The network is becoming critical to GPU utilisation, resilience and scale.
Whilst the AI data center conversation has traditionally centred on GPUs, servers and power, as clusters grow, networking is becoming an equally important part of the infrastructure equation and HPE’s / Junipers expanded collaboration with Oracle to deploy HPE Juniper Networking across Oracle’s global AI data centers proves this. The planned multi-year deployment includes Juniper PTX and MX routing platforms and QFX and EX switches – For AI data center operators, the significance of this goes beyond the size of the customer and shows HPE are pushing networking to the centre of it’s AI infrastructure strategy.
The network is now part of AI performance
Large GPU clusters generate enormous volumes of east-west traffic between compute nodes. HPE says Oracle’s AI infrastructure is placing increasing demands on bandwidth, latency, congestion management and fault recovery. The latest QFX switches are intended to support high-density connectivity, dynamic load balancing and congestion management for RoCEv2 AI backend networks. This matters because a GPU waiting for data is expensive capacity sitting idle. For operators, networking performance is becoming directly linked to the economics of AI infrastructure – the objective is not simply to connect GPUs; it is to keep those GPUs supplied with data and minimise the network conditions that can interrupt or slow distributed workloads.
Visibility is becoming as important as bandwidth
HPE and Oracle are also working on intelligent telemetry designed to provide deeper visibility into network and fabric behaviour – the companies say this can help identify queue build-up, packet loss, traffic imbalances and component degradation earlier, potentially reducing their impact on GPU utilisation and network availability, which points towards a wider change in AI data centre operations. As infrastructure becomes larger and more distributed, manually identifying network problems becomes increasingly difficult, operational intelligence that can identify bottlenecks and potential failures before they affect workloads could become an important part of maintaining predictable AI performance.
HPE is integrating Juniper deeper into its AI stack
The Oracle announcement follows HPE’s June launch of expanded AI networking capabilities – At HPE Discover, the company announced that QFX switches would be integrated into its HPE AI Data Center Solution, alongside compute, storage, software and services. It also introduced the QFX5140 for inference clusters and edge AI, and a QFX5252 switch tray designed for AMD’s Helios AI rack-scale platform. This strategy is moving beyond selling networking as a separate layer as the company attempts to make networking part of an integrated AI infrastructure stack — while adding AI-driven operational capabilities around it.
The timing is also notable – HPE reported that it’s Data Center Networking revenue reached $382 million in its fiscal Q3 2026, up 112.2% year-on-year, while overall Networking revenue increased 74.9%. These are HPE’s reported results, rather than an independent measure of the wider market.
Why this matters to data center executives
For AI data center executives the broader lesson is straightforward; as clusters scale from individual racks to large AI factories, network performance, congestion management, telemetry and fault recovery become infrastructure-level concerns rather than networking-team concerns. The HPE-Juniper strategy reflects that shift: AI data centres do not simply need more compute – they need a network capable of keeping that compute working efficiently, predictably and at scale.
