More agents still mean more computing power demand, but this new demand won't necessarily be met entirely by the most advanced GPUs. While cutting-edge model training and inference still require high-end chips, agents performing numerous routine tasks may have lower real-time requirements and can utilize more traditional CPUs or non-cutting-edge chips. As agents become more widespread, the demand for AI infrastructure may further expand to CPUs, memory, storage, networks, and task orchestration systems. AI transactions may also shift from being highly concentrated on GPUs to a broader range of computing infrastructure. (JPMorgan Chase report, September 23)
