Cloud Repatriation: Optimizing Infrastructure for AI at Scale

Key Takeaways

• Cloud repatriation is a strategic realignment, not a retreat from the cloud.

• 91% of organizations running private AI in production rely heavily on object storage (Freeform Dynamics survey of 504 organizations, commissioned by Scality).

• 81% say a private AI infrastructure under their own control is business-critical.

• Data movement, accessibility, and integrity are becoming just as limiting for AI as GPU availability.

• The winning model: deliberate workload placement across on-prem, hybrid, and public cloud.

What Is Cloud Repatriation?

Cloud repatriation is the strategic relocation of data and workloads from public cloud environments back to on-premises data centers, external data centers, or private clouds. It should not be understood as a departure from public cloud services — rather, it is a deliberate realignment of architectural strategy.

The focus is on modernizing core systems to meet evolving operational requirements. A recent survey of 504 organizations, conducted by the independent analyst firm Freeform Dynamics on behalf of Scality, illustrates how this shift is playing out in practice.

Cloud repatriation in practice: AI workloads are strategically distributed across cloud, on-premises and GPU infrastructure.

Why AI Is Driving the Shift

As organizations scale AI in production, they increasingly encounter constraints that are fundamentally data-driven. Growing workloads intensify the need for three things:

• Consistent storage performance: predictable throughput under heavy, parallel AI workloads.

• Robust data management: governance and integrity across the entire pipeline.

• High-throughput data movement: feeding GPUs fast enough to keep them utilized.

As a result, data and storage infrastructure is becoming just as critical to AI scalability as GPUs. Notably, 91% of organizations running private AI in production highlight significant use of object storage — a striking indication of the central role the data layer plays in operational AI systems.

Rethinking Architecture from the Ground Up

These findings challenge the long-held assumption that compute power is the primary bottleneck. While GPU availability dominates many discussions around AI scaling, more and more organizations are recognizing that the movement, accessibility, and integrity of data can be equally limiting. Cloud adoption initially offered a fast path to capacity expansion — often at the expense of a holistic architectural review.

As AI matures for production use at scale, organizations are deliberately rethinking their entire architecture: strengthening core systems, modernizing stable environments, and aligning workloads with the execution contexts best suited to them — cloud, on-prem, or hybrid.

From Cloud-First to Cloud-Smart

For years, cloud-first strategies shaped IT roadmaps, driven by the meteoric rise of LLMs and AI services on hyperscale platforms. Today, organizations are moving toward a cloud-smart approach: understanding the benefits of the cloud while making intelligent, dynamic decisions about infrastructure, platforms, and software.

Cloud-smart in practice means:

• The right solution mix: matching each workload to its optimal environment.

• Transparent cost control: avoiding unexpected expenses such as egress fees.

• Effective governance: managing and securing services consistently across environments.

Two Real-World Examples

• Banking: A bank stores sensitive customer and transaction data on-premises to meet strict security requirements, while using the public cloud for customer-facing mobile apps that benefit from global reach.

• Life sciences: Research organizations keep proprietary genomic data in private environments to protect compliance and intellectual property, while leveraging cloud platforms for large-scale simulations where elastic compute offers clear advantages.

Control, Compliance, Costs: The Three Drivers

• Control (data sovereignty): Sovereign AI solutions let organizations decide where and how data is processed. 81% of surveyed organizations call a private AI infrastructure under their own control business-critical.

• Compliance: Sensitive workloads remain in controlled environments — essential in regulated industries such as financial services and life sciences — while cloud-native tools continue to be used.

• Costs: Public cloud services can lead to unpredictable expenses through surprise egress fees. On-prem and hybrid setups make spending predictable and eliminate hidden costs.

Rebalancing Data and Compute

As models scale, the efficient delivery of data often determines overall system performance just as much as raw compute. This is driving a broader movement toward decoupling storage and compute: data is no longer statically tied to one environment but flows at high throughput across distributed workloads.

Object storage fits seamlessly into this architecture — managing large volumes of unstructured data and supporting training, inference, and fine-tuning. Rather than treating AI as an add-on to existing systems, organizations are adopting tiered hybrid architectures that support diverse access profiles across the entire AI pipeline.

Conclusion: Beyond Cloud-Only

No single environment can meet all workload requirements. On-prem and hybrid models now complement cloud infrastructure, enabling targeted workload placement. Cloud repatriation is a strategic pivot focused on optimization — not a retreat from the cloud. This adjustment enables organizations to build a resilient, agile, and future-proof AI ecosystem.