The switch revolution in the era of AI supernodes

Why switches are suddenly strategic

AI model sizes have exploded from billions to trillions of parameters, and that growth is shifting the bottleneck away from single‑GPU FLOPS to cluster networking. RDMA (Remote Direct Memory Access) and its GPU‑focused descendant, GPU Direct RDMA, are now the plumbing that lets dozens to thousands of GPUs behave like a single “supernode.” Without low‑latency, zero‑copy memory transfers, large‑model training stalls — jobs take longer and GPU utilization collapses. Who wins at the network layer directly shapes who wins at AI scale.

A protocol fight: InfiniBand vs. Ethernet

InfiniBand — led historically by Mellanox and later absorbed into NVIDIA — remains the latency and reliability leader, delivering sub‑2µs links and proven congestion control. But Ethernet has not stood still. RoCE (RDMA over Converged Ethernet) evolved to RoCE v2 to support routed deployments, and industry players are now trying to close the remaining gap. It has been reported that Broadcom, Microsoft and Google unveiled a Ultra Ethernet Consortium (UEC) 1.0 specification in June 2025 to push Ethernet performance toward InfiniBand levels. It has also been reported that Chinese cloud and infrastructure players including Alibaba (阿里巴巴), Baidu (百度), Huawei (华为) and Tencent (腾讯) have joined the effort — a sign that Ethernet’s ecosystem advantages and lower cost are drawing broad support.

Hardware race and market impact

The technical tug‑of‑war is translating into a hardware sprint. High‑density switches, 100–400G (and beyond) ports, co‑packaged optics and multi‑Tbps ASICs are suddenly the hottest products in data‑center procurement plans. NVIDIA’s Spectrum‑x and Rubin AI moves, and Broadcom’s Tomahawk‑6 102.4Tbps silicon, illustrate how cloud vendors and silicon houses are repositioning to serve supernode fabrics. Analysts expect Ethernet share in AI fabrics to grow substantially in the next few years, reshaping demand for switches, NICs, optical modules and cabling.

Geopolitics, supply chains and the play for self‑reliance

This is not just a technical contest. Supply‑chain concentration — particularly around NVIDIA’s InfiniBand stack and US semiconductor leadership — intersects with trade policy and export control risk. As a result, Chinese vendors and cloud providers are accelerating domestic stacks and standards participation to reduce dependency. The result? A global, multi‑front race: better latency and congestion control, lower cost per port, and geopolitical resilience. Who will dominate the next generation of AI supernodes — incumbents with a latency edge or the broad Ethernet ecosystem aiming for scale and openness? The switch market may hold the answer.