HPE reinforced the telco shift at MWC: the urgent mid-gen upgrade of fiber interconnect and longhaul networks for the AI era, plus the edge inference agenda just coming on-stream – as served by its PTX and MX portfolios, respectively, as supported by its modular hardware, custom ASICs, and AI management software.
Optimized AI infrastructure – new PTX series routers, powered by custom ASIC, offer up to 500 Tbps switching capacity and low packet loss for DCI workloads.
Edge and enterprise reach – latest MX series routers, including compact MX301, provide high-throughput routing for enterprise WANs and metro-edge AI inference.
Agentic-AI management – integrated Aruba Central and Mist platforms enable AI-driven automation, maintenance, and scaling across cloud and edge deployments.
Another replay of another MWC discussion; this time with HPE, which also confirms this RCR narrative that, even at MWC, mobile matters less, right now. “Correct. It’s not a mobile access story; the mobile access piece comes later,” says AE Natarajan, senior vice president and general manager for routing infrastructure solutions in HPE’s networking division. For now, the network piece in the AI infrastructure jigsaw connects the back-end, going east-west between data centers, and basically everywhere north of the RAN edge. His firm has just launched a new line of (Juniper-branded) PTX series routers, built on a fifth-generation version of its own application-specific integrated circuit (ASIC). It is a humdinger, by all accounts – an “engineering marvel”, no less.
Natarajan outlines a platform built for scale at the extreme end of current AI traffic flows – in data center interconnect (DCI) systems between GPU clusters in the cloud. Its custom chip (Juniper Express 5 ASIC), tuned for the bursty AI traffic flows, delivers a 49 percent jump in power efficiency, apparently – versus its forebear. It also supports more sophisticated traffic management. “It means fewer packets are dropped when you’re multiplexing multiples of these (GPU streams) in 400G or 800G pipes”, he says. Which translates in its PTX12000 line as a better GPU utilisation – fewer drops and less idle time for expensive GPUs. “We are at 0.3 percent – which means one packet in a 300 packet burst in GPU-to-GPU comms – whereas competitor products drop about 60 packets,” he says
These are modular routers, he says, built to scale network capacity for future workloads without repeated redesigns and upgrades. “You can do 800G today and migrate to 1.6T in the same chassis in the future,” he says; the set-up also supports 100G and 400G. It is designed for high switching capacity in a dense footprint, scaling from 345.6 Tbps in the eight-slot version (PTX12008) to over 500 Tbps in the 12-slot configuration (PTX12012) – to 1,854 800G interfaces at max density, with dynamic bandwidth allocation to boot. HPE has also introduced a compact fixed-form two-rack (2RU) router, PTX10002, for up to 28.8 Tbps across f multi-rate interfaces (100G, 400G, 800G), also promising better routing performance across large AI fabrics and distributed deployments.
A press note talks about “efficient scaling of AI clusters and WAN” – so the portfolio is good for longhaul systems as well as DCI. Its software layer has evolved, too. Its network management and automation platform, Juniper Routing Director, is now “agentic-AI ready”, the company says, allowing enterprise AI co-pilots to automate operations and troubleshoot WAN routing issues dynamically. It has moved “quickly” to integrate its new Juniper portfolio, it says, following its $14 billion purchase of the firm in July last year (2025). It has since combined “the best of” their management platforms (Aruba Central and Juniper Mist), it says, with a common agentic AI and microservices framework. Mist gets global NOC insights and Aruba gets a ‘large experience model’, as provided by each other.
The latter uses “billions” of data points from standard IT apps, plus “synthetic” digital-twin data to resolve issues. New WiFi-7 access points work across both platforms, as well. Indeed, HPE has an eye on the edge, even if its PTX angle is mostly for the cloud. Natarajan offers a “10,000-feet view” of the new AI migration. “Inferencing is coming up big time, and disrupting everything. Regular telcos, like at MWC, have to re-architect their networks to participate. Traffic patterns will be different – not just asymmetric, one way, but symmetric where rich data goes both ways. Plus agentic AI never sleeps so your load-balancing and multiplexing algorithms don’t work, and [data] doesn’t render itself for caching because it is quickly obsolete. So the network has to change – everywhere, including at the edge.
That edge story is where Juniper’s long‑established MX series universal routing platforms come in, going from compact units for edge inference and enterprise WAN aggregation to larger modular platforms for metro and edge routing. Natarajan points to the MX301 unit, as the “most power efficient and most compact”, with built-in security, both MACsec and IPsec. He says: “If you’re using ChatGPT, there is a 100 percent chance it goes through a Juniper MX router, either in Amazon’s Direct Connect or Microsoft’s Expert Cloud, or Oracle Customer Connect, or any of those – in hyperscaler clouds, regional clouds. Wherever you go, you most likely hit one of those. So we cover both the DCI between GPUs, as well as all the edge stuff that’s coming in.”
How far along are mobile operators, generally? What’s the discussion like at MWC? Some and some, the answer goes. “Some have actually gone the bold step, and it is unbelievable,” he says, citing an unnamed customer with a new AI cluster, already maxed-out and ready-to-expand. “They want to build their second and third,” he says. But it remains a major cap-ex exercise, while they struggled to monetize their existing 5G deployments, and they size-up mid-term 6G outlays. Of course, HPE is serving other sectors besides, all with major operatives on similar AI courses. “Big factories, industrial complexes, with private networks” – its PTX and MX systems will go into all of these venues, says Natarajan. “Yes, absolutely. They are becoming cloud‑scale enterprises, with their own requirements.”
He goes on: “It could be a big car manufacturer or a large financial institution – they have their own models, agents, data protection. They fit together logically in several use cases – manufacturing floors, wherever you need edge inferencing transformation in the network. Because you’re using AI or building out agentic AI, you need to transform the network.” In the end, the pitch is about scale and flexibility. HPE is not just shipping router boxes, but modular systems based on bespoke tech to connect and manage AI workloads end-to-end – from the hyperscale GPU clusters in the cloud to the telco and enterprise edge where inferencing, low-latency analytics, and AI agents are coming online. “What sets us apart is our control of innovation,” he says. “Because we build our own chips.”
He comments: “That gives us leverage. Could others build it? Absolutely. But most of the value – lowest power, most compact, load-balancing, inline security – I haven’t seen others do that yet. It comes from the chip – from the IP, from our silicon.”
