What is AI back-end networking? It is the network that moves data between the accelerators inside an AI cluster — the GPU-to-GPU traffic that never leaves the cluster — across the three networks of an AI data center. It excludes the front-end network that serves clients, storage and management. The back-end fabric is where AI networking spending concentrates: we published in June 2024 that data center AI networking will surge to nearly $20B in 2025, and as of October 2026 we see the market exceeding $1T annually in the early part of the 2030s. Knowing what sits inside the back end — and what sits outside it — is the difference between sizing this market and double-counting it. We walk the full curve in The $1 Trillion Dollar Networking Market (the early-2030s curve).
How Is Back-End Different From Front-End?
Scale-up plus scale-out fabrics are the back-end core, and the exact split depends on which traffic you count. In 2025, we forecast scale-out networking to exceed $8B with front-end networking above $5B, so the fabric connecting accelerators already outweighs the network serving users around AI sites. Add scale-up interconnect and the back-end share grows further. Buyers who fold the front-end into the back-end double-count shared infrastructure; buyers who exclude scale-up miss the fastest-ramping fabric in the cluster.
Two fabrics make up the back end. The scale-up fabric connects accelerators within a rack or domain, where NVLink, Ethernet scale-up variants, PCIe and UALink carry the traffic. The scale-out fabric connects racks into one GPU cluster fabric, where InfiniBand and Ethernet running RDMA move training data between nodes. Front-end traffic rides a separate physical network, which is why we report the two markets separately, quarter by quarter.
How Fast Is Back-End Networking Spending Growing?
The ramp is steep and public. The AI networking market reached $2B in 2022, with InfiniBand carrying 75% of the revenue, grew to over $3B in 2023, and reached over $15B in 2024 on its way to over $25B by 2028. As of April 2025, we reported Ethernet AI networking revenue surged over 150% Y/Y in 2024 with bandwidth up over 200% — the moment Ethernet took the back-end lead. RDMA switching has grown from a $1B market to one that will exceed $18B in 2028, with RDMA networking overall exceeding $22B by 2028.
Looking out the decade, the fabric layers separate cleanly. As of January 2026, we forecast AI scale-out Ethernet revenue to surge to over $100B by 2030, with scale-up switching exceeding $30B by 2030, optics adding $10B+, NVLink at $25B+, Ethernet scale-up at $8B+ and PCIe/UALink at $3B+. After rack-scale systems reset the architecture, we steepened the scale-up view to passing $100B by 2030 as of October 2026 — a revision made in public, on the same published curve that carries AI networking exceeding $1T annually in the early part of the 2030s.
How Was This Calculated?
The figures in this Explainer come from 650 Group’s published quarterly research on AI networking, data center switching, and interconnect markets, which track vendor revenue and port shipments by segment and speed each quarter. Each figure above links to the publication it came from, and the full catalog is on our research reports page.
650 Group researches several of the technologies we covered in this Explainer, including Data Center, Co-Location & Cloud and Enterprise, and more.
What to read next: the market sizing in How Big Is the AI Networking Market?, the speed ladder in What Are 400G, 800G and 1.6T Ports?, and the architecture split in A Tale of Three Networks: AI Networking Infrastructure.