The Future of Intelligent Edge Devices | Edge of Tomorrow – The Edge AI Debate

Pete Bernard (00:01.516)

All right, here we are. Welcome everybody to the Edge of Tomorrow. I think that’s what it’s called, right? Edge of Tomorrow. But I’m here with Scott Hanson from Ambiq. And we’re gonna spend a chunk of time talking about edge AI as usual, for me at least, and probably for Scott as well. I know he mentioned he talks to investors half the time he talks to folks about this stuff.

So, I think we’re both pretty well practiced in this topic, but I think we will learn some new things today, about Ambiq specifically and get a better sense of what’s actually happening in the space, the AI space that I know I consider to be the most important part of AI, which is happening where the data is created in the real world, where there’s real impact and all that good stuff. So, Scott, welcome to the podcast.

Scott Hanson (00:50.261)

Thank you, I appreciate you having me.

Pete Bernard (00:52.748)

Are you in Austin right now?

Scott Hanson (00:54.847)

Yeah, in Austin we’re right at the front edge of summer here, so it’s getting warm.

Pete Bernard (00:59.926)

Ooh, nice. Yeah, I’m here in Bellevue, Washington, and it is raining, but that’s not too surprising. So it’s getting a little warm in Austin. It’s raining in Bellevue. So everything is normal. No problems. but yeah, so before we get into kind of the nitty gritty, give us a little bit of a backstory, maybe origin story on Ambiq. And you know, you’re a publicly traded company, but you’re I think you’re originally a Michigander, right?

Scott Hanson (01:05.931)

Nothing new right.

Scott Hanson (01:11.605)

world is right, yeah.

Scott Hanson (01:30.015)

Yeah, yeah. So Ambiq was founded out of my PhD research at University of Michigan. We were building little cubic millimeter computers that would be used in the battlefield, used in medical implants. So we we had these little cubic millimeter devices that would have a battery, processor, sensor, radio, everything. And then you implant in someone’s eye in in a glaucoma patient or or some other patient type. So really cool, really cutting edge stuff. Got to the point where a number of companies said, hey, I’d love to see this commercialized. And we thought we were going to be a medical device company at first, right? But there’s the there’s one thing you learn is that that takes forever to to start a company. And there’s there’s only one type of company that’s like marginally faster than that. And that’s to build a chip company. And so that’s what we did is we built a microprocessor company. Ambiq today builds the world’s most energy efficient chips. The vision would be that everything from a 50 cent processor to a $50 processor and the whole spectrum in between is what we deliver and and the point is as AI has has broken out of the data center, we’ve got a ton of compute needs out at the edge, out in battery constrained devices, and Ambiq delivers that. So it’s been it’s been a fun journey. We got a really cool technology that I can I can talk more about if if interesting. But in twenty twenty six there’s no place more exciting to be than building ultra low power processors for the edge for for AI. So exciting times.

Pete Bernard (02:57.102)

Yeah, for sure, for sure. So how big is Ambiq these days?

Scott Hanson (03:01.795)

we’re roughly two hundred and twenty people. We’ve shipped in excess of three hundred million units over the years. So, you know, I’m I’m wearing a couple of our customers’ products right now. So all kinds of wearable devices, all kinds of industrial electronics, medical devices, smart home devices, and you know, the other metric on size is just like we just reported our

our Q1 earnings and and quarterly revenue was at about 25 million. So company’s growing really healthily and we’ve got a a nice outlook for the rest of the year. So I I’d say that we spent a number of years really trying to figure out what the technology was going to look like. Building a chip company is really hard. The cost of mistakes is really high and so you have to move a bit slower than software companies. But you know as we were developing our core competency in low power

Pete Bernard (03:32.964)

Nice.

Pete Bernard (03:44.579)

Yeah.

Scott Hanson (03:53.759)

This IoT thing was kind of emerging and then that morphed into what is now edge AI and and here we are today. It’s I think the timing is perfect.

Pete Bernard (04:01.346)

Yeah. Yeah. No, it’s a good point you make about chip companies, hardware companies. I mean, I worked at Microsoft, software company, although we built a lot of hardware too. And I actually worked in different groups there and it was interesting to see the culture of the different groups and how they worked in terms of attention to detail, right? So working in the hardware team, you know, you gotta get it right. So software team, you know, push it out there, you know.

Scott Hanson (04:22.315)

Absolutely.

Scott Hanson (04:27.327)

Well my yeah, my software team will get mad about with with me saying this, but you know, he here’s the here’s the consequence of of having a bad software or a a bug in software is that you recompile and you you spin it and and and deliver a new product. And in in chips, a recompile means I’ve got to tape out a whole new, you know, mass set. It cost me millions of dollars and there’s s a multi month delay. So yeah, it’s a different it’s a different mindset for sure. And actually one of the things when I when I first founded the company, I made certain that I had

Pete Bernard (04:29.954)

Pete Bernard (04:41.752)

No. Right. Right.

Scott Hanson (04:54.933)

far more experienced people than me building these chips. Right. I was how old was I? I was you know in my mid twenties starting this company. I w I had a PhD but nothing no real work experience. And so it was really important that I found people who had been building chips for for a career. And we did that at first and then over time we kind of we we started bringing in more junior engineers. But yeah, it was really important to have some smart people, experienced people along for the ride.

