Two decades ago, startups in the networking space could increase their chance of success by specializing, by dodging Cisco’s breadth and instead looking for a limited and Cisco-underserved area. Today, Cisco says it’s looking to sustain its success by essentially dodging its own broad incumbency. Niches were good in the past, but can niches add up to a broad market? Can they sustain a giant like Cisco, or even smaller security-specialized firms like Palo Alto?
Breadth commoditizes. Nothing can command an enormous TAM without losing complexity, differentiation, in favor of lower prices to broaden appeal. So it is now with networking. Cisco seems to believe that there are remaining spaces at the edge of breadth where you can still gain pricing power and find new opportunity. I think you can argue that the HPE/Juniper deal is at least partly hoping the same thing, but perhaps adding a bit of realism. Niches cannot add up to a broad market, so a networking company has to be able to deal with the commoditization of breadth. Can you do both niches and breadth? I think it’s clear that both companies see AI as a “niche” developing. Do they both see it as a defense against commoditization?
Commoditization has one clear impact and two potential mitigators. The clear impact is that prices fall, so sustaining and raising revenues means selling more. One mitigator is that if you have account control, you can hope to influence planning to increase your market share. The other is that you might find new benefits to justify new spending. To promote those, you also need account control in the form of strategic influence.
No network vendor has led in strategic influence in almost four decades. Networks deliver value, not create it. In the early 1980s, IP networks supplanted IT-vendor networks (notably IBM’s SNA) because they were cheaper and could link the information locked in data centers with more workers, customers, and partners. Pent-up information equaled pent-up opportunity. But what is pent up is targeted for release, and gets released eventually. That’s why no network vendor has dominated since.
Both Cisco and Juniper hoped that AI could change that, and it could and yet could not. AI networking in itself is a niche; you need very fast fabric-like connectivity to host AI, and hosted AI is the only big opportunity in the near term. What makes it something potentially greater in the long term is, as I said in my blog on Tuesday, is that it could figure in the creation of a hierarchy of digital twins that could then open new applications with new network delivery needs. It’s the application benefits, which are really about software and it, that build more network utility, and spending.
IBM, right now, is the sole master of enterprise AI benefits. That means that they have leveraged their market-leading strategic influence to build real AI value. So far, it’s primarily in the area of business analytics, and so far it’s largely confined to IBM’s own major accounts. HPE, second currently in strategic influence, has a broader base of potential accounts to leverage unless IBM is able to extend its influence beyond its traditional accounts to the massive Red Hat base. And that is the question, not only for HPE/Juniper and IBM/Red Hat, but also for Cisco and maybe for niche vendors like Palo Alto too.
If either Cisco or HPE make this about AI networking alone, they almost surely cede any opportunity to build network traffic growth, so they’re contending for a niche and giving IBM time to figure out how to broaden the business case for IT. If that happens, then the battle of the future might be IBM and Red Hat against Broadcom chips for white boxes and VMware as an IT platform. IBM can make chips, of course, and they could buy a network player, but if you own the business case and can drive strategic decisions, why try to win networks too? Let them commoditize, make Broadcom compete in IT applications rather than validating their current strength.
Cisco and HPE/Juniper are both at risk if they hunker down on AI networking as a special case of hosting. AI should be thought of as a sort of high-level programming language, something that can build applications, create information and functionality that’s valuable enough for companies to want to deliver. This application effect is what drives up IT and network spending. Hardware infrastructure is valuable only to the extent that it supports applications. It’s popular to say “software” here, and I’ve been guilty of that simplification myself, but it’s not all software that’s valuable in itself. AI is a platform tool, like middleware. It has to hold up something.
Whose path to success would be easiest? I think that it’s obvious that IBM has an easier time; they’re the success story for many accounts already. IBM/Red Hat has sponsored InstructLab, an open-source project that enhances LLMs by “subsetting” them to specific missions. This isn’t well-known (only about a quarter of CIOs recognized or mentioned it), but it shows how IBM is shaping the broad opportunity for AI to a form that Red Hat can then help them spread through the market. Red Hat Enterprise Linux AI incorporates this capability, which makes it a viable “platform” for self-hosting AI.
HPE probably has an easier time than Cisco, with or without Juniper, because they already have strategic influence and you need servers to host AI GPUs. The question is whether Juniper adds to the picture, and if so how much. I’ve heard, but cannot confirm, that Juniper has been working on operational SLMs using a technique similar to InstructLab’s.
In the near term, Juniper and it’s AI-native positioning is a positive, I think. At the very least, it means that HPE has a smaller risk that Cisco would try to exploit its own target accounts, by addressing the data center AI cluster networking needs. In the long term, it could also be additive if HPE can play off Juniper’s standard configurations for AI and AI lab initiative, which surely they could. To get rid of the “could” qualifier, though, HPE needs to step up to the notion that AI is a means and not an end. Otherwise, Enterprise Linux AI could become the platform, anonymizing the server and reducing the impact of strategic influence at the server level, which hurts HPE.
For all these vendors, there’s a balance to be struck. Their core business is commoditizing, offering fewer and fewer opportunities for the sort of revenue growth Wall Street expects. There’s a temptation to rush into “growth” areas, but the risk is that no such area could offset the core commoditization they face, and trying to deal with that by jumping on multiple growth areas raises the risk of losing focus. It’s pretty clear this issue will have to be faced, perhaps by all the big players, in 2025.
