Networking startups Cornelis Networks and Delos Data have unveiled rival technologies aimed at breaking Nvidia's dominance in "scale-up" interconnects, the fabric that lets clusters of GPUs act as a single giant accelerator. Announced at the AI Infra Summit, the moves reflect a wider industry push, alongside efforts from AMD and others backing protocols such as Ultra Accelerator Link, to build open alternatives to Nvidia's proprietary NVLink and NVLink Fusion technology as demand for large-scale AI computing infrastructure grows.

Cornelis, an Intel spin-off best known for its Omni-Path interconnects, introduced an open architecture called Active Compute Fabric (ACF), which builds programmable compute directly into the network to offload tasks such as KV cache management, mixture-of-experts data dispatch and failure checkpointing, functions similar to Nvidia's SHARP technology on its NVSwitch chips. The company claims that in a 100,000-GPU system, roughly half of all GPU hours are currently wasted waiting for data, costing around $1.68 billion a year and 500 GWh of power, though such vendor estimates should be treated with some caution. Cornelis also plans to launch 800 Gbps CN6000-series switches and NICs, while Delos Data, founded by veterans of Barefoot Networks, is separately pursuing its own approach to the same problem.