Enterprise AI shifts focus from model choice to platform control

As agentic AI infrastructure moves from proof-of-concept into production, enterprises are discovering that the technology introduces challenges beyond model selection. Joe Fernandes, vice president and general manager of the Artificial Intelligence Business Unit at Red Hat, outlined the strategic shift in a conversation with theCUBE's Rob Strechay. The core issue is not choosing between different AI models but controlling the cost, data exposure, and infrastructure that underpin production AI applications.

Fernandes identified two primary drivers of this pivot. First, token costs—the expenses incurred each time an AI model processes input and generates output—are rising sharply as organizations scale from simple chatbots to always-running enterprise agents. This transition from experimentation to production at scale makes reliance on public cloud AI services economically untenable. Second, data sovereignty and compliance concerns are now decisive factors. Enterprises must decide whether they can afford to send their sensitive data into public cloud environments, and many cannot due to regulatory requirements or corporate policy. "I think cost is a huge factor, but then there's also the data side: Are you willing to let your data go into these public cloud services?" Fernandes said. "I think those are the two things combined that make them start thinking, 'Maybe there's an alternative here.'"\n\nThis realization is elevating the role of platform teams within enterprise organizations. Where AI was once a specialized function, it is now becoming integrated into the core enterprise platform strategy. Platform teams must ensure not only reliability, uptime, scale, and security—traditional operational concerns—but also manage the autonomous behavior of agents. The defining characteristic of an agent is its autonomy, which introduces new operational questions: what network resources and file systems can an agent access, and how can enterprises trace and control agent actions? To address this, Red Hat has introduced the concept of agent sandboxes—isolated execution environments that constrain what agents can reach and do. Red Hat is extending its open-source approach to support this infrastructure, including active contribution to and maintenance of OpenShell, Nvidia's open-source sandbox runtime for AI agents. Fernandes described OpenShell as "a key project for us now in the agent sandbox space," positioning open-source collaboration as the path to driving innovation and establishing industry standards.\n\nThe broader implication is that agentic AI infrastructure will eventually be hybrid by necessity. Governments and organizations will require control over where models, agents, and data run, which Fernandes argued necessitates a multi-environment architecture spanning public cloud services, private data centers, sovereign clouds, and edge systems. This vision aligns with Red Hat's longstanding position that enterprise IT infrastructure will always be distributed and hybrid.