Dylan Patel (SemiAnalysis): The Datacenter in 2026: CPUs, RL Environments & Agent-Driven Workloads

The conversation centers on the shift in AI infrastructure bottlenecks from GPUs to CPUs, driven by the increasing complexity of AI models and agentic workflows. Dylan Patel explains that while GPUs were the primary focus initially, the rise of reinforcement learning and AI agents performing complex tasks like database calls and physics simulations has significantly increased CPU demand. This demand has led to shortages, impacting infrastructure stability, as seen with GitHub's recent performance issues. Patel notes that companies are even porting codebases to different CPU architectures, like ARM, to secure necessary capacity. The discussion also touches on the rising prices of CPUs, memory, and storage due to increased demand and manufacturing constraints, affecting the cost of building PCs and servers.