AI-ready Data Centre Networking: Why the network solution of the future for your data centre will differ fundamentally from your current architecture

1. Limitations of traditional data centre architectures

Many data centre networks currently in production have evolved over time. Typical characteristics include:

• Manual configuration of switches and routers
• Device-centred operation rather than a service- or application-oriented approach
• Limited visibility into traffic flows and performance
• Static design assumptions based on predictable load profiles

These approaches quickly reach their limits in AI environments. Training clusters with GPUs or specialised accelerators generate massive, synchronous data streams. Even the slightest packet loss or microbursts can significantly prolong training times or cause jobs to abort completely. At the same time, the complexity of operations increases exponentially when networks need to be manually adjusted or expanded.

2. New requirements arising from AI and data-intensive workloads

AI workloads are fundamentally changing behaviour within the data centre:

• Significantly higher east-west traffic between compute nodes
• High bandwidth requirements (100G, 400G and beyond)
• Ultra-low latency and deterministic behaviour
• Lossless or near-lossless transport
• Dynamic scaling of clusters within a very short time

A network must therefore not only be fast, but also predictable, adaptive and self-monitoring. This is precisely where an AI-ready architecture differs fundamentally from traditional designs.

3. Architectural principles of modern AI-ready data centre fabrics (Juniper focus)

In contrast to traditional, often proprietary designs, Juniper Networks’ modern approach is based on clearly defined, validated fabric architectures. These include, in particular:

• Spine-leaf fabrics (Layer 3 Clos design) for horizontal scaling
• IP fabric with EVPN-VXLAN as an overlay for maximum flexibility
• Collapsed spine designs for smaller deployments
• Dedicated AI fabrics for GPU clusters with optimised east-west traffic

These designs enable deterministic performance and can be scaled linearly – a crucial factor for AI workloads.

3.1 Intent-based Networking with Juniper Apstra

A key element is the transition from device-based configuration to intent-based networking. Apstra acts as the central, multi-vendor SDN management platform within the data centre networking environment.

With Juniper Apstra, the desired state (‘intent’) is defined, whilst the system automatically:

• generates and rolls out configurations
• Validates dependencies between fabric components
• Performs continuous validation (closed-loop)

This ensures that the network always corresponds exactly to the desired design – even after changes have been made.

3.2 EVPN-VXLAN fabrics as a scalable foundation

Modern data centre networks are based on spine-leaf architectures using EVPN-VXLAN as the underlay/overlay technology. Juniper extends this approach through deeply integrated automation and validation via Apstra, as well as optimised control-plane mechanisms. Apstra thus abstracts the complexity of the protocols used to such an extent that policies can be rolled out end-to-end in a service- and application-oriented manner.

Advantages:

• Virtually unlimited scalability
• Layer 2 and Layer 3 services across the entire fabric
• Clear separation of physical and logical topology
• Multi-tenant and multi-cluster capability (e.g. for applications, services, departments, business units, customers, …)

3.3 Telemetry and AI-powered assurance with Juniper Mist AI

Instead of periodic SNMP polling, Juniper relies on real-time streaming telemetry.

In combination with Juniper Mist AI and Datacentre Assurance, this results in:

• End-to-end visibility across the fabric, applications and user behaviour
• Automatic root cause analysis (RCA)
• Proactive detection of performance degradation

3.4 Optimisation mechanisms for AI fabrics (RLB, DLB, ECMP enhancements)

Special mechanisms for fabric optimisation are used for high-performance AI clusters:

• RLB (Random/Adaptive Load Balancing) for better distribution of flows
• DLB (Dynamic Load Balancing) to prevent hotspots
• Optimised ECMP hashing for even path utilisation
• Congestion-avoidance mechanisms for low-loss transport

These technologies are crucial for minimising incast, microbursts and head-of-line blocking in GPU clusters.

4. Automation and Lifecycle Management

Another distinguishing feature compared with traditional architectures is the fully automated lifecycle:

• Day 0: Design validation and simulation prior to commissioning
• Day 1: Zero-touch provisioning and automated fabric deployment
• Day 2: Continuous validation, updates and enhancements without any disruption to operations

These functions are centrally managed by Juniper Apstra and enable consistent operation throughout the entire lifecycle. A range of vendors can be integrated into Apstra. This also allows for the implementation of brownfield, migration and multi-vendor scenarios.

5. Platforms and hardware: Juniper & HPE integration

High-performance switching platforms are an essential component of modern AI-ready architectures. Juniper relies in particular on the QFX series, including, for example, the Juniper QFX5250.

Where required, these systems offer:

• 400G/800G interfaces for AI clusters
• Ultra-low latency switching
• High buffer capacities for bursty traffic

The collaboration between Hewlett Packard Enterprise (HPE) and Juniper has resulted in integrated solutions in which compute, storage and networking are closely interlinked – particularly for AI and HPC infrastructures. Traditional data centre network environments can also be implemented using these solutions (e.g. 1/10/25/40/50/100GE).

6. Comparison: Today vs. the AI-ready future

The following diagram provides an overview of the key differences between traditional and AI-ready data centre networking environments: