The modern AI factory faces a critical infrastructure divide. As models breach the trillion-parameter mark, the bottleneck is no longer the GPU itself, but the fabric connecting the cluster. For years, architects have been forced to choose between the raw, memory-semantic speed of proprietary interconnects and the ubiquity of Ethernet. The emergence of Ethernet for Scale-Up Networking (ESUN) signals the end of this binary choice, offering a standardized path that combines high performance with an open supply chain.
NVIDIA’s NVLink is the current benchmark for scale-up performance. NVLink delivers impressive performance in both bandwidth and latency which allows GPUs to function as a single logical memory domain. However, NVLink creates a walled garden; it requires proprietary hardware and
To counter this, the Ultra Accelerator Link (UALink) Consortium—backed by AMD, Intel, and hyperscalers—proposed an open alternative. UALink utilizes a protocol optimized for memory semantics (load/store operations) rather than Ethernet’s message passing, promising up to 1,024 accelerators in a pod. While UALink theoretically matches NVLink’s utility, it faces a critical hurdle: availability. Dedicated UALink switches are not expected until late 2026. Hyperscalers building clusters today cannot afford to wait for silicon that exists only on paper.
While UALink finalizes its specs, Broadcom has delivered a pragmatic solution: Scale-Up Ethernet (SUE). Implemented in the latest generations of Tomahawk hardware, SUE transforms traditional Ethernet into a lossless, low-latency fabric suitable for AI. By stripping down headers and implementing components of the Ultra Ethernet Consortium(UEC), like Link Layer Retry (LLR) and Credit-Based Flow Control (CBFC), SUE aims to achieve sub-250 nanosecond latencies and eliminates packet loss caused by buffer overflows.
Crucially, SUE is available now. Because UALink switches are delayed, companies like AMD are tunneling their protocols over Ethernet (UALoE) for their immediate rack-scale deployments.
The industry’s convergence point is ESUN (Ethernet for Scale-Up Networking), an Open Compute Project (OCP) workstream supported by AMD, Meta, Broadcom, Arista, and others. ESUN takes the innovations of SUE—lossless transport and optimized framing—and standardizes them. It ensures that Layer 2 and 3 behaviors are consistent across different vendors, turning the scale-up network from a black box into an interoperable utility.
ESUN will likely become the standard because it leverages the massive economies of scale of the Ethernet ecosystem while accommodating the performance needs of AI. It allows operators to build "AI Factories" using best-of-breed components rather than single-vendor verticals.
While NVLink offers peak performance today and UALink promises open memory semantics, ESUN delivers the winning combination of performance, openness, and immediate availability. By standardizing the physical and transport layers, ESUN transforms the scale-up domain from a fragmented battleground into a unified ecosystem. If history can be used as an indicator, the sun will continue to rise on Ethernet.