Modern AI workloads demand more than just powerful GPUs—they require an entire infrastructure stack that works in sync. That’s the premise behind Mirantis’ AI Factory Reference Architecture, which CTO Shaun O’Meara and VP Randy Bias unveiled in their latest TFiR appearance.

Why GPUs Aren’t the Whole Picture

According to O’Meara, most teams focus on compute, but AI infrastructure has become more complex than that. “It’s not just about GPUs,” he said. “It’s about the full infrastructure stack. We’ve returned to something that looks more like the mainframe era—where applications are tightly coupled to the hardware.”

This tight coupling means infrastructure needs to be built with precision—ensuring that every layer, from networking to storage to orchestration, is optimized for AI workloads.

From Enterprise to Supercomputing Principles

Bias adds that AI infra is evolving beyond the typical enterprise data center. “These systems have more in common with high-performance computing than cloud-native design,” he said. That’s why the reference architecture emphasizes non-blocking networks, east-west traffic optimization, and latency minimization—hallmarks of supercomputing.

A Living Blueprint

Mirantis has published a comprehensive white paper—already 65+ pages long and growing—which serves as a living architecture guide. It distills lessons from decades of infrastructure engineering and applies them to today’s most demanding use cases.

For teams building serious AI workloads, this clip is a solid primer on the why behind infrastructure complexity—and what’s needed to solve it at scale.