What is a “technology advance?” To the media, Wall Street, and most vendors, it’s anything that’s new. Hey, this is the community that gave us the Great 5G Fable, after all. But “advance” usually means moving forward, which implies progress against a credible goal. There’s not a lot of that going on, and much of what claims to be is just click-bait. So when Juniper, who’s not impressed me much with their positioning and announcements, said they were launching a new fusion between AI and networking, a two-dimensional one in fact, I think I can be forgiven for skepticism. I was wrong, and the “why” is the most important thing to investigate here.
What Juniper announced was “The First AI-Native Networking Platform”. Not only the first, but the only one, so they said in their analyst briefing, and the two dimensions are “The AI-Native Platform” and “The AI Data Center”. Nobody, especially me, would believe that there’s not an element of opportunistic AI-washing going on here, but what’s unusual and surprising is that there is a pretty strong foundation of truth here too, and truth that might well play into the HPE acquisition of Juniper.
The AI-Native platform is a realization of Juniper’s Experience-First Network concept. The idea is that networks are getting a lot more complicated every day, companies are relying on them more every day, and operations is getting more and more behind the curve in assuring the fundamental truth that it’s the experience that networks deliver that matters. How can you do that when you don’t understand what’s being delivered and how it relates to that experience? That’s a question that Juniper believes AI is uniquely capable of answering.
The first step in validating that claim is to broaden the scope of what Juniper’s AI can see, and use 7 years of multi-domain network data to build an AIOps model. Juniper has a cloud-and-microservice-based analytics service that can now ingest more data than ever, including data from Apstra, their virtual-data-center technology. Routers, switches, firewalls, the Apstra data center, and Juniper applications can all feed the data pool. Juniper’s Mist assurance products address a full range of equipment, and the Marvis virtual network assistant, AI-based, allows natural language queries into network conditions and also into the knowledge base of technical documentation. Full-scope AIOps, and because Apstra is a virtual-data-center technology, it can handle the variations in vendors you’ll almost surely find in every data center.
The second step is, I think, the most interesting piece. Marvis Minis is a software component that can be deployed on Juniper access points to generate what appears to be user traffic. In effect, it creates a kind of digital-twin at the experience level rather than at the actual user level, and Juniper calls this a “digital experience twin”. Enterprises today tell me that a great majority of their problem remediation activity is stimulated by user complaints, which is hardly ideal. The Marvis Minis let enterprises invoke simulated users to uncover problems.
Simulation injection right at the access edge is interesting, of course, and valuable. What I wonder is whether “digital experience twin” is just a handy term, or whether there’s an element of digital twinning involved. Could the Marvis Minis draw on real-world conditions in some way, as an actual digital twin might? Could it even be stimulated by real users’ activity? Or even represent real users too? Juniper didn’t have an answer to that, but it sure would move the ball in a lot of ways, one of which we’ll talk about below.
The second leg of the new Juniper announcement is the AI Data Center piece, and there are a number of interesting pieces to this. At the top level, it’s about validating Ethernet as the fabric for AI connectivity within the data center. The keystone of this part is Juniper’s new PTX model, which is an 800G switch that’s the first use of Broadcom’s Tomahawk 5 chip.
It’s not surprising that Juniper is pushing Ethernet versus the Nvidia-native InfiniBand strategy given that Juniper doesn’t make InfiniBand switches, but if you think about it, the move strongly suggests that Juniper believes there will be a boom in self-hosted AI among enterprises. In fact, they say that their modeling shows that private data centers would be cheaper than public AI for “major use cases”. The simple fact that they’ve modeled it, and that they say Ethernet as a 50% TCO advantage over InfiniBand, says they’re taking this very seriously. We’ll get to that point too.
All of this, including the AI-native positioning, is pretty good. It might even be great, but making it so would require a few tweaks and alignment with a potential future trend that Juniper alone has little chance of driving. However, HPE could drive it, and of course HPE and Juniper could (and probably will) become one in the near future.
The biggest challenge the AI strategy faces is the multi-vendor nature of IT and network infrastructure. There is no question that a complete Juniper solution would enable “AI-native” to radically change operations. In fact, having a few pieces of maverick gear scattered about wouldn’t likely be a major problem, as long as that gear was confined to areas near the network edge, and in particular to non-critical sites. As more non-Juniper gear intrudes, though, it would tend to break down AI continuity. I think Juniper is already working to enhance their multi-vendor network features, and to expand Apstra’s ability to use data center virtualization across a wider spectrum of equipment.
I also think that advancing the AI-native concept into security, and to expand the use of digital-twin technology (Marvis Minis) to both real and virtual users would be a strong positive. The broader the set of benefits, the greater the overall operations impact. And, of course, the easier it would be to justify lifting out some older non-Juniper equipment to improve things even more. HPE could have a positive impact here, but it could also have a negative one in that M&A always distracts everyone, particularly for fear of layoffs.
The big question remains; is this enough to stem the slipping investment in networking by enterprises. While operations efficiency is a valid issue, I can’t find any indication that it’s powerful enough to do more than slow down commoditization. The only certain pathway out of the current broad-based business tech commoditization is a set of new applications that open new business cases. As I’ve pointed out in past blogs, I believe that the real-time-linked point-of-activity empowerment applications are the only credible candidates but there’s always the chance of something new coming along. A new application like real-time digital twinning could be what at least some HPE planners have in mind for the deal, and it may even be that Juniper’s characterization of Marvis Minis as experience digital twins is a nod to a future move. We’ll see.
