First, there was machine learning. Then, generative AI. Now, we’re in the midst of the agentic AI revolution. But agentic AI is a waypoint rather than an endpoint. So, what’s next? The answer is physical AI.
But what is physical AI, in practical terms? What is it good for, how mature is the technology, what kinds of networks will it need and how can we all prepare for what’s coming? Nvidia and Nokia have answers.
Catch the video at top, listen to the audio edition on the go or read through our transcript below.
New this season: You view this and future episodes on YouTube right here!
This podcast is written and hosted by Diana Goovaerts. It is edited by Diana Goovaerts and Matt Rickman. Liz Coyne is our executive producer. Special thanks to our guest, Nvidia SVP of Telecom Ronnie Vasishta, and to Nokia for providing a clip of their Chief Technology and AI Officer Pallavi Mahajan.
To learn more about the topics in this episode:
- Here's why physical AI is a big deal
- Will AI agents really raise the network traffic baseline?
- T-Mobile’s John Saw on AI, RAN and that big deal with Nvidia
- Nvidia wants 6G to be made in America - with Nokia's help
This transcript has been lightly edited for clarity.
Diana Goovaerts, Fierce Network (DG):
We all know AI-based movie villains like M3gan, Terminator’s Skynet and VIKI from I robot are fiction. After all, AI is just something that runs in data centers, right? RIGHT?
Wrong.
I’m Diana Goovaerts and this is The Five Nine.
First, there was machine learning. Then, generative AI. Now, we’re in the midst of the agentic AI revolution. But agentic AI is a waypoint rather than an endpoint. So, what’s next? The answer is physical AI.
Physical AI refers to AI that is applied by machines to understand and enable complex interactions with the world.
But what is Physical AI, in practical terms? What is it good for, how mature is the technology, what kinds of networks will it need and how can we all prepare for what’s coming?
We’re here today with Nvidia’s SVP of Telecom Ronnie Vasishta to get the details and sort the AI fact from fiction.
Ronnie, great to have you.
Ronnie Vasishta, SVP of Telecom, Nvidia (RV): Nice to be here, Diana.
DG: So I want to start with a baseline question. Uh, what is physical ai, uh, and how is it different, say, from an LLM? What does it look like in the real world? Let's kind of level set here.
RV: Yeah, that's a good question. So think about physical AI as AI that perceives reasons and acts in the real world. So rather than taking maybe text as a input, the input to the AI is cameras, lidar, depth sensors, audio. And then what AI can do, the physical AI can do, is comprehend those inputs and the output are control signals and motions that can drive real world actions. So it is quite different than a large language model. But as we come on into this conversation, you'll understand that there's obviously similarities in being able to transform a series of inputs into a series of output, meaningful outputs, um, based upon the knowledge of the environment.
DG: So I think one of the things about physical AI that's interesting to me is, correct me if I'm wrong, it lives on devices that interact with people in the real world as opposed to say on a server in a data center. Is that, am I thinking about that correctly?
RV: To some extent, yes. But then if you think about the, let's say just initially the size of the model, but also the other aspects around latency or the ability to provide, you know, performance per watt, as you start to go through that training of physical AI, a lot of that training is done in larger data centers as you go through the simulation.
And we'll come onto this maybe a little bit later, but as you come onto the simulation of that world foundational model or the model that represents the world a lot of that simulation can be done in a data center where you have, say, a digital twin of the physical world that's physically accurate. When you come onto the inference of the model and taking in some of those real world environments that can be done on device. But we're also – so think of it as like different computers that act within different parts of the workflow of physical AI.
DG: Okay. I want to touch on the whole 'connecting those different pieces of the models via a network' in just a minute. But let's talk about physical AI as a technology, right? You mentioned the world foundation models. I think there may be only a couple of those right now. So my, my big question is how mature is this as a technology and what kind of use cases are we working towards in terms of why are we even doing this? Right? So what is the end goal and how mature are we so far?
RV: Yeah. So, you know, we talk about the next era of AI being the physical AI era, which by its mere nature says that it's a newly introduced concept, but it's already being deployed. If you think about autonomous mobile robots that operate within a factory environment, or if you think about some of even autonomous vehicles on the roads, essentially a robots that are understanding the real world environment around them to make decisions instantly to on how they operate. So the physical AI actually is already being deployed in real world environments.
In terms of how that growth continues, you know, we see going into a lot of edge or deployment use cases, humanoid type robots, delivery robots, drones, things that operate in the real world that have to understand the real world and doing real functions and jobs in the real world.
DG: So I guess my next question is you were kind of talking about the different elements of the models where different parts of the computing happen. So, let's kind of drill into that for a quick second. Once these physical AI robots or AI-enabled robots are out in the real world, how much is done on device versus via the cloud via a cloud connection?
And if there needs to be a cloud connection, what are the implications for network connectivity? Let's start there before we drill into more of the network piece because I have so many questions.
RV: So a lot of the training, as I said, of the world foundational models happen in the cloud because of the compute requirements and then the simulation of the data, whether it be synthetic data of use cases that can't always be seen in the real world but are use cases to help train the – think of it about training the brain of the robot – happen in the cloud in a digital twin environment as well.
