Nvidia's Spectrum-6: The Ethernet Play That Kills InfiniBand's Moats

I saw the price action before I read the specs. Nvidia unveils Spectrum-6 — 102.4 Tb/s of Ethernet switching for "gigascale AI factories." The market yawns. But I've been watching the order flow in AI infrastructure for months. Something's off. Here's the context. For years, InfiniBand was the only game for large-scale GPU clusters. It's fast, proprietary, and expensive — a perfect lock-in. Nvidia owns both Quantum (InfiniBand) and Spectrum (Ethernet). Why cannibalize your own cash cow? Because the data shows institutional buyers are voting with their wallets. Meta, Oracle, Cisco, Nebius — they're all signing up for Spectrum-6. They're not buying InfiniBand for the next gen. That's a signal. The core mechanic is brutal. Spectrum-6 packs 102.4 Tb/s switching capacity — top-tier silicon. But hardware is table stakes. The real alpha is in Nvidia's software stack: CUDA Net, NCCL, and their BlueField DPU orchestration. They've tuned the congestion control for RoCE v2 specifically for AllReduce communication patterns. Translation: when you run a 10,000-GPU training job, the network becomes the bottleneck. Nvidia is selling the entire pipe — from GPU to NIC to switch — and guaranteeing the latency profile. No third-party vendor can match that integration. I've audited enough smart contracts to know vertical integration = margin compression for competitors. Now the contrarian angle. The retail narrative is: "Ethernet is open, so this helps AMD and Intel GPUs." No. Look at the fine print. The Spectrum-6 firmware and driver layer include proprietary optimizations that only fully unlock when paired with Nvidia's SuperNIC and H100/B100. You can plug an AMD GPU into the network, but you won't hit the advertised throughput. It's a soft lock — not as tight as InfiniBand, but strong enough to keep customers inside the Nvidia ecosystem. Pain is just tuition; I paid in full during the Terra collapse when I ignored code-level risks. I'm not ignoring this. The real losers are Arista and Cisco — their general-purpose switches can't compete on AI-tailored performance. Bet against them. Takeaway: watch the optical module suppliers — 800G/1.6T demand just got a catalyst. But if you're long Nvidia, this isn't a new story. It's confirmation they're building the rails for the next compute paradigm. I didn't come here to make friends. I came here to make alpha. The rest is noise. We don't trade narratives. We trade order flow. Spectrum-6 is an order flow event, not a headline.