Pete Bernard (05:18.85)

Yeah. Yeah. No, you’re right. I mean, yeah, and people that really know the operations and the timescales and the equipment and you know, the testing and the validation. I mean, all that stuff comes after the great ideas, you know, in terms of executing them. So that’s pretty critical stuff. I guess one of the questions we can start out with here, and we’re gonna talk a little bit about kind of where this edge I vision is heading, and maybe we can even do a little one hundred level edge AI for those that are not.

as embedded in it as we are. and then we’ll talk about the vision. Then we’re going to talk a little bit about sort of where the action is today. You mentioned you’re shipping a lot of units. So there’s a lot of stuff going on out there. and then, you know, one of the interesting things will be sort of we hear a lot about agentic AI and physical AI and generative AI and

Scott Hanson (05:57.354)

Yeah.

Scott Hanson (06:05.872)

Mm-hmm, mm-hmm.

Pete Bernard (06:07.67)

Analog AI and neuromorphic AI and embodied AI. So it’s like every day there’s another thing. so we’ll talk about that as well. But let me actually let me s scroll back a little bit to your product statement about development. I guess the question is, you know, what I find is that this edge AI space is moving so quickly. I was actually talking to someone from Google this morning and we were just like trying to

top each other on the newest thing that we’re what’s happening. And it’s like every week there’s a new thing. You know, as we’re talking about like making chips or making hardware. I mean, it’s a long development cycle. It’s almost, it feels like the the AI innovation cycle is outpacing the product development cycle a little bit. And it’s it’s hard to sort of like almost reconcile the pace of innovation with what it just takes to make and ship a product. I mean, how do you guys think about that?

Scott Hanson (06:53.579)

Yeah, indeed.

Well, I a couple of ways to to answer that. So first of all, the the chip team moves at a certain pace, right? And and and offers products, but we have a software team. In fact, there’s on any given day, there’s more people writing software at Ambiq than building hardware, right? There’s there’s our software team, there’s our AI team, which is really a software team, there’s our field applications team, which is really mostly software engineers helping our customers. And the development cycle on that software is a lot quicker than the hardware, right? So that’s part of how we keep pace. but

You know, part of it too is that is is we’re aiming ahead of where our customers need to be. You know, they tell us where they need to be this year and next year and five years or, you know, maybe fuzzy idea of three to five years from now what they need. And so we’re attempting to to to kind of lead the receiver, so to speak, on that and have more than than they need. and there’s one thing we know is that is that, you know, software engineers and AI engineers will always fill however many cycles you’ve got available, however much memory you’ve got available, they’re gonna they’re gonna fill that up.

the the other thing I’ll say is I’m a big fan of general purpose architectures. So what I mean by that is you know y y you talked about neuromorphic earlier, you mentioned neuromorphic early. There’s a lot of specialized AI acceleration techniques that require custom hardware or custom network structures. I really want to avoid all that. I want to build a general purpose machine that is capable of of running multiply accumulate operations hyper-efficiently.

and you know what’s cool about our particular platform at Ambek, we we have this core platform we call SPOT, subthreshold power optimized technology. And without giving you a whole like circuits one on one lesson here, we’re we’re running transistors at extremely low voltage to get huge energy savings. And my tran that transistor doesn’t care if it’s running a general purpose CPU, an NPU, DSP,

Scott Hanson (08:46.731)

you know, paseban processor for a radio. And so, you know, in that way what I can build is I can build a great NPU or a great vector accelerator that just runs any network style that I want. So in that way we can keep pace a little bit with with what’s going on out there.

Pete Bernard (08:57.593)

Really.

Right, right. But don’t you think that things I mean, you know, in the long arc of computing history, like processors do get more specific, right? I mean we go from CPUs to GPUs to NPUs.

Scott Hanson (09:10.793)

They do, they do, and you know, if you look at if you look at what our processors have have gone through, the first processor that we built, it was called Apollo and and it had a microcontroller core ARM Cortex class core that could run at twenty-four megahertz and it had a few peripherals and it’s I mean it’s kind of laughably simple now that I look back at it. That’s only, you know, I don’t know, twelve years ago, right? our latest chip, the one that we’re at we’re actually taping out this year and and launching formally next year is called Atomic.

And it has a GPU, a CPU, it has an NPU, it has a DRAM in package, it has all kinds of other, you know, computing course, a cryptographic accelerator on board. So we’ve become quite a bit more specialized. So you I can I once once out of mount my mouth I can say, yeah, let’s keep it general, but yeah there’s there’s some, you know, specificity needed for sure.

Pete Bernard (09:54.168)

Right.

Pete Bernard (09:59.63)

Well, I mean, even at the you know, the the mighty NVIDIA and Jensen Huang, who’s never wrong about anything, you know, was all into GPUs for a while and then they bought Grok because they realized that, you know, doing sort of language acceleration was kind of a bit of a different beast, you know. So it feels like yeah, everything is as general as possible until it needs to be more specific, right? And then I guess that’s the the so you know, in defense of neur the neuromorphs, the you know, sometimes

When have very specialized applications, right, that, you know, really require orders of magnitude, more power efficiency. And it’s not it’s not really an open and generic platform. Like maybe there’s room for some more specific innovation for those specific applications. I don’t know. But

Scott Hanson (10:47.657)

Hey, I I I believe it. Yeah. No doubt. Hey, one one thing I will say though is I I never bet against Moore’s Law. So I know that, you know, on the very bleeding edge, we’re still Moore’s Law is slowing down and everyone’s always wondering, hey, are we gonna get are we going sub two nanometer? Are we gonna go sub one nanometer? But you know, the edge is living a couple generations back from that, right? we started at one eighty nanometers, our newest products are in twelve nanometers, and there’s so far to go from there. And so, you know, my point is that the the kind of conventional digital

Pete Bernard (10:50.52)

You know. but

Pete Bernard (11:09.154)

Mm. Mm.

