According to the data centre market research firm, agentic AI and inference are driving demand for front-end switch sales in data centre networks differing from those associated with training workloads.
According to the data centre market research firm, Dell’Oro, agentic AI and large-scale inference are beginning to push AI-driven networking demand beyond specialized accelerator fabrics and into the front-end networks connecting CPUs, storage, services and general-purpose infrastructure.
In a newly published report it says that more than half of growth in sales of datacentre front-end switches will come from AI-related opportunities.
The architectural shift reflects the different infrastructure requirements of inference compared with large-scale model training. Dell’Oro says that the traditional assumption of a XPU to CPU ratio of 10:1, no longer applies to all emerging deployments, with some environments transitioning towards a 1:1 ratio.
XPU is an umbrella term for all processing units like CPUs, GPUs, TPUs or NPUs. CPUs became increasingly important for orchestration of data movements in inference networks. Additionally, networking for KV caching storage racks is also needed for inferencing applications.
Dell’Oro identifies Accton, Arista, Celestica, Cisco, HPE/Juniper, H3C, Huawei and NVIDIA among the vendors positioned for the opportunity.
