AI-Native Networking For AI Factories

Author: Ahmed, Juniper Networks (now part of HPE) Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Juniper Networks, now part of HPE, presents its AI-native networking strategy designed from the ground up to address the unique demands of AI workloads -- massive elephant flows, extreme sensitivity to packet loss (which causes large tail latencies and GPU performance degradation), and order-of-magnitude higher bandwidth requirements versus traditional workloads. The solution centers on three product families: QFX switches, PTX routers (both using merchant and custom silicon for supply chain diversity), and SRX 4700 firewalls (industry-leading performance per RU). Juniper was among the first companies to submit MLPerf results to ML Commons, validating their AI networking with real customer workloads in their HPE-Juniper AI Lab using gear from AMD, NVIDIA, VAST, WEKA, and Broadcom.

The operational layer leverages two key acquisitions: Apstra (intent-based networking for rapid design and deployment with end-to-end visibility from fabric to NIC and autotuning) and Mist (Marvis AI platform with natural language assistant, agentic AI for troubleshooting and resolution, and predictive analytics for optics failure prediction). Networking is now recognized as critical not just for training but also for inference, driven by reasoning models, agentic AI, and mixture of experts architectures. Juniper secured the Gartner leadership position in data center switching and ranks #2 in the AI fabric use case (behind NVIDIA). Customers include DigitalOcean and Vultr, both deploying AMD and NVIDIA GPUs over Juniper networks.