Scott Hanson (11:17.579)

processor is just gonna get better and better and better and and cost per transistor, performance per transistor, power per transistor is all improving. So

Pete Bernard (11:24.898)

Right. True. And, you know, the software architectures are getting more efficient too, right? Like vector symbolic architecture, VSA, things like that. People are really like rethinking. You know, I I I like to use the term that I think we’re still in the steam engine era of AI, right? We’re still sort of brute forcing it. And and you know, there’s going to be some re architectures that happen that give us multiple order magnitude efficiencies because cur current course and speed.

Scott Hanson (11:29.738)

Absolutely.

Pete Bernard (11:50.722)

And I know we’re we’re kind of edge folks, so we don’t really bother with data centers, but the idea of building these kind of nuclear powered data centers that kind of suck up the water table to do this stuff, it’s not sustainable. It’s not, you know, there’s an asymptote there and maybe we’re we maybe hit the asymptote, I’m not sure. but there’s gotta be, you know, more than just kind of the brute force compute that’s currently happening. It’s like, well just build a build a f five football field data center. It’s like, well, that doesn’t really sound like the right way to go, but

Scott Hanson (12:19.381)

You know, we we we we talk publicly about our hardware having a two to five X power advantage compared to conventional approaches and and I I would say in practice it’s often five to ten X power advantage. What I’m telling the software teams, the the AI software team and and our other software developers is find a way to give us another five to ten X power advantage, right? Combine all that together and give us a a huge advantage. And I think it’s there. You know, we we get with customers and I’ll I

they’ll they’ll show us their network architecture that that’s running on this in some obfuscated way or anonymized way and and there’s almost always some way to dramatically speed things up. You know, reshape this layer or we’ll we’ll write a you know a custom kernel for this other layer here or hey this particular layer is not mapping efficiently the hardware and and and we’ll approach it in some other way. So there are almost always ways to dramatically improve the the execution at the software level as well.

Pete Bernard (13:12.918)

Mm-hmm. Yeah, for sure. For sure. So you mentioned about kind of shipping some stuff. Like the if we talk about the the reality of the scale that’s happening right now, you talk about hundreds of millions, and you know, that’s Ambiq and you can throw in, you know, within the foundation we have about 70 different companies doing stuff. And so, and that’s probably just a portion of the ecosystem. So, I mean, it’s really starting to scale out there. What what what do you see that’s really

Scott Hanson (13:25.545)

Yeah.

Pete Bernard (13:39.618)

triggering a lot of the the drive for scale these days, like and and what what’s still what are still points of friction?

Scott Hanson (13:48.955)

well what’s what’s driving scale? What I’ll start by saying is we we saw edge AI deployments going back to twenty sixteen, twenty seventeen on our products, right? We’ve got various generations of our Apollo processors. in in the last ten years, the the majority of the units have been used in you know wearable devices. at which which have morphed into more health tracking devices that that have a cloud-based agent that’s that’s monitoring everything you do and and giving you advice.

we saw the first use of of AI there and they’ve just only gotten more sophisticated, right? They they they have built systems now that might have in excess of five AI models running on your device simultaneously, and they and and they have ambitions of of going really, really far into the future. But what’s notable to me is that in the last couple of years we’ve seen whole basically every other customer we have is is talking about it or deploying AI.

just in our our last few earnings calls, we’ve disclosed that that in excess of eighty percent of our units shipped are running some kind of AI model. And that’s exciting, right? And it’s it’s not just the wearables. They are certainly the the most aggressive and earliest users, but we’re seeing everything from water leak detectors and smart home deployments to machine health monitors. That’s that’s probably one of my favorites is the the machine health monitor that’s doing anomaly detection. Right? You got a you know, a cans a

Coke can size sensor that’s monitoring vibrations and sound and and whatever and is looking for changes in that and says, Hey, I just saw a change in vibration. Better come out and fix me. I think I might be failing soon. So that’s a that’s a cool use case. all kinds of medical devices that are, you know, looking for falls, you know, off of somebody falling, looking at heart health. and and so yeah, a lot of really interesting deployments that are that are starting to happen. And, you know, that y yes about points of friction.

Pete Bernard (15:24.249)

No.

Scott Hanson (15:43.275)

On my side, the point of friction is just there’s so many diverse use cases that the AI team can’t possibly keep up with it. And so what we’re focusing on is the class of problems that we can uniquely solve. specifically, we like to say that we’re giving our customers picks and shovels so that they can go mine gold, right? I I’m not gonna be in the gold mining business, I’m in the picks and shovels business. So I’m gonna give you the best kernels out there, we’re gonna give you compilers and run times that that are hyper efficient.

we’re gonna give you what we call AI development kits, which is maybe includes some example AI models that we’ve pre-trained, but but has a lot of the other, you know, building blocks that that you need to to get going and and build your own model. So what what we’re seeing is a lot of the chip vendors out there are trying to get you to play in their proprietary sandbox, right? They they build some beautiful

service where you, you know, you you it’s auto ML, you push a button and out comes a a a model, right? And I think for a a certain set of customers, that’s a wonderful solution. But you know, we’re finding that real real AI engineers, real data scientists, real model developers, they want to use the flows that they’re familiar with. And so we you we want to give them kind of a modular system where we’re giving them picks and shovels. We say use them as you see fit, slot them into your into your flow and and get going.

