The Scene
Juniper Networks makes switches, routers, and security equipment, which help businesses keep their networks up and running but also may sound like technology from the 1990s.
Yet CEO Rami Rahim says the current artificial intelligence phenom has long been a part of his business and it has now taken over every aspect of the firm, from automating network maintenance to making graphics processors work more efficiently while training huge AI models.
In the edited conversation below, we talked about the AI transformation and, of course, how it’s helping the Aston Martin Formula One team network faster.
The View From Rami Rahim
Q: So where does Juniper sit in the grand scheme of AI things?
A: There’s a lot of hype, but AI is not new to Juniper. We’ve invested heavily in it. We’ve implemented it already in our products.
AI has already started to help us in operations. This is where artificial intelligence has the potential to make the network disappear. Those network managers have the proverbial pager on their belt that gets calls at 3 a.m. because something is not working. We want to completely eliminate that burden.
We’ve leveraged AI operations to crush the number of network incidents and make heroes out of the network operators. We have a customer, a global retailer, that has seen the number of store issues and point of sale issues drop dramatically by over 90%. We have a global software company that has seen day-to-day issues drop by over 90%.
Q: Can you name the customers or are you being purposely vague?
A: I’m being purposely vague, but I can name the global software company. It’s ServiceNow. And there are many others. For instance, colleges around the globe have a demanding customer — students who want the network on all the time with high bandwidth, but very little IT staff to run networks. That’s the perfect solution that leverages AI to basically have the network run itself, so you don’t need to have large IT teams.
In fact, I was just talking to a college CIO a few weeks ago — a customer — who told me that they don’t need to have expensive skilled IT people. They’re using college students to actually deploy and run their networks.
Q: Should experienced IT professionals be worried?
A: I don’t think so only because I think experienced IT professionals should not be getting issues in the middle of the night. I think skilled IT professionals should be thinking about how to advance the strategy of the organizations that they serve with a killer experience for the end user, not keeping the lights on.
Q: Are there other things that will change for consumers?
A: We want to accelerate the deployment of new networks. What used to take many months, maybe even a year to deploy, we can reduce down to a few weeks. And the reason is usually when you’re deploying these networks, they need to be configured and tuned, especially if it’s a Wi-Fi network. There’s all these sorts of radio frequency tuning that needs to happen in order to achieve the right experience for the end users. This is now all done by robots, by software. That’s ensuring the experience of things like Zoom or Teams calls is always great for the end user. And then there are all sorts of amazing new capabilities that you can introduce on top of a cloud-delivered, AI-driven network, such as location services.
Our solution has the ability to understand where people are, let’s say, a retail environment. Based on that information, they can provide a better, more digitally-engaged experience for those people. It can offer directions at a college campus to get you not just to the building, but even to your classroom.
Q: That kind of stuff already exists at places like Target and Home Depot. But are you saying the reason that you don’t have that at, say, a college campus, is because it’s expensive and complicated and now you’re making it a turnkey solution?
A: It can be expensive. It’s also cumbersome. If you look at the legacy solutions that require these separate beacons you have to attach to corridors that need batteries, they need replacements and maintenance over a period of time. The solution that we’ve actually built is based on virtual beacons. The Wi-Fi access points themselves also have the ability to understand where people are and provide the necessary telemetry and analytics.
Q: So it’s like triangulation from the access points?
A: It’s triangulation but it also requires some pretty sophisticated antenna arrays in these access points that we have built to achieve a level of accuracy. That makes these use cases possible.
Q: So I can’t just do this at home with my UniFi access points around the house?
A: You need these Mist access [a line of access points Juniper sells] points that have the specialized antenna arrays. I installed it in my house, so I know where people are in my house. Just for fun. Not that it’s actually useful but just because I want to be on top of the technology.
Q: You’re dogfooding your tech.
A: We say sipping our own champagne.
Q: What about generative AI and large language models? Is Juniper part of that revolution?
A: The new opportunity for Juniper is to offer the infrastructure to train and run inference. There’s a lot of talk about GPUs today. What people need are clusters of GPUs. In fact, they’re clustering 1000s of these GPUs. A cluster is really just a network -— a low latency, high performance network that carries a ton of traffic. We have the silicon technology, the management software that ties these GPUs together in a seamless, high performance way.
Q: So are your customers companies that want to do on premises inference on AI models? Who is that customer?
A: There are more and more customers that want to do on prem inference, but there are also cloud customers that want to do training on large language models themselves. Not hyperscalers. There are smaller customers that are cloud providers that were offering CPU services to the end user that are now pivoting to offer GPUs as a service.
Q: Can you give me an example of a company that would be doing that?
A: Allied Digital would be an example. I’m not saying they’re customers. This is an example of a non-hyperscale cloud provider that is offering training and inference services to end users.
As for end users for those companies, the sky’s the limit. In time, every company is going to be leveraging artificial intelligence in certain ways.
Q: So how much efficiency can you gain by using better networking techniques in GPU clusters?
A: What those people who train and apply inference care the most about is the efficient use of the GPUs. You do not want to connect a bunch of GPUs to a cluster and only use 50% of the processing power of those GPUs. You want them to be able to overcome periods of congestion. Congestion essentially becomes a bottleneck to sending traffic to the GPU. And you want to reduce latency, because latency is the determining factor for job completion time. And job completion time is sort of the duration of time it takes to learn something. The software is all designed to achieve non-blocking behavior, as close to 100% utilization of all the GPUs as possible and ultra low latency to minimize job completion time.
Q: Before the dawn of these giant transformer models, were you thinking about this kind of problem or is it a newer thing?
A: Most of the technology and know how that is required for the GPU cluster problem is already in-house. There are some new capabilities that ethernet requires in order for it to be ideal for this clustering problem and we’re solving those. Most of that will be in-house and some will develop over time.
Q: So you don’t have PhDs in-house trying to figure out the physics of how energy travels from one GPU to the next?
A: Not necessarily nuclear scientists, but definitely engineers. We’re working on some very cool things that I think could lead to real breakthroughs in the attributes of a network or GPU activity.
Q: Have you seen anything in universities or basic research that’s exciting?
A: Since you asked, Juniper published a paper with MIT on artificial intelligence a couple of years back. It leverages certain programmability aspects of our silicon technology in a way that helps minimize job completion time. So I think the breakthroughs are coming in ways that we understand and also in ways that we don’t yet understand.
Q: So you’re enabling LLMs and generative AI but is Juniper actually using those things?
A: We’re talking about things like input improving the customer experience by having an intelligent virtual agent or more personalized in-product help, or AI to speed execution. For example, a coding copilot, or legal contract reviews, and so on. This is also going to be incredibly disruptive and exciting, and enabling us to just move faster and increase the quality of our product offerings, and maybe even innovate in ways that we don’t even understand is possible.
Q: Now, what is that T-shirt you’re wearing?
A: We sponsor a Formula One team, Aston Martin. We don’t do this just to get a logo on a shirt. We do this to basically get more brand awareness, demand generation, and to invite customers and prospects to races.
Q: How long have you been sponsoring F1?
A: This is our second year and Aston Martin has improved a lot. They were an underdog and now they’re on a journey to a championship, hopefully, in a few years. They use our technology in their garages, in their brand new factory they built in the U.K. So it’s a technology partnership as well.
Q: Are you using any AI to help them?
A: You bet. They rely on our network to basically achieve better latency characteristics when they’re pulling all of the telemetry from the cars around the track, pulling them into their command center in that factory so that they can determine what’s working well, what’s not working well, and make adjustments.