{
  "id": 7673624,
  "title": "AI networking startups race to replace Nvidia's NVLink",
  "url": "https://urgent.news/2026/09/15/ai-networking-startups-race-to-replace-nvidias-nvlink-7673624",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-15T20:56:48.000Z",
  "source": {
    "name": "The Register Science",
    "slug": "the-register-science",
    "url": "https://www.theregister.com/systems/2026/09/15/ai-networking-startups-race-to-replace-nvidias-nvlink/5296672"
  },
  "original_language": "en",
  "account": "As Nvidia's influence expands with its NVLink Fusion technology, rival networking vendors are racing to introduce alternative interconnects and switches to the market. At the AI Infra Summit, Delos Data and Cornelis Networks officially entered the scale-up networking race. Scale-up fabrics, like Nvidia's NVLink, allow multiple GPUs to behave as one massive AI accelerator. To compete, rivals such as AMD have adopted emerging protocols like Ultra Accelerator Link. Currently, these protocols are largely run over standard Ethernet switches. For example, AMD uses Broadcom's 102.4 Tbps Tomahawk 6-based switches connected to custom I/O dies on the MI455X. While purpose-built UALink switches and physical interconnects are still lacking, Cornelis and Delos aim to fill that gap soon. Cornelis, a HPC-centric networking vendor spun out of Intel in 2020, introduced the Active Compute Fabric (ACF) at AI Infra. This standard seeks to establish an open architecture for scale-up and scale-out networking, integrating programmable compute into the fabric. Cornelis aims to bring its tech to a broader audience through the open Ultra Ethernet protocol and scale-up networking. The ACF expands on technologies like Ultra Accelerator Link and Ethernet for Scale-Up Networking, setting a baseline for in-network compute capabilities across various hardware, not just Cornelis' CN-series parts. Key features include KV cache offloading, Mixture-of-Experts expert dispatch, Message Passing Interface offload, and in-fabric checkpointing for failure recovery during training and large workloads. These in-network accelerators could significantly reduce idle GPU time, potentially saving substantial costs. Delos Data, founded by former Barefoot Networks and Intel executives, aims to tackle the physical layer with network reference designs under its Nonstop AI portfolio. Delos offers data interfaces in three form factors, including an I/O die capable of over 30 Tbps of aggregate bandwidth or 8 TB/s in either direction. This is more than double the interconnect bandwidth of Nvidia's or AMD's latest accelerators, which cap out at 3.6 TB/s. Delos' chiplets are protocol agnostic, allowing customers to use ESUN, UALink, or other protocols. This approach enables chip designers to focus on AI aspects of their accelerators while licensing Delos' chiplet design.",
  "summary": "Intel spin-off Cornelis and newcomer Delos Data pitch open alternatives for scaling AI beyond the rack",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "The Register",
        "title": "AI networking startups race to replace Nvidia's NVLink",
        "url": "https://urgent.news/2026/09/15/ai-networking-startups-race-to-replace-nvidias-nvlink",
        "published": "2026-09-15T20:56:48.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}