james regan

Electrical packet switching hits a hard wall at ~8 radix for 50 Tbps GPU links, forcing multi-tier switch fabrics that kill latency, power, and reliability. Nanosecond optical circuit switching enables single-hop, million-XPU scale-up fabrics, collapsing the scale-up/scale-out distinction and slashing transceiver count, power, and failure domains.

GPU/XPU egress bandwidth (50 Tbps and rising) is diverging from electrical switch roadmaps (200-400 Tbps chips), making the network the limiting factor for AI cluster scale. This shifts capex focus from compute to interconnect and favors radical photonic solutions over incremental electrical or slow optical circuit switching.

Keeping data as light through the switch fabric eliminates billions of optoelectronic converters, transceivers, and electrical switch ports. This directly reduces power draw, component count, supply chain pressure (evidenced by Nvidia's $2B+ optics bets), and infrastructure failure rates that plague large AI clusters.

European VC mindset of capital efficiency and incremental milestones fails in winner-take-all AI infrastructure markets where US competitors raise and deploy aggressively. Success requires hiring top talent at market rates, designing into high-volume supply chains from day one (avoiding pilot-line traps), and preserving equity for the multi-round capital journey.