How Many Will Actually Get Built & Is Energy AI's BIGGEST Bottleneck? | Positron AI Co-founder

AI infrastructure is shifting from compute-bound training to memory-bound inference, a transition necessitated by the transformer revolution. Thomas Sohmers, co-founder of Positron AI, identifies a critical "memory wall" where memory bandwidth has failed to scale at the same rate as compute, creating significant bottlenecks. Efficient inference relies on KV caching to avoid redundant computation, though this strategy introduces complex memory storage demands. While public opposition to data centers often cites environmental concerns, these facilities are essential economic drivers, and current anti-data center sentiment represents a strategic geopolitical disadvantage. Scaling laws remain the primary path to AGI, with future models evolving into autonomous agents that will trigger massive increases in global token consumption. As enterprises integrate local models with cloud-based frontier intelligence, the cost per unit of capability will continue to plummet, fundamentally transforming business productivity and the global data economy.