AI supply-chain thesis — mapping bottlenecks, focus companies, and supply-chain exposure for investors.
**Bottleneck theme:** Networking / Retimers **Focus:** $ANET — Arista Networks, Inc. Arista is the only Ethernet-AI fabric vendor with proven hyperscaler depth at GB200/GB300 scale. Where InfiniBand still owns the smaller, latency-critical AI training clusters, Ethernet is winning the bigger and faster-growing class of large-scale AI factories — and Arista's EOS software stack, deep buffering, and merchant-silicon strategy (Broadcom Tomahawk and Jericho ASICs) have made it the default switching vendor at Microsoft, Meta, and Oracle AI deployments. The Ultra Ethernet Consortium (UEC) and Broadcom's continuous switch-ASIC roadmap reinforce Ethernet as the long-term scale-up and scale-out fabric for AI. The investment case is hyperscaler AI capex compounding plus a structural fabric-share shift from InfiniBand to Ethernet. The bear case is concentration: top-2 hyperscalers drive a large share of revenue, and any in-sourcing pivot (or an NVIDIA-Spectrum-X share gain at the high end) would be highly material. Trades at a premium multiple to the broader networking complex — much of the AI-fabric optionality is already in the price. Pair with $AVGO (the silicon under Arista) and $CRDO (AECs to extend Arista racks).
Browse all AI supply-chain theses · Macro trends · Industries · Product categories