Ultra Ethernet 1.0 turns commodity Ethernet into an AI back-end fabric
A 560-page spec gives Ethernet InfiniBand-class RDMA and congestion control; the 2026 work item is Programmable Congestion Management
Summary
The Ultra Ethernet Consortium (UEC) shipped Specification 1.0, a 560-page, vertically integrated stack across NICs, switches, optics and cables, giving commodity Ethernet the features that made InfiniBand the AI back-end default: native RDMA, advanced flow control, path diversity, congestion awareness and ultra-low tail latency, while keeping standard Ethernet optics and tooling. It is the scale-out complement to UALink's scale-up fabric, and the cost-led alternative to Nvidia's InfiniBand for million-GPU clusters. The headline 2026 work item is Programmable Congestion Management, code a congestion-control algorithm once in a standard language and run it on any UE-compliant NIC, decoupling the fabric's most contested layer from any single vendor.
By the numbers
- 560+, pages in the UEC 1.0 specification.
- 2026, first PCM (Programmable Congestion Management) work item.
- 5, stack layers covered (NICs, switches, optics, cables, transport).
- 1.0, current spec generation, released ahead of 2026 refinements.
Why it matters
InfiniBand's lock on AI back-end networks is Nvidia's quietest moat. An open Ethernet stack with RDMA-class performance lets hyperscalers wire million-GPU clusters on multi-vendor gear at commodity-optics prices, shifting where the networking margin sits and giving Broadcom, Cisco and Arista a credible AI-fabric pitch.
What to watch
- Programmable Congestion Management landing in shipping NICs.
- UEC-1.0 deployments in named hyperscale clusters vs continued InfiniBand.
- How UEC scale-out and UALink scale-up divide the fabric stack in practice.