Two startups announced competing approaches to scale-up networking this week as rivals seek open alternatives to Nvidia’s NVLink Fusion. At the AI Infra Summit, Intel spin-off Cornelis Networks and newcomer Delos Data presented architectures and hardware aimed at scaling AI beyond a single rack and reducing reliance on proprietary NVSwitch-based fabrics.
Scale-up networking: Cornelis’ Active Compute Fabric
Cornelis, spun out of Intel in 2020 and built on the company’s Omni-Path technology, introduced the Active Compute Fabric (ACF), an open architecture intended to standardize in-network compute across scale-up and scale-out deployments. The company said ACF integrates programmable compute into the fabric and expands on efforts such as Ultra Accelerator Link (UALink) and the Ethernet for Scale-Up Networking protocol to set a baseline for in-network capabilities across different vendors’ hardware.
Cornelis is preparing 800 Gbps-capable CN6000-series switches and NICs and plans to support the open Ultra Ethernet protocol. The company highlighted several in-network acceleration use cases: KV cache offloading to store and retrieve model state across sessions; MoE expert dispatch to cut duplicate transfers for mixture-of-experts models; message passing interface offload for HPC workloads; and in-fabric checkpointing for failure recovery. Cornelis said these in-network accelerators “allow the network to operate on data as it moves through the system.”
Cornelis also presented modeled savings potential: “In a 100,000 GPU system, Cornelis modeling of public data shows that roughly half of all GPU hours are spent waiting for data, worth about $1.68 billion a year in wasted capacity and 500 GWh of power.” The company added that “Communications overhead and synchronization can leave expensive accelerators underutilized.” As usual, take these claims with a grain of salt.
Delos Data’s protocol-agnostic I/O and optics
Delos Data, founded by former Barefoot Networks and Intel executives, unveiled its Nonstop AI portfolio of network reference designs that span co-packaged interconnects, near-package optics and more traditional NICs. The company said its I/O die offers more than 30 Tbps of aggregate bandwidth — about 8 TB/s in either direction — more than double the interconnect bandwidth of Nvidia’s and AMD’s latest accelerators, which the source cites as capping out at 3.6 TB/s. Delos noted that actual competitiveness will depend on when these I/O chiplets are deployed, since they must be integrated into accelerator packages and require significant co-design.
Delos emphasized protocol agnosticism: its chiplets do not require a specific protocol and can work with ESUN, UALink or others, and customers can license Delos’ chiplet designs rather than adopt a closed ecosystem. Its near-packaged optics (NPO) technology integrates a 10-plus Tbps data interface — described as roughly 2.5 TB/s bidirectional in Nvidia terms — with optics engines from leading suppliers. Delos proposed using copper inside racks and user-serviceable NPO modules between racks as systems scale from rack-level (72 GPUs) to row-scale and beyond, where optics are likely unavoidable. The company is also developing a 400-plus Gbps NIC and said its Nonstop AI software and telemetry enable configuration, monitoring and dynamic rerouting of switched fabrics or meshes to speed recovery from link failures.
Market context and funding
The announcements come as vendors and hyperscalers seek alternatives to NVLink Fusion and its NVSwitch-based scale-up approach. Today, emerging scale-up protocols such as UALink are often tunneled over standard Ethernet switches; for example, AMD is using Broadcom’s 102.4 Tbps Tomahawk 6-based switches to connect to custom I/O dies on the MI455X. Collective acceleration features similar to Nvidia’s SHARP have been implemented to varying degrees by major networking vendors.
Investors appear to be backing competition: Cornelis announced approximately $205 million in funding to bring its next-generation scale-up and scale-out networking products to market, while Delos Data said it has raised more than $100 million with support from Matrix, Playground and Socratic Partners, among others.
