The Biggest AI Deployment Nobody Talks About | Samsara CEO Sanjit Biswas

Yeah, absolutely. Well, first, Matt, thanks for having me on your show. I would say physical AI is really the application of AI to the physical world. So if you think about the infrastructure of our planet, it's way more than just the roads where you might see a Waymo robotaxi. It's the construction sites. It's the electrical grid. It's really kind of like all the plumbing that's under the street. It's everything out there. The interesting challenge, I think, with physical AI is that it's not digitized, right?

This is a frontier where you don't have decades of bits that you can reason over and tokenize and quickly ingest. And that makes it really fascinating because the amount of value that's trapped is really significant. So we think about it in a few different ways. We think about digitizing things like location from GPS tracking, of course. We think about using cameras as sensors. So you can use that to understand the physical world in a pretty rich way, especially when you ingest lots and lots of basically video footage.

If you think about how many frames you get from that, it's really significant. And then there's a lot of other kind of data sources, whether it's like weather sources of what happened with precipitation on all the roads. You can think about speed limit data. There's lots and lots of physical aspects of the world that, when you put them together, when you fuse them together, there's a huge value unlocked.

Well, it's interesting you mentioned centuries because that is like the right timescale to think about physical infrastructure, right? All the way back to, like, the Roman era, there's pretty significant infrastructure out there. So many of these processes of, like, how do you maintain a roadway, have been in place for a very, very long time. A lot of that process was manual, right? Like, let's go inspect the condition of the road. Let's understand when it was last worked on. Let's kind of dig it up and see what we find.

If you think about the ability to digitize that and then use sensor data, it's a huge unlock. So the question becomes, how do you get the data? Then how do you process it? And then how do you come up with a meaningful insight or really an action? Like, what should we do about it? And that's, I think, now possible. The last, call it, two decades was around reporting. Like, how do we ingest the data and give you a really cool table so you can look at it and reason about it, figure it out?

Now, what's awesome, really, in the last two, three years, is that AI is able to reason about this kind of information. They can look for other context clues and then give you the insight. And now we're actually seeing agentic AI, of course, which can take an action for you, maybe schedule the work to be done or start performing some of the work itself.

I think it makes complete sense where all of this AI wave has started, which is in the digital world. We had all the bits. We had petabytes of data to reason over. And there's a great training set. So think about all the trillions of tokens that were needed to get these models bootstrapped. The physical world is much messier. And there's also a hardware component to it. And there's kind of this saying: hardware is hard, right? This stuff has to hold up in the environment.

It's got to relay the data over unreliable networks. It has to be deployed to the front lines. And that requires a lot of messy work: physical hardware installations, getting millions of frontline workers to adopt new technologies, integrate it into their day-to-day work. And it's basically not as much low-hanging fruit as what we've seen in the digital world. All that being said, it's a massive part of the global economy. All these industries, they make up about 40%, 50% of world GDP.

And so it's an area where you can have tremendous impact, but you have to really roll up your sleeves and get much more involved than connecting to a large database that may have already existed.

Absolutely. So, of course, you're dealing with human life in a lot of physical operations, and that's an area for tremendous impact. So if you can build safety systems that keep workers safe, that's a great thing. But you also have to be careful that you don't somehow introduce risk into the picture. There are other sides of that too, which is these are digital technologies. So we want to make sure they're hardened from a cyber perspective, like they're not introducing cybersecurity risk.

But really, practically, the physical world is a pretty dangerous place. Think about a construction job site. There's a lot of earth-moving equipment, multi-ton, really dangerous stuff. It's kind of low visibility. And so the operators that are operating that equipment are taking some risk. The people on the site are taking a lot of risk. So it's inherently a risky environment. And our question has been, can we find ways to make it less risky using data? So we see that the risk is the opportunity as opposed to the challenge.

