Cloud repatriation: optimising infrastructure for AI at scale

As organisations scale AI in production, particularly those pursuing private or sovereign AI strategies, they increasingly encounter constraints that are fundamentally data-centric. That is one of the clearest takeaways from independent research from 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 report highlights how this shift is playing out in practice. Growing workloads amplify the need for consistent storage performance, robust data governance, and high-throughput data movement, making data and storage infrastructure as critical to AI scalability as GPUs. Notably, 91% of enterprises that are running private AI in production reported in the aforementioned survey the meaningful use of object storage, impressively underscoring the centrality of the data layer in operational AI systems.

This reality challenges the long-held assumption that compute is the primary bottleneck. While GPU availability has understandably dominated AI scaling discussions, many organisations are discovering that the movement, accessibility, and integrity of data can be equally constraining.

Rapid cloud adoption initially offered a quick way to extend capacity, but it often came at the expense of a holistic architectural review. Now, as AI matures into sustained production at scale, organisations are intentionally revisiting their entire architecture. Their focus is now on strengthening core systems, modernising stable environments, and aligning workloads with the most suitable execution contexts, whether cloud, on-prem, or hybrid. Cloud repatriation, in this sense, reflects a strategic evolution: it is a precision-driven approach to operational design rather than a simple reversal from cloud. 

For years, cloud-first strategies have shaped IT roadmaps, driven by the rise of LLMs and AI services delivered through hyperscale platforms. However, in today’s evolving landscape, organisations are moving beyond a cloud-first mindset, with a clear focus to optimise the infrastructure.

Being infrastructure-savvy combines an understanding of the cloud’s benefits with the ability to make dynamic, intelligent choices about services across infrastructure, platforms, and software. It requires selecting the right combination of solutions, maintaining clear visibility into pricing to avoid unexpected costs, and ensuring services are effectively managed and secured. Equally important is the deliberate placement of data and workloads, so they achieve the optimal balance of performance, control, and efficiency at scale.

For instance, a bank may choose to keep customer and transaction data on-prem to meet strict security requirements, while using the public cloud to power customer-facing mobile applications that benefit from global reach. Similarly, in life sciences, organisations might retain proprietary research data or genomic datasets in private environments for compliance and intellectual property protection, while leveraging cloud platforms to run large-scale simulations or accelerate drug discovery pipelines where elastic compute provides a clear advantage.

Cloud repatriation is also emerging as a natural response to the inherent limitations of hyperscale public cloud offerings for long-term AI deployment. While the cloud offers scalability and flexibility in the short term, many enterprises are finding that their evolving needs for data residency, high-throughput performance, and predictable costs are better met by tailored hybrid or on-prem solutions. This shift reflects a growing realisation that optimising for long-term AI goals requires more than just relying on cloud elasticity: careful control over the infrastructure that supports AI at scale is emerging as a game changer.

Data sovereignty, compliance, and cost management are key drivers to bring AI workloads back on-prem or into hybrid environments, especially in regulated sectors such as finance and life sciences. Sovereign AI solutions let enterprises control where and how data is processed, supporting regulatory adherence while enabling advanced machine learning capabilities, with 81% of enterprises reporting in the recent survey that private AI infrastructure under their control is critical to success.

 

Cloud repatriation addresses both governance and performance concerns, allowing sensitive workloads to remain in controlled environments while still leveraging cloud-native tools. At the same time, cost considerations are significant: public cloud services, though flexible, can lead to unpredictable expenses through egress fees and constant data movement for AI workloads. Moving workloads on-prem or into hybrid setups helps organisations predict and manage spending, avoid hidden fees, and optimise infrastructure for both efficiency and compliance.

As models scale, it becomes more and more obvious that the ability to feed them data efficiently often determines overall system performance just as much as compute availability. This is driving a broader shift toward decoupling storage and compute, allowing each to scale and evolve independently. In this model, data is no longer statically tied to a specific compute environment but can be accessed flexibly with high throughput across distributed workloads.

 

Object storage naturally fits into this architecture as a way to manage large volumes of unstructured data, supporting training, inference, and ultimately fine-tuning without requiring close coupling between where data lives and where it is processed. Rather than treating AI as an add-on to legacy systems, enterprises are adopting tiered, hybrid architectures, supporting diverse access patterns across the AI pipeline and thus avoiding bottlenecks.

 

Businesses are increasingly recognising that no single environment meets all workload needs. On-prem and hybrid models now complement cloud infrastructure, enabling intentional workload placement. Cloud repatriation is experiencing a heyday, but far from a total rejection of the cloud, this trend is strategic and focused on optimisation. It is an approach that allows organisations to create a more resilient, agile and ultimately future-proof AI ecosystem.