So that then when you go to deploy – and Nvidia has different platforms that we use for deployment, whether it be within an autonomous vehicle or whether it be an autonomous mobile robot – when you go to deploy these models and infer and really drive use cases, sometimes there will be a requirement to connect to infrastructure outside of the device itself.
But the device itself has to be able to operate autonomously. So, you know, the connection can't be something that absolutely is required for the operation of the robot, but there does need often to be a connection for downloading new models in terms of new training, new experiences, uploading information that's coming from the robot, whether it be video information or location information.
So that connection actually is important, but it's not absolutely necessary in all cases.
DG: I think that this is something that telecoms providers and vendors are already thinking about. So Nokia just had one of its investor days and it was talking about moving its standard from five nines of reliability, which is typical in the telecom world, to six nines in part to support the needs of 6G, to support the needs of new use cases like physical AI.
Clip of Pallavi Mahajan, Chief Technology and AI Officer, Nokia:
Now there is another disruption waiting to happen in this AI super cycle, and I can actually see it coming. Because AI is now on the cusp of the next wave of transformation, which is physical AI. This is where the boundaries of the physical and the digital world are going to get blurred.
Now think about it. This is the autonomous vehicles, drones, AR/VR glasses, intelligent factories, healthcare. This is the world where critical, essential services will demand that the network should always be on, where sub split second decisions need to be made, and every millisecond matters.
And with physical AI, we will have robots standing hand to hand with humans, with heavy machinery. And now just imagine these robots, you know, the safety and the motion control loops that they need, they need decisions to be taken in the matter of microseconds. Now, one network slip and the robot can actually miss a safety stop. So in these environments, it's about safety.
Now, let's let's step back and look at, you know, this whole transformation that I'm talking of. Gen AI was three years back. 2025 is the year of agenic AI. And now I can clearly see that physical AI is knocking at our doors. And this evolution of AI will exponentially change the dimensions of the networks themselves, whether it is bandwidth, whether it is capacity, whether it is latency, whether it is reliability, all of these KPIs are changing with AI.
DG: You mentioned earlier cars on the road. It seems important that they have the most up-to-date software. It seems important that they can upload the data that they need to continue driving safely or operating safely. So even though the connection may not need to be constant, what kinds of networks and what kinds of network attributes do we need to enable physical AI?
RV: Good question. You know, one of the things that, if you think about robots or you think about autonomous vehicles as robots or even base stations as robots, as you start to operate the radio base station with beam forming and antenna orientation, for all of those robots time synchronization is very important.
And nowhere better in the world is time synchronization developed than in the mobile networks. So the benefit that mobile networks provide in terms of understanding, as I said earlier, the 3D space and time is a mobile network. Now there will be some applications where you can have connectivity that is actually quite important.
As I said, there's somewhere, you know, the connectivity is a nice to have, especially as you're starting to understand like fleet management. You know, you're not gonna make a left or right turn decision based upon a connection to a radio base station, for instance, or infrastructure nearby.
But you do want to understand where all those autonomous vehicles are and fleet management, whether they be robots or other things. And for that, you do often will need connection and you'll often need a real time digital twin to be able to reorientate and replan those things. One thing that's very important as it comes to the mobile network operators and the equipment providers is that time synchronization of the network is important in terms of the SLA or the service level, agreement of the application.
So if you have, for instance, a drone where you need to know where the drone is and the drone is actually being managed on delivery, there will be a local throughput of intelligence and the way that the robot can reorientate its surroundings, but there will also be a connection that's very important in terms of latency. Now, latency isn't all about the distance. Sometimes it's distance, but also there's a compute latency requirement. The compute latency requirement is actually sometimes even, how much memory, local memory is there? How is the pipeline of the video pipeline, for instance, comprehended within a video language model? And for that, often compute should be more localized and dimension for the application.
DG: Okay, so I know, I know you're talking a about one of telecom strengths is time synchronization. Let's take a beat and explain that. What is that and why is it so important before we move forward?
RV: Yeah. So all the networks around the world are synchronized on time. And when you start to think about, in the real world, you have a number of autonomous entities, whether it be robots or vehicles or other things, and they're communicating not just with their environment, but with each other and with the orchestration platform that's understanding where they are and what they're doing.
Of course, you want to make sure that all of them are synchronized in time and that's where networks add a tremendous benefit because you are able to, within microseconds and milliseconds, be able to understand the communication, but also the location.
DG: Perfect. One thing we're talking about is networks and their importance and some of the attributes they bring and why those are so important to physical AI. My next question is what do telecoms need to do to prepare for physical AI, if anything? Do they just need to continue maintaining their networks? Do they need to upgrade them? Should they be thinking about this as something they can potentially monetize? Right? If time synchronization is so important, maybe they can sell some sort of network slice with an SLA package around it. Talk to me about the opportunity here for telcos as well as what they need to do to keep up.
RV: Okay, so this is really, I think, a new era for telcos. The monetization opportunity that exists even for generative AI for telcos, outside of physical AI, is there today. Think about token generation, many applications, even if it's a voice application where you are talking to your phone or talking to your smart glasses, et cetera, you know, you've got one and a half seconds of latency before the conversation becomes somewhat weird and, you know, not real. And telcos can enable that to happen.