Pete Bernard (17:05.25)

Yeah, I think if I think if companies are building something really unique, they need to sort of invest pretty much down to the metal if they’re really gonna do stuff. yeah, there’s always gonna be folks that, you know, yeah, sure, take some stuff and put a logo on it and ship it, but that’s usually not pretty cutting edge. Actually, speaking of interesting, I’m gonna ask you about your favorite deployments, but actually just as a little anecdote, I bought this thing called a droplet the other day. So have this house in Massachusetts and

It’s a pretty simple thing. It’s like just a little clamp. It clamps onto your water main. And, you know, basically based on vibration sensing, right, it’ll detect water flow through your house. And then, you know, you say how many faucets or have an outdoor shower there and a lot of other stuff. And then you just go around and you flush the toilet and you watch the meter thing. You tag that activity as toilet, you know, and or outdoor shower. Okay, that’s the outdoor shower. And then so I did that and you know, walked around the property and did that. And then now I get

Scott Hanson (17:48.139)

Mm-hmm.

Scott Hanson (17:54.377)

interesting. Interesting.

Pete Bernard (18:01.804)

remotely. So I just I just flew back last night from Boston. So now I can see who’s using what in my house from a water flow. Just for like v just the vibration patterns of the water main are detected and and basically if that’s you know.

Scott Hanson (18:15.613)

Incredible. Incredible. And you know what’s what’s exciting is that next level up where, you know, you’re you’re probably in the loop of that, but it could just as easily go to an agent, a cloud based agent, that’s sitting there and looking at it and and kind of telling you a story at the end of the day of, hey, it looks like pretty normal water usage today, or hey, you I noticed that there was some extraneous water usage. Let’s get somebody out there. Or I ordered a a a a repair person to come out and and take care of that.

Pete Bernard (18:36.264)

yeah.

Yeah, yeah, yeah. No, I mean if I I assume that these guys have like an MC C P server or something that I can so I use this other app called Little Bird, which is like a a Mac OS E you know Little Bird. So and I have that connected to a bunch of C P servers. And so every morning I get my daily briefing and I get all my C P data kind of, you know, with the natural language, like, Hey, here’s what’s going on. So definitely Little Bird could say, Hey, here’s what’s going on with your house in Massachusetts and not only the water, but you know, the energy and the

Scott Hanson (18:44.819)

There you go. Yeah. Mm-hmm.

Scott Hanson (19:01.512)

Yeah, yeah.

Pete Bernard (19:08.036)

electricity and the heat and you know, take the ring cam footage and be like blah, blah, blah. And yeah. So it we’re we’re getting there. So I I c I can definitely see like these days, I mean, these are all con consumer applications, but I guess to to your question to to kind of your business, are you seeing the volume more in the consumer space with like kind of the the wearables like you’re talking about? Or is are you seeing more in the commercial space or how’s it sort of translating?

Scott Hanson (19:34.165)

Well we we we certainly s saw the first deployments in the consumer space, in the consumer wearable space. And right now, I would say the most aggressive roadmaps are coming from that space as well. And I’ll I’ll specifically call out anything body worn. So whether it’s it’s a it’s a watch or you know, a band like the like this whoop band I’m wearing, or smart glasses, probably the most aggressive roadmaps we’re seeing. Yeah, there you go. The most aggressive roadmaps I’m seeing are in the smart glass space because

The minute you add a camera, you can do some really exciting things, right? So, and yeah, I of course I can have cool UI with with gestures and and all sorts of other things. But for me it’s more it’s more about I can tell who I was talking to, with the microphone. We can, you know, man or listen to what we’re talking about. I can take give me summaries, build action item lists, and and kind of a so the the VC

Pete Bernard (20:06.254)

Okay.

Pete Bernard (20:12.088)

Sure.

Scott Hanson (20:28.213)

Josh Wolfe recently talked about so-called life courting as his next big thing, life recording, right? And and I love it. I think, I do think that all wearables are moving toward that, right? And they’re all doing it in their own unique way. Some of them are coming at it from a purely health standpoint, some of it coming from a you know, kind of a I’m a smartwatch vendor, but all of them are are converging on this life courting concept of of we’re gonna we’re gonna keep track of the context of your life. It’s gonna be health data, it’s gonna be who you’re talking to, what you’re talking about.

Pete Bernard (20:49.941)

Mm-hmm.

Scott Hanson (20:57.755)

what what is is going on with your family, et cetera. And you know, we could get into a whole side conversation about the ethical implications of all that, but I for one am excited about about what we can do there.

Pete Bernard (21:01.026)

Mm.

Pete Bernard (21:08.162)

Yeah, no, I think it’s definitely possible. And, you know, just like using when I use Little Bird, it can act it accesses my emails and accesses my, you know, my Google Meets. And so it g it knows kind of what I was talking about and all that stuff. So I you get a little taste of that. I do agree that like especially getting into the kind of the wearables, hearables, and plantables space. then you you know, there’s some societal norms that are gonna have to bend a little bit in terms of recording. Like even my phone, like I could plop it down.

Scott Hanson (21:32.757)

Mm-hmm, mm-hmm.

Pete Bernard (21:37.859)

You know, I meet people at the coffee shop here, place called Soul Food Coffee House. If you’re ever in Bellevue, happy to take you there. And you know, always have my meetings there, try to get out once a day. and you know, I can plot my phone down there and it can record the conversation just like a Google meet and it can give me the summary. But it’s a little bit odd, you know. It’s like I’m recording this conversation, you know. So, like at some point we will as a society just sort of default that everybody is kind of recording everything that’s going on and

Scott Hanson (21:42.795)

There you go, okay.

