Scaling an AI project to production depends not only on the model chosen, but also on having a ready cloud infrastructure and an operational foundation capable of supporting it.

Many companies have already moved beyond the pilot phase with artificial intelligence, but turning those projects into real, scalable, and sustainable solutions remains one of the biggest challenges. For Miquel Pujol, global director of infrastructure services at SEIDOR, the issue is not just which technology is adopted, but whether the organization’s infrastructure and operational foundation are truly ready to support it.

When AI moves beyond the pilot phase and begins to be integrated into business processes, the requirements change. The company needs an infrastructure capable of meeting demands in terms of capacity, availability, and security, but also of ensuring well-governed access to data. That’s where cloud infrastructure establishes itself as a necessary foundation for scaling AI. As Pujol points out, it allows companies to “adjust resources based on demand, integrate advanced data and AI services, and operate with greater flexibility.”

However, having the infrastructure in place is only half the battle. “There’s often a lack of the operational foundation needed to turn a pilot into a business solution. A pilot can work with a small number of users, limited data, and a controlled environment. The challenge arises when connecting it to critical systems,” adds Pujol.

 

There is no reference architecture

Every company starts from a different situation, and every workload has different requirements: some workloads require proximity to the data, others need scalability, and still others demand intensive processing power.

That is why Pujol rules out generic approaches: "The key lies in combining public cloud, private cloud, on-premises environments, and managed services in a coherent way, with an architecture designed from the ground up for data, AI, and cybersecurity." This approach strongly drives the development of the sovereign multicloud model—a trend that is gaining momentum precisely because it allows organizations to tailor their architecture to their actual needs without relinquishing control.

The responsibility for getting that combination right does not rest solely with the in-house technology team; rather, it is advisable to seek expert advice to refine the approach. “The company needs to know which combination of solutions is right for it, how to integrate them with its current systems, how to operate them securely and efficiently, and how to measure the value generated,” says the SEIDOR specialist.

 

Efficiency from the design stage, not as an afterthought

AI models do not behave like traditional workloads. A complex query can consume hundreds of times more resources than a conventional transaction. Many companies discover this discrepancy once the project is already underway, and the impact on their bill is difficult to reverse without redesigning part of the architecture. It is one of the factors that most influences migration and cloud strategy in environments where AI is already part of day-to-day operations.

Adding more servers or purchasing more capacity without a clear strategy can exacerbate the problem rather than solve it. “A well-planned architecture allows you to scale resources, avoid overcapacity, optimize workloads, and reduce unnecessary consumption,” notes the head of infrastructure services at SEIDOR. Therefore, efficiency is not a secondary goal to be addressed only when costs spiral out of control—it is a design decision that must be made from the very beginning.

That operational foundation must also be supported by mechanisms that enable the infrastructure to be managed efficiently and sustainably as projects grow.

There are two specific tools that can help move in that direction. The first is the adoption of FinOps practices, which transform cloud spending management into an ongoing dialogue between technology and business: the goal is to understand what each decision consumes and whether the return on investment justifies it. The second is selecting providers and infrastructure based on energy efficiency criteria from the outset. “AI forces us to innovate more rigorously: it’s not just about having more capacity, but about using it better—through planning, automation, control, and a long-term vision,” Pujol summarizes.

 

Data first, tools second

In the face of these challenges, the starting point must be the data and the processes you want to transform. Once that scope has been defined, it’s time to review the architecture: which systems are critical, which workloads can be moved to the cloud, which environments should remain on-premises or hybrid, and how security and continuity are ensured.

Only then does it make sense to discuss which AI technology to adopt and how to integrate it.

As Pujol concludes: "The recommendation is not to start with the AI tool itself, but rather with the foundation that will make its adoption in production feasible: governed data, scalable architecture, built-in security, and a clear roadmap for business impact."

 

Cloud & AI Infrastructure 2026: The Event to Gain a Global Perspective

Cloud & AI Infrastructure 2026, taking place on November 4 and 5, 2026, at IFEMA Madrid as part of Tech Show Madrid, is the premier event for companies seeking to evolve their cloud architecture toward more agile, scalable, and AI-driven models. It brings together more than 60 exhibitors, nearly 60 speakers, and over 3,000 professionals, including technology, infrastructure, data, and business leaders. The program features real-world case studies on multicloud, AI Cloud platforms, automated FinOps, cloud resilience, and data governance—all designed to help organizations transition from pilot projects to production environments with tangible results.