Cornelis Networks Launches Active Compute Fabric with $205M to Unclog Scale-Up AI Networking

At the AI Infra Summit, AI infrastructure startup Cornelis Networks introduced its Active Compute Fabric, a specialized open network architecture engineered to alleviate cluster bottlenecks across both scale-up and scale-out AI environments. Alongside the architecture launch, the company announced a strategic collaboration with Qualcomm Technologies focused on rack-scale inference and confirmed a $205 million funding round led by IAG Capital Partners to scale production of its CN5000 and CN6000 network switches. For cloud network engineers and infrastructure architects, network fabric performance has become the decisive factor in AI cluster utilization. While GPU and TPU accelerators have scaled rapidly in compute density, networking layers often leave expensive silicon idle while waiting on distributed data synchronization and collective operations. Cornelis's Active Compute Fabric embeds programmable compute directly into the network transport layer, enabling in-flight data processing, lossless transport, and hardware-accelerated offloading of collective operations as packets traverse switches. This development reflects a major architectural transition across modern cloud and high-performance networking: the convergence of compute and switching. As AI model topologies outgrow single-server chassis, scale-up interconnects and scale-out spine-leaf fabrics must handle distributed inference and training workloads with near-zero jitter and strict congestion control. The push toward open, programmable network fabrics also signals growing enterprise demand for vendor-neutral alternatives to proprietary interconnect ecosystems like NVIDIA InfiniBand and NVLink. In practice, infrastructure teams designing on-premises AI data centers or hybrid clusters should evaluate how in-network compute architectures reduce communication tail latencies compared to standard RoCEv2 (RDMA over Converged Ethernet) implementations. While adopting programmable fabrics like Cornelis's requires evaluating compatibility with existing orchestration and accelerator stacks, the potential to significantly boost effective GPU utilization makes in-fabric processing a vital architectural paradigm to watch as cluster sizes expand.