Dylan Patel on the infrastructure powering the AI revolution | The Next Big Thing

The AI infrastructure landscape is undergoing a fundamental shift as reasoning and agentic workflows replace standard chat-based models, driving unprecedented demand for memory and CPU capacity. SemiAnalysis founder Dylan Patel explains that this evolution necessitates a move beyond GPU-centric architectures, as agentic tasks require significant CPU power for environment checking and code execution. Memory remains a critical bottleneck, with prices soaring due to limited capacity growth relative to doubling demand. Furthermore, energy availability has emerged as the primary gating factor for data center expansion, forcing a transition toward on-site, behind-the-meter power generation using diverse industrial technologies. While networking content is also rising, the industry faces complex supply chain dislocations, with co-packaged optics delayed in favor of continued copper innovation in the near term. These infrastructure constraints are reshaping the entire semiconductor ecosystem, requiring deep technical analysis to navigate.