Pete Bernard (22:07.682)

You know, that’s always been sort of a a hot button issue. I mean, we still have people that don’t like the f you know, automatic checkout at the supermarket. So, you know, there have there’s issues with that. So this this is going to be that’s probably the it’s not a technical hurdle there as much as the life cording will be more of probably a societal thing. Unless it’s like your personal recording, which I do, I record all my vitals and all my stuff. But once you start recording in public and who you’re interacting with and stuff, I don’t know. Is it

Scott Hanson (22:15.659)

Sure. Sure.

Pete Bernard (22:35.02)

Someone I I know the EU will pass some legislation that says you have to have a big red light on and the thin you’re the center of your glasses or something.

Scott Hanson (22:42.343)

And you know, what’s w from a from a technological standpoint, we’re basically there, or at least from my perspective, if if if we don’t have all the technologies today to to build these devices, we will in the next year or two. And so I feel good that at least the technology side of his done. It’s it’s the it’s all that other stuff. It’s the societal embrace of the stuff, it’s it’s what government regulations are gonna come down on top of it. and and frankly, I I do think some of that needs to be there, right? these are dangerous technologies if if deployed incorrectly and y you know, companies

Pete Bernard (22:58.626)

Yeah. Yeah.

Pete Bernard (23:10.244)

Surreptitiously.

Scott Hanson (23:11.391)

Companies, I I know this well as a as a now publicly traded company. Our our obligation is to deliver returns for our shareholders, right? And and as a founder, I feel obligated to be ethical and and lawful about everything that we do. But y you know, I don’t sometimes there’s a there’s a need for outside bodies to to regulate these things and make sure that we’re doing things that are right for everybody, right?

Pete Bernard (23:33.411)

Little bit. Some guardrails. Well, you know, funny s not a funny story, but interesting anecdote. You know, when camera phones were first introduced, they’re actually invented by Philippe Khan from Borland. You remember Turbo Pascal? So but he invented the camera phone. Long story. In Japan, caught on first, right? Maybe everyone was taking pictures of their food like twenty years ago and we all thought that was kind of kind of crazy. And we all do that, but

Scott Hanson (23:54.475)

Mm-hmm, mm-hmm.

Pete Bernard (23:57.699)

The the the Japanese government introduced legislation that you had to make a camera sound, a little click click or whatever. That sound had to be played whenever you took a picture with a camera phone. And that was enforced by Dokomo and their handsets and stuff like that. And that was just so that no you couldn’t take pictures surreptitiously. Right. which, you know, obviously you can imagine some bad situations there. So so that was like an early thing where it’s like, Hey, you need to let people know I’m taking a picture ’cause you hear that little sound.

Scott Hanson (24:08.469)

Mm-hmm.

Interesting.

Scott Hanson (24:20.841)

Yeah, of course.

Pete Bernard (24:27.598)

So I wonder, yeah, as we sort of get into this life courting thing, you know, what what what we’ll have to do there to, you know, hold up a little sign. I am recording. Walk around. But we’ll I’m sure we’ll see that in the Bay Area anytime soon. so yeah, so you got a lot of cool things being deployed. So I guess one of the questions would be so what happens next? Like we’ve talked about the lot, you know, we’ve got these haribles, wearables, implantables.

You know, the the the buzzwords, I think we were talking about before we hit the record button, you know, obviously there’s agentic AI and there’s, you know, generative AI and language models and and embodied AI and analog AI and physical AI, of course. And like how do you see the evolution of let’s say call it machine learning that’s running right now in some of these embedded wearable systems to something more profound? Like where where does that start fit in?

Scott Hanson (25:27.327)

So I I think l let me start by saying agentic AI is the vision in the in the sense that there will be agents that are managing the collection of of all kinds of data and deciding what’s relevant and making decisions based on that. And and it will give there will be an illusion that that agent lives in these devices on our bodies, in the the devices in in our building, in our homes and in our factory floors.

and it’s kind of it’s it’s not as important where the agent actually does reside because we have connectivity everywhere now, right? So I it it takes almost no time to get to the cloud.

Pete Bernard (26:01.689)

Well, we sort of have connectivity everywhere. I mean, you know, not exactly.

Scott Hanson (26:06.737)

Not everywhere, not everywhere. Okay. But but the the there I I think that what we’ll see is that there are there are cases where it’s necessary to have that that agent on the edge, but I think for the most part where you’re gonna find is a system similar to what actually a lot of our customers have today. So both Woop and Garmin have this. If you look at all the leading wearables vendors at at Aura or Google or whoever, they’ve they’ve got wearables that are collecting data, collecting all kinds of vitals, and that wearable is deciding what’s interesting, right?

They’re identifying arrhythmias, the the wearable is identifying that you I I went to a CrossFit class at noon today, and then it’s it’s sending that macro data, let’s say, up to the cloud, and there is an agent or there is an LLM with an agent wrapped around it taking in all this data and and and you know asking questions. I’ve actually noticed that my devices have started asking me questions. Hey, I’ve noticed that you’ve consistently got a workout happening at noon. What are you what are you going for? And if I say, I’m training for

a marathon in October, it’s gonna go, okay, here’s a marathon training plan that I think you can use, right? This is happening. And what was interesting to me is I went to a smart home conference a couple weeks ago and they were having all the same discussions in that space about agent plus LOM in the cloud, monitoring your whole smart home. You’ve got this collection of heterogeneous devices that are all feeding data. It’s your your your smart door lock, it’s your

your security sensors all over the home. It’s your your thermostat and your smoke alarms and your you know smart fridge and everything delivering data up to the cloud and then and then getting a home level view of what’s going on. It’s you know everything from what what’s the health of the people in your home to to how secure is your home, et cetera. So it I think we’re converging on agents being and and LLMs being present everywhere. agents and LLMs in the cloud, they they can’t see or hear, right? They need senses.

