Michael Lane, VP of Networking at Hyve Solutions, presents a framework for next-generation switch architecture requirements for AI hyperscale data centers at OCP Global Summit 2025. The talk covers five key networking requirements — bandwidth, latency, telemetry, security, and topology — and explores emerging trends including DLC-cooled switches, software-defined networking evolution toward AI-based orchestration, and optics transitions from EBO to CPO. A forward-looking roadmap projects switch ASIC scaling from 51.2T to 204.8T and topology migration from Clos to Dragonfly between 2025 and 2030, with a call to participate in OCP's Open Systems for AI (OSAI) initiative.

Key Takeaways

• AI cluster switches must support ultra-high bandwidth (multi-terabits per rack), sub-microsecond latency, and zero-trust security — all simultaneously, without the usual latency-security tradeoff being ignored.
• NVIDIA's NVL72 rack outputs 72 Tbps across 180 connections and 2,880 fiber connections; the NVL144 doubles GPU count and thermal footprint, making DLC-cooled switches essential by 2026.
• Broadcom has halved switch latency for AI workloads, from 600-750 ns down to ~250 ns, demonstrating that silicon vendors are actively co-evolving with hyperscaler requirements.
• Dragonfly topology will become necessary for large AI clusters from ~2027 onward, though its primary disadvantage is cabling complexity for operators and ODMs building out physical infrastructure.
• Switch thermal management is shifting: 1.6T optical transceivers can exceed 24W per port, making DLC-cooled 2RU switches mainstream by 2026 and LPO/CPO the key mitigation technologies.
• SDN orchestration must expand beyond network devices to cover GPUs, servers, and DLC management; AI-based predictive orchestration is projected to replace pure SDN starting around 2026.
• Switch ASIC bandwidth roadmap: 51.2T (2025) → 102.4T (2026-2028) → 204.8T (2028-2030); SerDes shift from 100G to 200G tightens SI constraints on switch PCB designs.
• OCP's Open Systems for AI (OSAI) initiative is forming as a consortium to define open infrastructure standards for AI workloads — Hyve Solutions is a participating member.