise AI pipelines are structured, from raw ingest through inference and archival. Tiered Storage for AI: Scalable Performance and Cost Control — How to match storage performance tiers to the right stages of your AI pipeline without overspending on flash. What Is an AI Data Center? — The infrastructure components that separate purpose-built AI environments from general-purpose data centers. Data Lake vs Data Warehouse: Architecture, Cost, and Use Cases — Understanding which storage architecture fits which analytical and AI workload, and when to use both. Object Storage Use Cases — A practical breakdown of where object storage fits across backup, analytics, AI, and cloud-native applications. Multi-Cloud Storage: Architecture, Benefits, and Strategy — How to design multi-cloud storage architecture without creating the fragmentation that undermines AI reliability. Cloud Repatriation: Strategy, Costs, and Guide — Why enterprises are moving AI data workloads back on-premises and how to evaluate whether it makes sense for your environment. Solving Big Data and Analytics Challenges with Object Storage — Real-world examples of petabyte-scale analytics workloads running on object storage, including a 1PB/day travel services deployment. Multidimensional Scaling for AI and Cloud Storage — Why scaling capacity alone isn’t enough for modern AI workloads, and what dimensions actually matter. Hybrid Cloud Backup Explained: How It Works and Why It’s Used — The data protection dimension of hybrid cloud strategy — because AI datasets that can’t be recovered are AI datasets you can’t trust.