Pete Bernard (27:54.127)

Hmm.

Scott Hanson (27:59.709)

And so the the us the the people vendors out at the edge like us are are you know making that happen. I’m I’m giving giving them eyes and ears so that they can see what’s going on in the real world. So it’s it’s almost less important to me where that stuff where where the agent resides, it’s more important to me is is what is the system high level architecture and and I one thing I know for certain is that that those edge devices, they are running AI because you can’t just take tons and tons of raw data and send it all to the cloud. You know, speaking of your connectivity problem earlier.

Pete Bernard (28:09.614)

Mm.

Pete Bernard (28:26.467)

No. Right. Mm.

Scott Hanson (28:30.121)

I certainly can’t, you know, jam a a camera a multi camera feed, microphone multimicrophone feed, biosignal feed all up to a cloud and then have that be true for millions and millions of users. So a AI will certainly be everywhere and then that we could sit here and debate about you know where where do the LLMs lie? Do they become SLMs that are at the edge, or is there an agent somewhere along?

Pete Bernard (28:40.013)

No. Yeah, I agree.

Pete Bernard (28:50.179)

Yeah, I mean it’s interesting you mentioned that because like we started every six months we do a kind of multi-day generative edge AI live stream. It’s like 10 to 15 talks, tech talks, just back to back. A lot of academia, a lot of engineering folks. We’ve done that. We just had our sixth one, I think. And so we’ve been doing it for a few years. And it’s really interesting to see the evolution of what is state of the art in generative edge AI, basically, every six months, you know, for a few years. And

Scott Hanson (29:16.203)

Mm-hmm. Mm-hmm. Sure.

Pete Bernard (29:19.933)

And you know, I mean, as you know, running multi billion parameter SLMs on a phone or an edge device is not that big of a deal anymore. Like it’s very derigo. And so whereas, you know, a couple of years ago it was like, my God, can you actually run a really slow chat bot on a Raspberry Pi? You know, so we’ve come a long way even in that period going back to our Moore’s Law conversation. So I definitely think there’s a gravitational pull that will move these things as close to where the data is created as possible.

And I wouldn’t be surprised to see sensors that have their own MCP servers in them or MCP interfaces that could be queried by agents that could be running. That agent could be running on your glasses or could be running on your phone. and maybe as a last resort, it gets to the cloud. I mean, one of the things that we’ve seen too is that, you know, running tokens on the clouds can kind of add up. So a lot of these companies that are like, We’re gonna have AI for this and AI for that, all of a sudden their cost of good, their margin goes to like zero.

Scott Hanson (29:55.657)

I believe it, yeah.

Pete Bernard (30:19.257)

because they’re just just, you know, writing huge checks to Microsoft and Amazon. and and so it’s like, well how you know, how do you move it out?

Scott Hanson (30:25.739)

We we had a you know, it’s funny, we w we we spent a a lot of years pitching the value of edge AI, right? And it’s latency and and there’s power benefits and all sorts of things. But w yeah, the the the cost one actually came to the forefront recently. We had forefront recently. We had a customer who was resistant to to doing AI on the edge and then all of a sudden they came to us and said, Okay, we gotta move it all out there and we said, Well what changed? And it was exactly that. It was their AWS bill was through the roof. Yeah. And I get it too, right? It’s it’s amazing how quickly those AWS bills start adding up.

Pete Bernard (30:48.909)

They got the bill. Yeah.

Pete Bernard (30:56.163)

Yeah, no, it’s interesting. And AWS is one of our strategic partners, by the way. So shout out to AWS. Love those guys. Wonderful. But yes, you know, the OpEx can really start to add up. And, you know, use always tell people like the your AI project always starts in the cloud. And then, you know, it gets optimized and commercialized and migrated to kind of a lowest cost, you know, operational cost to to make money on it. And that’s when it tends to kind of get jump the data center into the real world. but yeah, I think.

Scott Hanson (31:01.203)

It’s a wonderful service, Very important.

Pete Bernard (31:26.467)

You know, getting these language models and agentic AI running and edge equipment is you know, kind of happening. I mean, we we have a lot of interest in physical AI. I was curious, like Ambiq, like how how have you been involved in the physical AI space? Is you have humanoid robots walking around there? What’s going on?

Scott Hanson (31:45.683)

Yeah, they’re all walking around just outside my door here. no, it’s a it’s a really important market that we see as a midterm opportunity and and there’s two pieces of it that are really interesting. So you know you break down these humanoid robots and there are tens of motors that that have to run and all those motors need motor control and and the more complex the control, the more complex the algorithm, the the more need for compute and and we’re seeing

Pete Bernard (31:46.821)

Just outside of Kingdom.

Scott Hanson (32:10.879)

I’ve seen a lot of academic work talking about the use of of AI, the use of of AI models to assist in in motor control. So that’s one really important place. But then the sensing too. And and and there’s there’s you know centralized sensing where you’ve got cameras, but but then there’s distributed sensing where we’ve got pressure sensors and and all sorts of other things. And and the Ambek can play a role in in all that and will play a role. And the way we’re getting there is that we’re starting with with mastery of you know devices like like these and and smart home devices and certain industrial devices.