You ask the question around what the telcos need to do. Firstly, the topology of the intelligence nodes. So AI factories, larger AI factories, but then more localized point of presence or mobile switching offices. These are all great locations for - think of it as the AI factory. It's the taking data in and outputting tokens, monetizing those tokens on things like a service level agreement.
That could be time to first token, could be tokens per watt, or to a number of tokens per dollar. You know, telcos need to start thinking about, 'Hey, I've got this infrastructure. I can use it in a point of presence to actually manage physical AI.' To do that, I might need slicing as a way of delivering a guaranteed bandwidth or guaranteed jitter, guaranteed packet loss to that particular robot or those fleet of robots. I need to think about the infrastructure as, how do I ensure that I can have compute latency that's guaranteed in that point of presence, [such] that sending the data, say the video data, all the way back to the cloud doesn't make sense.
It just costs a lot of money to send that data, ingress, egress and just transport. So I can do all of that locally. [That's] a new way of thinking for telcos that's definitely, I think, monetizable.
DG: It sounds to me like you're talking about deploying edge compute or AI kind of not on the RAN, but near the RAN. Am I thinking about that correctly?
RV: Yeah, absolutely. So think about intelligently connecting intelligence. The points of intelligence are driven by the user and use application. And it can be even all the way up to a base station, right? You can have a computer at the base station. In fact, Nvidia, we recently at the GTC conference announced a compute platform called ARC Pro.
ARC Pro is a computer that will process tokens as well as be able to run the full 5G and 6G stack. That same computer can be all the way out at a base station. We can also have computers at points of presence that do the same thing, and you can aggregate base stations into that computer at mobile switching offices or baseband hotels.
So yeah, the topology of the network will support the application and will be driven by the application, and all of these computers can be orchestrated together such that you can have a network of intelligence.
DG: Kind of like the scale out principle or scale across that Nvidia has been talking about quite a bit in the data center realm. I do want to ask a question though, right. So it, I think the ARC PRO system includes GPU technology, correct?
RV: Yes, absolutely.
DG: The thing that I've heard in telecom circles is that they don't see a need for GPUs necessarily out at those base stations. And the primary concern that I've heard about right now is cost. So can you maybe address why this kind of compute is required, right? Like, why can't we just do it with the CPUs that are already out there and how telcos can think about balancing the upfront cost of GPUs, which no offense are kind of expensive, with the potential revenue streams?
RV: All right, so you asked a really good question, Diana. Let me take the fundamental premise that you explained and first refute that. GPUs come in all shapes and sizes. You know, the benefit that NVIDIA has is the CUDA - which is our programming environment - development environment is across all of these types of GPUs.
Of course GPUs can reside in data centers and perf per watt, you know, is great in terms of the heavy computational requirements. But we have GPUs that sit on robots, you know, small, low power, you know, 40 type watts or even less. We have GPUs that go in laptops. So the performance capability of these GPUs now is increasing generation after generation after generation, and then within generation, because now these are all software defined. You can improve performance even within generation.
So now think about this new world. You can have a low cost - think of it as a performance per dollar - GPU at a base station that complies with the performance per watt requirements of a base station. So now you get the performance, now you get the TCO that a base station actually has traditionally been met by developing custom silicon.
That's a very new but real capability now and as we go forward in generations, that's only gonna improve. So the first thing is there is an opportunity now to have a completely software defined GPU running a full 5G, 6G stack at a base station. Now that you have that GPU in there, you can do two other things.
You can use it for AI applications as we just talked about, but you can also use AI to improve the signal processing of the RAN network itself. So this opens up a whole new area of research and development, which we're already starting to see benefits of for things like layer two scheduling, for SRS beam forming.
All of these capabilities that now you have this GPU there that still conforms to the requirements of cost and power of a base station can do other things as well.
DG: I don't wanna fall too far down the AI-RAN rabbit hole. Ronnie, I'll have to bring you back for another discussion on that specifically. But what kind of hurdles are standing in the way of physical AI today? So, is it model development? Is it the network piece that we've been discussing? Is it something else? Is it getting people to trust putting AI in robots?
RV: Yeah, that's a great question. You know, I would say it is not the network per se today. It would be more, you know, the model training, the deployment of the models. There's a lot of data that's required in training and simulation within digital twin worlds, physically accurate digital twins.
There's still work that's happening. And then of course, the cost effective deployment into certain applications as we start to go to enterprise and factories and go beyond factories into consumer as well.
DG: Okay. If we have to end it on a single big takeaway for physical ai, what would you tell everyone?
RV: The topic of physical AI is talked about as the next era of AI. Of course, the opportunity set is huge. We are already starting to deploy. So even though people talk about it in the future, it's really starting now. I think that's the key takeaway: physical AI is gonna be around us all the time and it give us so many more opportunities.
DG: All right, Ronnie, I think we will leave it there. Thank you so much for your time today.
RV: All right. Thank you very much, Diana. Been a pleasure talking to you.