So Samsara is a technology company serving the world of physical operations. So think about those construction companies, the energy utilities, the supply chain and logistics companies that power the planet. We help digitize their operation. So that's a combination of hardware. So think GPS trackers, dash cameras, asset trackers, all kinds of different devices, cloud services to ingest all the data, and then now AI and applications to really close the loop, right? To help people take some kind of action.

Mm-hmm. Or ideally automate the action that's needed. What we've found is it's helpful to start with just tangible, real-world problems and then expand over time. So where we started was around fleets of vehicles, almost all of these industries that have tens of thousands of vehicles that they need to perform their work. But over time, we've expanded now into those frontline operations. And we're able to fuse all this data together from different sources on our platform, third-party sources, and unlock tremendous amounts of value for the customer.

That's right.

That's correct.

Okay.

Yeah, so on the data points side of things, these are numbers that feel abstract even to me, and I live them every day. But we're talking about 25 trillion data points—GPS, video, third-party API integrations, all kinds of data flowing into the system. Millions and millions of vehicles, for example. We're talking about millions of frontline workers that are using our apps every day. And in terms of impact, that's the other set of data points we look at, which is, well, how is all this technology having impact in the world?

We believe we helped prevent about 380,000 car crashes, road accidents, in the last year. That's meaningful to us because, as engineers and product builders, we are able to have significant impact in the world this way. We've helped avoid the emission of billions of pounds of CO2 by helping do things like optimize routes and reduce engine idling—things that seem simple, technologically simple perhaps, but the execution matters a lot. What's cool is you see that real-world impact.

You got it. Yeah, so there are so many different factors that produce risk. And we're very excited about autonomy and robotaxis and everything that we're seeing on the frontiers. But there are a lot of these industries, like in heavy-duty trucking or construction, where people work very long shifts. Maybe they've been out in the field for 10, 12 hours. It's been hot, and so they're exhausted. So fatigue is definitely one of them. You also tend to see more accidents in general at night because the roads are less visible.

You see accidents in foggy conditions, after it snows or rains, things like that. So we're able to help prevent a lot of risk by warning the driver when we see the risk increasing. We can provide some real-time feedback, and that helps them be much more alert, much more aware. And we can also coach away some of the bad habits that people develop. This is an interesting stat, but in the U.S., approximately 10% of people don't regularly wear their seatbelt.

And that varies by state. It varies by industry. But that's one of the biggest things you can do to improve your risk outcome, is just simply put on the seatbelt.

And it makes sense because sometimes people are doing a quick trip, or they're distracted. But that little reminder helps save lives, right? So that's one simple one. Putting down your mobile phone is the other one, right? When you take a look at your mobile phone—lots of people have this habit—your car, if you're driving, can move the length of a football field. And that is hard to think because you're like, "Oh, I'm just taking a quick look to see what that message was about."

But then you look up, and you've moved 100 yards, right? That's the kind of risk avoidance that we can create with real-time alerting.

That's correct. Actually, during our PhD. So my co-founder, John, and I met at MIT as PhD students over 20 years ago. I now caveat it because we're just old. But it was a fun sort of research project that we worked on. This was around the time that Wi-Fi was emerging as a new technology. We built a research project called Roofnet. So we covered essentially the city of Cambridge, the area between MIT and Harvard, with free Wi-Fi in the early 2000s.

So that was really exciting. It's like a hands-on, kind of very practical research project. We did a bunch of academic research on routing protocols and how to build the network. But the first company, Meraki, came out of that project, which was—we thought it was tremendously cool, this idea that Wi-Fi could connect so many people, just incredibly useful. We wanted to help other people build big networks. And so we essentially took that research, and now I would use the word distill, like we condensed it down to run in a box that other people could build networks out of.

And then we started essentially making that product available. So that was Meraki. To be honest, we kind of thought of it as a project. We weren't even thinking of it as a company. We kind of bootstrapped the business in Boston. We ended up moving to California. And it was fascinating because this is 2006, like 20 years ago. Wi-Fi was a brand-new kind of nascent technology. And there were some real challenges, right? How do you do guest access?