Pete Bernard (32:18.062)

Yeah.

Pete Bernard (32:37.604)

Mm-hmm.

Scott Hanson (32:40.009)

And all of the the the AI software and AI hardware that we’ve developed there are transferable to the to the humanoid and and broadly physical AI opportunities. I think I think it’s gonna take some time, frankly, for that market to fully emerge, right? There’s a lot of big technical problems that need to be solved. energy, by the way, being a huge one of them. you know, early on in Ambiq history, we we consciously avoided

Pete Bernard (32:52.901)

Yeah, that’s right.

Scott Hanson (33:05.685)

places like drones. and not because it wasn’t a good use of of our technology, but the motors were such a a dominant source of energy usage in these systems that it was like, let’s we gotta solve this problem, right? But then, you know, if you look at the latest DJI drones, these things are are ultra complex, right? They’ve got these these super complex sensory systems that that can do you know all sorts of interesting things. And sudden so suddenly the electronics have become a a a appreciable portion of the overall energy budget.

Pete Bernard (33:13.701)

Mm-hmm.

That’s right.

Scott Hanson (33:34.953)

In fact, I I know some people who were founding a garbage collecting robot company and I was talking to them about about where the energy was used in their system and even for something as big as a garbage collection robot that’s wandering around parks and picking up and compressing garbage, the the the processor stack that’s running that’s running all the AI was actually consuming more energy than than the motors themselves. So that I thought was pretty remarkable and

Pete Bernard (34:02.243)

Yeah.

Scott Hanson (34:02.396)

represents a really big opportunity for Ambiq. So we will we will be there. It’s just a matter of time.

Pete Bernard (34:04.644)

Yeah.

Okay. Yeah, I think it’s like physical AI, that’s kind of the un untold story right now is the energy consumption and the battery usage. And you know, you can’t just and again, love NVIDIA, but you know, throwing a jets in everywhere and seeing what happens, like it’ll run for like twenty minutes. so so yeah, I think that’s huge in terms of bringing power efficiency to physical AI. We have a company that does actually specific like

solutions for drone motor control and optimizing that. And they claim to get like a 50% improvement in flight time just by optimizing the the motors for the drones using AI. So yeah there’s a lot of opportunity there and last thing you want is a humanoid robot with a dead battery just kind of sitting there on the sidewalk and it’s like that’s recipe for disaster. I was gonna before we run out of time we still have some time but

Scott Hanson (34:43.369)

No kidding. Wow.

Scott Hanson (34:56.555)

Absolutely.

Pete Bernard (35:02.957)

I know in a previous conversation we had talked a lot about career development. And, you know, one of the things that we’re invested in as as an educational nonprofit and our scholarship fund is career development, because at the end of the day, you need smart folks to come in and build this stuff, like you said. and so what is is Ambiq been involved in career development and university engagement? I mean, you mentioned, you know, Michigan and stuff like that, but is that

Scott Hanson (35:29.311)

Yeah.

Pete Bernard (35:30.169)

Part of your thing or is there an Ambiq thing on that? Or how do you guys approach career development and sort of investing in the next generation?

Scott Hanson (35:36.895)

Well, you know, as an first I’ll say as a newly public company, it’s it’s something we’re figuring out as we speak, which is exact exactly why you and I are talking, right? you know, when I first built the company, and I think this is true for every entrepreneur, you’re worried about next week. You know, will will we live to next week or next month or next quarter or whatever it is. and so it’s hard to think too much about the future. And and I w the one thing I wish I could go back and tell myself ten years ago is hey,

The company’s gonna be here in ten years, so plan accordingly, right? That means invest in in in young engineers, bring bring them up, teach them what they need to know. and we’re doing that now, right? Because because we will be here ten years from now and twenty years from now and thirty years from now. So back to your question, yeah, we’re we’re engaged with a couple different universities both here and and elsewhere. So in in Singapore, for example, we’ve we’ve recruited a number of people out of the local universities there.

we have an office there. we’ve got headquarters here in Austin, Texas, and have been pulling actually we we just had a I don’t know, ten or fifteen interns that are all starting right now from various universities. We do have a University of Michigan intern, which is which is wonderful for me. We’ve got some UT representation as well here. So any of those longhorns out there listening, please please send your resume. And so yeah, I we’re we’re looking at universities as a a variet a variety of universities as good recruiting grounds for us.

Pete Bernard (36:48.836)

There you go.

Worse.

Yeah.

Pete Bernard (37:03.043)

Yeah. Hmm.

Scott Hanson (37:03.343)

and look in a in a world where we we’ve been talking about AI at at our end users you know our end users devices, but we’re using AI within the company as well and these younger engineers are are quite a bit more fluent in it than than the ones that that have been with us for a while. And so I like having fresh talent that can come in with new ideas and you know, who sits down and and they’re just you know, coding agent native in the way they are fluent in in the way they work, right? So

you know, I I had a a a neighbor the other day asking about whether her son, her high school son, should go take a a coding class. And I said, Well, I think it’s unlikely actually that your son will ever write code in the way that any of us has has written code in the past, but he should go take it, he should learn it, and then he should learn how to tell use an AI agent to to to do exactly that task, right? So I think

I I think it’s it’s great. It I hope universities are teaching this. We do sense that the younger engineers are all quite a bit more fluent and ready to to dive in on

Pete Bernard (38:07.609)

Yeah. Yeah. No, that’s interesting that you say that. I mean, I was actually looking at the data recently because we we actually had a live stream from UT San Antonio this morning, actually on our YouTube channel, talking about neuromorphic computing, speaking of longhorns. but it was interesting to me, and I we just we just went through commencement weekend with our daughter and everyone’s, you know, about trying to get jobs and stuff. The the actually in the US, the unemployment rate for recent college graduates, g can you guess what the unemployment rate is today for

Scott Hanson (38:16.52)

cool.

