✦ Ethernet Everywhere: Tomahawk, Thor & JerichoNarrow moat

Broadcom's networking bet is that AI clusters wire together the way the internet did — on open Ethernet, at every scale. The arsenal shipped fast: Tomahawk 6, the first 102.4-terabit switch, in production volume — two tiers of them can fabric 128,000 XPUs; Tomahawk Ultra at 250-nanosecond latency aimed directly at NVLink's scale-up role, with systems arriving through late 2026; and Thor Ultra 800G NICs at the endpoints¹. Jericho4 extends the wager across buildings — lossless fabrics interconnecting more than a million XPUs between data centers².

Every custom-XPU win pulls this networking through — the OpenAI racks are scaled entirely on Broadcom Ethernet — and every Ethernet win makes the next custom chip easier to justify: the two franchises sell each other. The opponent is formidable and fights on both boards: Nvidia's NVLink guards the scale-up socket while its Spectrum-X ships Ethernet too, so Broadcom must win the open standard and the share inside it. Watch scale-up decisions in next-generation clusters — Tomahawk Ultra displacing NVLink anywhere would be the signal — and watch the Ultra Ethernet ecosystem's cadence against Nvidia's annual clock.