Startups Challenge Nvidia's NVLink Dominance with Open Networking

As Nvidia expands its ecosystem via NVLink Fusion, new competitors are attempting to carve out a market for open-standard scale-up networking, as first reported by The Register. Reporting highlights two primary challengers debuting technology at the AI Infra Summit.

Cornelis Networks, an Intel spin-off, introduced the Active Compute Fabric (ACF). This open architecture integrates programmable compute into the fabric to handle functions such as MoE expert dispatch and KV cache offloading. Cornelis claims that in a 100,000 GPU system, roughly half of all GPU hours are spent waiting for data, which the company estimates represents 500 GWh of power and $1.68 billion in wasted annual capacity. To support this, Cornelis is launching CN6000-series NICs and switches capable of 800 Gbps.

Meanwhile, Delos Data—founded by former Intel and Barefoot Networks executives—is targeting the physical layer. Through its Nonstop AI portfolio, Delos is offering network reference designs, including an I/O die that claims aggregate bandwidth exceeding 30 Tbps (approximately 8 TB/s in either direction). The Register notes this is more than double the 3.6 TB/s interconnect bandwidth found in current accelerators from AMD and Nvidia.

While AMD has adopted the Ultra Accelerator Link (UALink) protocol—currently tunneled over Broadcom’s 102.4 Tbps Tomahawk 6-based switches for its MI455X—purpose-built UALink hardware remains elusive. Delos aims to avoid the "walled garden" approach of Nvidia's NVLink Fusion, which requires the use of NVSwitches for scale-up networking.