Volpe predicts shift to circuit-switching for AI data centers
Hanabi Capital’s Mike Volpe argues that current statistical multiplexing architectures are ill-suited for predictable AI workloads. He predicts a move toward circuit-switching designs that could reduce hardware costs and disrupt incumbents.
Mike Volpe argues that current data center networking architectures, based on stochastic statistical multiplexing like Broadcom’s Tomahawk switches, are ill-suited for AI workloads. He notes that AI data centers exhibit predictable, constant traffic patterns with high link utilization, unlike the variable traffic of the internet. Volpe predicts a shift toward circuit-switching or time-division multiplexing architectures that assume full utilization and predictability, potentially reducing the need for complex buffering and lookup logic. This architectural change could make networking hardware significantly cheaper, creating an opportunity for new entrants to disrupt incumbents like Cisco and Arista. Volpe points out that Google has pioneered optical circuit switching in its toroidal network designs, though he questions the slow switching speed of approximately one minute currently observed. He suggests that because the network is becoming an integral component of the compute system, innovation in this area may need to come from GPU vendors like NVIDIA or Google, who can ship complete rack systems, rather than standalone switch vendors. The current innovator’s dilemma means existing vendors are unlikely to abandon their profitable statistical multiplexing architectures for a simpler, cheaper alternative.