Edge Continuum: Where AI belongs from sensors to cloud (with Azita Arvani)

It was great to have a chance to share the stage with Azita Arvani, Stanford Sloan Fellow at Sensors Converge 2026 to talk about the "Edge Continuum" as it relates to edge AI and what I call the Perception Edge (Sensors + Edge AI).

Azita and I talk about what the edge continuum is, what is shaping the present and future of the perception edge, where AI makes sense across the edge, and what the role of sensors and edge AI will play in physical AI.

Thanks to David Drain and the Sensors Converge team for setting us up with this spot to share with this year's attendees! Congrats on another great event!

Transcript

Leonard Lee 0:00

Good morning, everyone. my name's Leonard Lee. I'm the executive analyst and the only analyst at, Next Curve. I have the pleasure of being on stage here with a good friend, Azita Arvani, who's formerly a CEO North America of Rakuten, has a, a storied history at, Nokia and a bunch of other places, right? She's like a big name in telco. And so why don't we get into this? what is Edge Continuum?

Azita Arvani 0:32

let me start by saying that it's a great time to be in sensors. You ask me why? Number one, we are in physical AI era, right? so we've done a lot of things in digital AI, now physical AI, and sensors are the gateway to that physical AI. number two is that when it comes to physical AI, there is a big limit in how much data is already available, right? So on internet, you can scrape the internet and get all these great things and some not so great things for digital AI. But when it comes to physical AI, we still don't have that much data. So sensor data is super important. and number three is this is-- we are moving past the AI training era to AI inference era. We actually have silicon that's AI inference specific. So we've seen the announcements from Google with TPU8i and, AWS Inferentia, et cetera, et cetera. So now we have silicons that are much lower power, inference specific that can go closer to the sensors. So I think it's a great time to be in sensors. now, what is Edge Continuum? At this conference, and generally, people think about edge AI as on-device AI, and then the, the other AI being going all the way to the cloud. So for example, you know, we were talking, you know, it was-- this is not a smart glass, but if it was a smart glass, we would have a lot of sensors around the periphery of this, and then they put a compute device at the nose of this device. So it takes some of that sensory information and processes it on this device, and then it sends some of it to the phone, right? So that's where we think about edge AI or even Aura devices. For those of you who monitor your sleep, some of that, not all of that sleep data goes to your phone, only the stages of your sleep goes to this so that it's outcome-based. So- And then there's the cloud. I'm here to say that there is another piece of this in between the, device/phone all the way to the cloud. So we can have the sensors, and it could come to a hub. You could have, compute at that hub, or you could have like a private, 5G hub there. You could have a, a hyperscaler edge, private there. You could have telcos that have many edges that have computes that are underutilized. It could be used for your AI. and then you could go to the cloud. So there's this whole edge continuum that starts from sensing to then, some sort of a hub, then various stages of edge that gets to the cloud. So when you think about your use case, it's really end-to-end and thinking about where you want to put that, distribution of that intelligence. And it doesn't have to be one or the other. It could be a hybrid, scenario

Leonard Lee 3:49

let's talk about some of the things that we're seeing here at the conference that inform... should inform folks on where some of that placement makes sense, right? Obviously, sensors go everywhere. They're across various industries. when you look at intelligent-- quote-unquote, "intelligent systems," you're looking at a wide range of architectures, and also the economics are very different for different situations, whether a smart car, smart factory, you're trying to instrument a machine, smart building, there's this whole cornucopia of tools that you have, but then also the number of use cases are very, wide-ranging compared to, let's say, what we're seeing in these AI data centers, for instance, which are becoming great, use cases for edge AI as well as IoT. Because when you look at, the operations aspect of a data center, there's a lot of physical, environments that need to be managed in order for these, data centers to be resilient, reliable, and support some of these, very expensive workloads, training workloads, right? So, sensors, IoT, AI, become important. So what are some of the things that y-you're seeing here that are informing the way you're thinking about that continuum?

Azita Arvani 5:13

Yeah. So, a couple of things. First, for example, in, in the, in company that I work with, these devices now have GPUs on them. for one reason, they go to places where they cannot be necessarily connected, so they cannot depend on connectivity all the time, so they have them on device. So now we have a lot of sensors like LiDARs and cameras and so forth. what I'm seeing, here is the LiDARs, instead of being one big expensive LiDAR, we have, smaller, lower cost LiDARs, and you can have multiple of them that can then be on device, and that brings the cost of that down. And then you could just imagine that if the cost of the sensors go down, then maybe you could have more budget for AI compute on device. So that's, that's one thing that I'm seeing. But- If you think about the, a venue, for example, I live in Los Angeles. Clippers just, Steve Ballmer, the owner, had a, started a new Clippers arena in Los Angeles. It started in towards late twenty twenty-four, and there are sensors everywhere. Everywhere. They know w-whether you're sitting in your seat, how much you're cheering, how-- If, if you're loud, then you get more rewarding, all kinds of things. There's grab-and-go, stores and everything and anything that could be digitized, visualized is, is done that way. When you walk in Have these face recognition things, so you don't even have to have your phone. So that's on-device AI that you have it. And then when you, with these, grab-and-go type of stores, those are all connected via fiber to a local compute on the venue. So then that's where the AI is sitting. so and then you have these low-level sensors, like for example, in the seat, to see if you're in the seat, how noisy you are, and those are connected via Bluetooth and, UWB to a hub, and that hub can then have, AI running on it that could figure out what to do with, with the information it gets from the seat. So even in one venue, you have multiple ways that you could deal with sensor and where you put the AI for those sensors.

