Breaking through the Network Wall with High Radix Switching

Author: Mike Capuano, Chief Business Development and Marketing Officer, Eridu Collection: 2026-OCP-EMEA-Summit-Executive-Sessions

Summary

Mike Capuano, Chief Business Development and Marketing Officer at Eridu (originally referenced as "Arrcus" in speech but title slide shows "Eridu"), presents the case for a 10x radix increase in AI networking silicon. The core argument is that AI networking is the primary bottleneck to GPU utilization — GPUs sit idle up to 60% of the time, networking consumes >25% of IT power, and network cost represents 20% of IT spend. Current state-of-the-art cloud switches (51.2T SOTA, 102.4T announced for 2H 2026) were designed for cloud workloads and are fundamentally inadequate for AI scale-out and scale-up demands driven by exponentially growing GPU FLOPS, Mixture of Experts models with trillions of parameters, agentic workflows with hundreds of API calls per prompt, disaggregated prefill/decode, and KV cache transfer across racks.

The proposed solution — a hypothetical 1 petabit (1P) high-radix switch — replaces 30 conventional switches with a single chip by collapsing 1,000+ optical transceivers and inter-switch fiber into a 2 mm die-to-die silicon connection. This yields 70% power reduction, 40% cost reduction, single-hop latency, and massively reduced cabling. For scale-out, the comparison is stark: a two-tier cluster using 102.4T switches achieves a pod of 1,008 GPU packages in 14 racks using 6,144 total switches; the same 1P switch design achieves a pod of 10,244 GPU packages in 142 racks using only 672 switches — 9x fewer switches for 10x more GPUs per pod. For scale-up, NS3 packet-level simulation with partner Scala Computing models a 5,120 GPU scale-up domain: the 1P switch delivers a 5.26x speedup in collective communication completion time over a 230T switch on a 500B-parameter MoE model, with AllToAll improving 10.64x.

Eridu was founded in Silicon Valley in 2024, exited stealth in March 2026, has raised $260M through Series A, employs ~150 people, and is developing new high-radix switching silicon using advanced packaging. The call to action is industry collaboration: optimizing flat scale-out architectures, engaging OCP workstreams, and working with UEC, UALink, and Ethernet-based fabric standards.