Scaling AI Up, Out, and Across: Panel hosted by Roy Chua, Founder and Principal Analyst @avidthink

Scaling AI Up, Out, and Across: A Decision Framework for Networking Transitions Reshaping AI Infrastructure Economics The AI industry will spend over $700 billion on infrastructure in 2026, yet even well-run organizations convert only about 20% of that into model-advancing work — with networking overhead among the primary constraints. Next-generation data centers won't be defined by faster chips, but by who moves data most efficiently between them at scale. The network is now three coupled design problems — scale-up, scale-out, and scale-across — each with different physics, economics, and operational realities. Get it wrong and you're burning GPU cycles — at today's capital intensity, that's lost capital and slower time-to-model. Whether you're building pre-training fabrics at scale, deploying sovereign inference at telco or neocloud scale, or running enterprise AI pilots, you face the same core question: how do you get more effective utilization out of expensive GPU/XPU clusters? The panel will discuss and debate three principles that will shape how operators think about AI fabrics over the next 18 months.