Pete Bernard (38:37.263)

recent college grads with thirty percent, forty percent, what do you think the number is?

Scott Hanson (38:37.801)

I don’t want to. I don’t even know. It’s probably high.

Scott Hanson (38:43.209)

I d are you suggesting it’s in the thirty, forty percent range? I have no idea. No idea. No no guess.

Pete Bernard (38:45.305)

No, I’m just asking, do you do you have a guess? Okay. So it’s o actually it’s only five point five point eight percent. Yeah. No, I’m sorry, five point six percent. I correct myself. Five point six percent, which is not too bad since the the overall unemployment rate is three point nine percent. Last year was actually five point eight percent. So it’s actually gone down the unemployment rate for college res since one year to the next. So it’s really interesting when I see that data, and then you see a lot of the kind of catastrophizing about

Scott Hanson (38:52.169)

No kidding. that’s much better than I thought it would be. That’s interesting. Wow.

Scott Hanson (39:00.479)

No.

Scott Hanson (39:10.132)

Interesting.

Pete Bernard (39:14.605)

AI is taking everyone’s jobs and stuff like that. But like you said, I think people are using AI now natively in how they work, right? which is different than using AI instead of work. Right. and I think that that’s really kind of where the future is. And I think advice for everybody out there, like you said, is yeah, learn how the stuff works, but then learn how you’re gonna use AI to help, you know, do do that stuff as well. And I think that’s the right combination. But

Scott Hanson (39:40.263)

Exactly. Yeah. Yeah. Well, I what I what I should say too is that I’ve seen that the bigger companies I can’t speak for these bigger companies who are downsizing because of AI. I’m trying to upsize. I’m trying to use AI to make myself look bigger. I want to look like instead of a two hundred and twenty person company, like a four hundred and forty person company, right? And then after that go bigger, right? So I mean for the moment it it’s a fantastic opportunity for us to effectively grow our workforce.

Pete Bernard (39:43.79)

Yeah, cool.

Pete Bernard (39:56.281)

Yeah. Sure.

Pete Bernard (40:05.091)

Yeah, no, for sure. And I would say the same here. I mean, we’re a small AI startup ourselves and we have a worldwide footprint. And the only reason we can do that, and we just actually passed a hundred thousand subscribers on YouTube, is because we are using AI platforms to make ourselves look a lot bigger. and and I think that’s the cup the companies that will be successful, the ones that use AI to, like you said, grow faster and get bigger, as opposed to, hey, I’m gonna save some money by laying off some old people and

Scott Hanson (40:17.146)

cool.

Pete Bernard (40:34.767)

Getting AI agents. That doesn’t really work. Cool. Well, it sounds like you got your hands full over there with edge AI. So that’s exciting. And any sort of parting words of wisdom around maybe let’s let’s turn it into something call to action. What’s the call to action for those that are listening and are saying, like, I this is exciting. This is the future. I want to go do more of this. Like, what what should they go do?

Scott Hanson (40:42.899)

Indeed. Indeed.

Scott Hanson (40:57.559)

Yeah, I I’ll say I’ll say t yeah, two two things. One is I think building an edge AI model these days is a lot easier than you think, right? and that I I got clued into that years ago and we started hiring our first AI engineers. It felt like something where I need to hire a a team of PhDs to come build these models and train them and and so forth. I it’s never been easier than than it is now. and it will only get easier from here to to build these models. So I’d say dive in.

there’s all kinds of of great tools available for that. I’m sure they can come to you guys to to to learn what all the latest and greatest tools look like. I mean I know that we’ve got a a number of engineers who just kind of sit there with their desktop gaming PCs and their training models left and right, and it’s and it’s great. So I don’t be afraid of of taking that on. Now, the second bit of of advice I’ll give is do be aware of the resource implications, right? we worry about memory, we worry about

Pete Bernard (41:37.805)

Yeah.

Scott Hanson (41:51.199)

Performance, we worry about power. You gotta pay attention to all those things. and what we’re finding at you know at the biggest of companies is there’s one group that that builds and trains the models, and there’s a whole other group that deploys the models on the hardware and and they may not be talking as well as as you think, right? and so if you’re an AI model, if you’re a data scientist, you’re an AI model developer, whatever, just be aware of the of the underlying resource usage and and

Pete Bernard (42:06.788)

Right.

Pete Bernard (42:16.441)

Mm-hmm.

Scott Hanson (42:17.033)

y you if you want help in that, come talk to Ambiq, right? We’re thinking a lot about how to give our customers tools to do that. it’s it’s a hard thing and that it’s but at the same time it’s fun for us, right? I I think that’s about as fun as it gets.

Pete Bernard (42:26.083)

Yeah. Makes sense. Well, yeah, efficiency is the new currency. That’s what they say. So cool. Well, Scott, thanks again. Thanks for joining us. Hopefully everyone here learned a little something and we’ll talk soon.

Scott Hanson (42:31.627)

You got it.

Scott Hanson (42:39.893)

Thanks, Pete. Appreciate