What does it actually take to scale AI inference to millions of users in production? In this conversation, Guy Currier, Chief Analyst for Visible Impact & Research Director at the Futurum Group, sits down with leaders from AMD, CharacterAI, and DigitalOcean to pull back the curtain on one of the most demanding real-world AI deployments in the market today.

CharacterAI’s Chief Architect walks through the memory bottlenecks that were limiting their ability to serve larger, more immersive conversational AI models. AMD’s MI325X accelerators, with their substantial 256GB HBM capacity, enabled CharacterAI to fit the same models on half the hardware. But getting there required DigitalOcean and AMD to work in close lockstep — resolving early memory access issues, writing and upstreaming vLLM kernels, and ensuring the full hardware-software stack was production-ready before CharacterAI ever flipped the switch.

But the hardware is only part of the story. This discussion also covers the software stack, the collaborative process, and the organizational factors that enabled a complex platform migration in roughly three months. Whether you're an AI engineer wrestling with inference costs, an architect evaluating GPU platforms, or a technical leader navigating a major infrastructure transition, this discussion offers rare, ground-level insight into the decisions and tradeoffs that separate a smooth deployment from a painful one.

Scaling AI inference to millions of users is no longer theoretical. In this session, leaders from AMD, CharacterAI, and DigitalOcean sit down with Guy Currier, Chief Analyst for Visible Impact & Research Director at the Futurum Group, to walk through what it took to make it work in production.

At the center of the story is a hard constraint: memory. CharacterAI’s Chief Architect shares how memory bottlenecks limited their ability to run larger, more immersive models, and how AMD MI325X accelerators with 256GB of HBM enabled them to run those same models on half the hardware.

Getting there wasn’t plug-and-play. AMD and DigitalOcean worked side by side to resolve early memory access issues, contribute vLLM kernel optimizations, and bring the full stack to production readiness before launch.

This session goes beyond hardware to unpack the software, collaboration model, and decision-making that enabled a complex platform migration in just three months. If you’re evaluating inference infrastructure or managing the tradeoffs between cost, performance, and scale, this is a practical look at what it really takes.