AI will reconfigure every job, but faces 3 key constraints: CISCO’s Jeetu Patel at Davos

We are probably witnessing one of the most profound platform shifts that we'll have ever experienced in human history. And so every job will get reconfigured. Every workflow is going to change, and the way in which we live our lives is going to be very different. Right? So then you say, okay, if that's the case, what are the things that might hold it back? Right? Because what ends up happening is, in the short term, most people grossly overestimate the impact of these technologies, but in the long term, they tend to massively underestimate the impact that these technologies have. So you start thinking about this and going, what are the constraints? And we think there are three major constraints. The first constraint is you just don't have enough infrastructure. There's not enough power, compute, and network bandwidth in the world to fulfill the needs of AI. It's massively in scarcity right now. So that's the first constraint. The second big constraint is there's a trust gap. And what I mean by the trust gap is, if you can't trust these systems, you're not going to use them. And so security and safety become prerequisites for the adoption of these systems. Right? And then the third area is what we call a data gap, where right now, so far, these AI models have been trained on human-generated data that's publicly available on the internet, and that'll continue, but we are starting to run out of human-generated data being publicly available on the internet. So what people are doing is the way that these models are evolving to keep the scaling laws going is you have synthetic data, and then you'll also have machine data, which will be a very, very prominent area because as agents get more and more prolific, you'll have the agents will generate logs of activity that'll go into machine data, and so you'll have a lot of time series data that you need to make sure that you factor in and merge that with human-generated data so you can have additional insights that come out. And so those are the three big constraints: infrastructure, trust, and data.

And then you say, "All right, so if those are the constraints, what can be done to go out and overcome those constraints?" And that's where you're starting to see all of the opportunity, huge amount of investments that are being made. Like if you think about data centers right now, that, you know, you're not bringing power to the data centers, you're building data centers where the power is. You're trying to make sure that, you know, every country has a need for token generation because your ability to generate tokens will be directly proportionate to your ability to have economic prosperity as well as, you know, kind of protect your national security. And so these tend to be profound implications that need to be thought through. And Cisco is at the center of it all because we provide networking, compute, and connectivity infrastructure for AI data centers. We also provide safety and security, not just for cyber defense, but also for securing AI itself. And then thirdly, we have, you know, a platform that we had acquired called Splunk, which is one of the world's leaders in machine-generated data and being able to make sure that you can use time series data to be correlated with human-generated data so you can get better insights. So in each one of those areas, we've got, you know, billions of dollars of investments. And, you know, we're really looking forward to making sure that we can actually progress the cause for AI.

Yeah, I think you'll start to see, you know, green data center buildouts that are very, very prominent. You know, I think the other area that we haven't really talked that much about in data centers that'll over the course of the next few years get really prominent is what happens with space data centers. Because if you think about what the constraints are right now, when you think of data centers, all the cooling that's required for the data centers, 90% of the weight in a rack tends to be the cooling infrastructure. You know, it's not the compute, it's the cooling that actually causes a huge amount of the weight. When you go into space, you've got solar, and the intensity of solar is much higher than what you have on Earth. And so the cooling can tend to be at a very different level of economic proportion than what you get over here. And then the energy availability is much higher too. So I think there's space data centers that are also being built out at the same time. So I think there's going to be innovation across a multitude of different directions on efficiency of power to power the data centers, on making sure that you've got cooling apparatus done in the right way, to make sure that you've got the right level of connectivity of the data center so you can do it where the power is available. For example, one of the areas that we've done a lot of innovation is not just, you know, the way that data centers worked is you had these GPUs and the models would sit on a GPU and then get trained. As the models got bigger, they didn't sit on a single GPU. They then needed to sit on multiple GPUs that were together. So you actually created a server with multiple GPUs. But then the models got bigger than what could sit on a single server. So then you had a series of servers that actually got tied together in a rack, and so you had the entire racketworked. But that got to be not sufficient either because the models kept getting bigger. So now instead of a rack, you have multiple racks within a data center that you started networking together, and now you're at the point where even that is not sufficient. And so what are we doing? We're actually taking data centers that might be hundreds of kilometers apart and creating an ultracluster with them so that there's coherent clusters across data centers that might be hundreds of kilometers apart that can beworked together where the latency is still low enough that it operates like a coherent cluster. And that's where, you know, Cisco has built specialized network ASIC chips so that you can have technologies like deep buffering for long-range, you know, kind of communication. You've got optics, all of these kind of capabilities that you need to have to make sure that you can actually overcome the constraints of power and overcome the constraints that are needed so that you can deliver the right level of infrastructure apparatus for AI needs. And this is only going to compound because as agents get to be working 7 by 24 on average, you'll have maybe 10 to 100 agents per human. If you have 8 billion humans on the planet, that you might have between 80 and 800 billion agents, and they're all going to be working 7 by 24. So imagine the level of kind of infrastructure requirements that are going to be needed. It's going to be non-trivial.

They're actually starting to get built out right now. And so the thing that's really interesting about this market is how quickly the entire industry seems to continue to keep moving, and it's moving fast across a multitude of different directions and dimensions. So I feel like, you know, one of the things that you'll find hard is in a few years, maybe in two years or three years, it'll be impossible to predict long term because it's not because your ability changed, it's because the compression of innovation is so much that now what used to happen in 10 years is now happening within 6 months. And so it gets very hard to predict what happens, you know, 3 years, 5 years, 7 years out because the compounding effect of that is pretty profound.