AI Networking — Strategy
Updated 6/19/2026
Where AI Networking is heading over the next 12 months, grounded in product-axis evidence and verbatim demand from the last 90 days. The judgment column is the engine's read — operators verify and refine.
Product trajectories
Lossless Ethernet / RoCEv2 AI Fabric (Ultra Ethernet) — fabric architecture / protocol 🔥 hot
Opportunity: GPU training/inference fabrics demand RDMA-class semantics on Ethernet; UEC 1.0 targets ~1M endpoints and ESUN extends Ethernet into scale-up. Community discussion is explicitly tracking the InfiniBand → UE transition.
Architectural inflection where three incumbents are committing PM and silicon talent. ESUN + UEC give it standards backing. Open competitive question vs NVIDIA Spectrum-X / MRC and reintroduced Omni-Path.
800G / 1.6T AI Ethernet Switching Platform (51.2T–102.4T class) — data center switch ↗ rising
Opportunity: Hyperscaler GPU clusters scaling past 100K endpoints need merchant-silicon, lossless 800G/1.6T Ethernet at radix levels InfiniBand cannot match cost-effectively (fact e070c95f Tomahawk 6 102.4T; fact 9b6493fe Arista $750M FY25 AI back-end).
Clear breakout product: Arista, Cisco, and Celestica (around Broadcom Tomahawk 6) are all converging on 51.2T → 102.4T Ethernet as the de-facto AI back-end fabric, with hyperscaler design wins already monetizing. Remaining gap is congestion control / lossless behavior parity with InfiniBand at scale.
Active Electrical Cable (AEC) / Active Copper Cable for AI Racks — high-speed copper interconnect ↗ rising
Opportunity: Inside-rack GPU-to-switch links need lower power/cost than optics at sub-3m reach; AWS Trainium and Nvidia rack-scale designs are productizing this exact slot (facts ba213d85 1.6Tbps AEC; cb28d21e Semtech >100% YoY datacenter).
Hot, concentrated: Credo (AWS anchor, ~80% revenue concentration per fact 169a5e0e) and Semtech CopperEdge are the two pure-plays. AEC is now an established AI rack-scale product, not a thesis — the open question is multi-customer diversification before optics-at-1m become viable.
Optical DSP & SerDes Retimer Silicon for 800G/1.6T — mixed-signal silicon ↗ rising
Opportunity: 200G-per-lane PAM4 is the bottleneck for 1.6T pluggables; demand for SerDes/retimer IP is structural (fact 23be5595 Credo SerDes hiring; fact ba213d85 200G/lane DSP).
Quiet but critical layer under the 1.6T transition. Credo monetizes through merchant DSP, Celestica integrates at the module/system level. Risk: hyperscaler in-house SerDes (AWS, Google) could disintermediate.
NVIDIA Spectrum-X / InfiniBand AI Fabric — InfiniBand + Ethernet AI fabric (NIC + switch + protocol) ↗ rising
Opportunity: Vertically-integrated alternative to merchant Ethernet at gigascale; community is actively comparing MRC vs UE (demand_query 21f9b813).
Incumbent leader — biggest AI cluster deployments to date sit on this stack. Pressure comes from UEC-based Ethernet ecosystem; NVIDIA is countering with custom MRC transport.
Cisco Silicon One AI/ML Switching ASIC — switching ASIC → steady
Opportunity: Cisco's bet to compete head-on with Broadcom Tomahawk in the AI back-end; fact d4537830 confirms FY25 AI orders ~$2B (2× target) and fact 5bb3849e confirms ongoing silicon hiring.
Hot for Cisco specifically — one of the few non-merchant-silicon answers to Tomahawk 6. Risk: hyperscalers prefer multi-source merchant silicon, so Silicon One must keep winning on per-port economics.
What the market is asking (last 90d)
- AI interview with Fabric AI - what has been your experience?
- I have decided to go long on MicroLED’s and photonics. FABC, KOPN and POET are my picks
- FABC (Fabric.AI) is having some unusual price action today
- More and more lawyers are getting caught using AI without fact checking their sources
- Ask HN: How long before we get "coding agent in a box"?
- Ask HN: Why is there no "Response Contract" for CXL memory fabrics?
See the Products and Hiring modules for the full landscape and who's investing in which direction.