Leonard Lee 7:35

A couple things that I'm noticing, and these are probably, like, trends that folks in the sensors indus- industry and industry already know. But from a physical perspective, and device perspective, there's consolidation happening. So we see lidars that are multifunctional, right? And so you're starting to, be able to make these things multipurpose, right? And then on the virtualization end or the digital end of, or the AI-ish end of things, so if we're to think about edge AI, we're seeing, new possibilities in sensor fusion, so different classes of, of, sensors coming together to create either net new sensory capabilities or augment or make more flexible and extensible an existing, sensor set. one of the things that Karthik mentioned, of going from just sensing things to understanding, so infusing some of that cognitive capability much closer to the sensor, which I think is a counterpoint to what we traditionally have thought of in the IoT world, where we're sending streams of data up, right? And so, going forward, as we look at the edge continuum, my view is this, is that what you folks are working on here and what we're seeing here is a counterpoint to what we've traditionally thought of in terms of how we architect, intelligent systems or cognitive systems, what have you, whether it's a smart factory or smart home, whatever. And, to your point, this is all about localizing things on device for all the benefits that it, it has ar-architecturally been represented to, bring in terms of security, privacy, latency. Right.

Azita Arvani 9:24

And then one other thing I wanted to mention that I think, I haven't heard that much being talked about here is about synthetic data, right? so we, we would like to have insights and outcomes instead of the whole amount of data being transferred from sensors to, next stop along the way. However, if you do want to keep that sensor data for archival purposes or further training or some other way, you could actually, compact that, that, that data, Via a model at the source, and then just transfer the model over. You could think of it like, freeze-drying your data and then transferring it over, so then the cost of transfer is much lower, and then you unfreeze it at the destination. So that's another thing that's available, and those companies are there, and I, I advise one specific data company, but there are, others as well. So you should think about that as well. So it's not just, you either s-send all of the sensor data over or you just send the outcomes. You also have a way of, capturing the essence of that data for further, AI training-

Leonard Lee 10:42

Yeah

Azita Arvani 10:42

or other uses.

Leonard Lee 10:43

Yeah, we see some of that out there in terms of like digital twining and simulation, right? so with synthetic data, The way I look at it, it's like once removed from the sensor. So if you're doing tw- digital twinning, you're only gonna be able to capture live data, real-time data that's coming off of, sensors, right? and you're making a great point. Next step with the synthetic data, you're gonna create a model and in a simulation, and that's what we see in automotive, right?

Azita Arvani 11:12

But what I'm saying is not using synthetic data to generate new cases, which is one use case of synthetic data, but actually synthetic data to capture what's already generated.

Leonard Lee 11:22

Yeah. Okay. And so we have one minute, less than one minute. Any final thoughts? Any, points that you wanna-- points of inspiration that you want to share with our wonderful audience here?

Azita Arvani 11:33

Yeah. the thing about sensors, as we heard from the last presentation, sensors are not just there to sense, but they're there to think and understand. And there are so many tools available now, and Edge is really not just one place. It goes, in-- You have many, many options. And actually, the challenge is: how do you use that option? And these could be dynamic as well. They don't have-- You don't have to start somewhere and just be stuck with it. You have orchestration software that can move your workloads around depending on, your SLA latency costs, all the things that we talked about.

Leonard Lee 12:09

And, my closing remark here is, a bit of inspiration for everyone here, a reason to get excited. So you've seen, like with the AI trend, how from a, a silicon perspective, we got excited about the accelerator, and then it became about networking, and then it became memory. This is all about how the-- as we move away from-- the focus moves away from, pre-training and post-training, all this stuff toward inference and agentic and now physical, you have all these constraints, right? Well, when you get to physical, you guys have let the sensor be the constraint, right? And so it's very possible in the very near future that, sensors will matter a lot because you need the IoT, and you need that understanding in order to hydrate these physical AI models. And then this ambition to create world models, it's all gonna be based on physics and, observations of the physical world, and that only happens through sensors. So, anyways, love to leave that with everyone. And, and hey, Azita, it's great hanging out with you as always. Same

Azita Arvani 13:20

here as always.

Leonard Lee 13:21

It's good. And I, I love the topic and, look forward to continuing this conversation about the edge computing one. And thank you everyone for attending, and we appreciate your, attention.

Azita Arvani 13:31

you.