The AI infrastructure supply chain has a new pressure point, and it’s not where most investors are looking. Semianalysis, the semiconductor research firm led by Dylan Patel, has identified deepening bottlenecks in server rack systems that are fueling renewed interest in Supermicro as a primary beneficiary of the crunch.
Supermicro’s stock moved higher on the back of the research, which paints a picture of an AI buildout increasingly constrained not just by GPU supply but by the physical infrastructure needed to house, cool, and power those chips at scale.
The bottleneck is shifting
In a report titled “CPUs are Back,” Semianalysis argued that reinforcement learning workloads are placing enormous strain on CPU resources. GPU accelerators have advanced so quickly that they’re now outpacing the CPUs they depend on, creating a mismatch that could become a real headache for data center operators by early 2026.
Dylan Patel expanded on this theme during an appearance on the Dwarkesh Podcast, noting that scaling bottlenecks in AI are progressively moving towards logic, memory, and power. He pointed specifically to semiconductor production capacity, including TSMC’s N3 node, as a critical long-term constraint for the industry.