How do you do networks at scale? How do you deal with people starting to use YouTube, which was brand-new back then? Hard to imagine, right? But that kind of exposed us to how fun it was to solve real problems. And we had a huge, deep background in networking. We had a lot of friends from grad school that we recruited to start that company. And so we got off the ground quickly. We started seeing these devices get out in the world.

And Meraki ended up kind of just growing and growing and growing. It was doubling in revenue every year. So that was the beginning of our entrepreneurial journey. It was a little bit of an accidental start.

Yeah, you're very right. We had never spent time in a loading dock or a warehouse or in a construction yard, but we were always fascinated by them. And I think that was really the key. It's just nerdy curiosity of, well, how does the electrical grid really work?

I have an electrical engineering background. I was just always kind of fascinated by this. Or supply chain, if you're just curious about, well, that Amazon package, how far did it travel? Where were the goods stored? All of those kinds of questions were fascinating to us. And similar to Meraki, we weren't intending to start this company right out of Cisco. We'd been kind of on this pretty intense run, but the curiosity kind of got the better of us, and we started reading lots of books about this and just trying to learn about the world.

The challenge with physical operations, though, is you can't learn about it in a book. You actually need to go on-site. And to do that, you need a reason. You need an excuse, essentially. And so we said, well, maybe we can be helpful to these industries, right? Because, like I was saying earlier, the infrastructure of our planet is so massive, there have to be interesting problems to solve there. So we kind of started this company market-first.

Very steep learning curve. And I have to say, as a second-time-through entrepreneur, I'm really glad we had that experience because had we gone back into IT, we probably would've overweighted our prior experience and said, hey, this is how it's done, or this is how we did it at Meraki. With Samsara, it's a different customer. We serve the world of operations much more than the kind of technical buyer. We sell direct, so we interact directly with our customers versus via channel.

And that reset was enough for us to kind of go back to beginner's mind, which I think is also very important for most companies.

Yeah. And I would say every single one of those layers has some connective tissue attached to it. So if you think about the hardware, I've got hardware on my desk, of course. And so this would be an example of one of our—

This is what we call an asset tag. So you could put this on a piece of construction equipment, right? It's got an accelerometer in there, so it can tell how much it's been moving. It's got a Bluetooth radio that's a little bit more powerful than what you've probably used on the consumer side. So your AirPods have Bluetooth. This is industrial-grade Bluetooth. It's got a battery inside. And then it's built to be super tough. So you can beat this thing up.

You can drive over it with a truck. It'll continue to operate. That layer has hardware, but also has firmware that's running on it. It's got network connectivity. I mentioned Bluetooth. So this connects to the millions of Samsara gateways, tens of millions of phones and handsets that can act as kind of a relay point for us. And then we're able to get that to the cloud in a secure way. So that's the data capture side, going from motion, like the accelerometer, into the Bluetooth layer, into the cloud.

But from there, you need to organize it because you've got signals coming from all over. You need to be able to operate on it in a pretty methodical way. And that's what's going to feed the AI, because if you give the AI pretty noisy data, you'll get less signal-to-noise ratio. So we need to get clean data in. And then, to use your analogy, that's the brain, right? That's where we store it, we operate on it. You can surface those insights to the end user, or the agentic piece is you can just take an action, right?

Maybe change a safety setting, right? Like, say, hey, we're going to ask our entire fleet in New York, because it's raining, to increase the following distance versus on a bright, sunny day, right? That kind of change would have normally required a human in the loop. We're now finding that the AI can do it very consistently and can do it at scale that people wouldn't be able to get to because you'd need someone just sitting there monitoring all the settings for thousands of vehicles.

Not very practical, so it just doesn't get done. And that's maybe the arms, the action side of things.

Yes, this is Bluetooth Low Energy, if you're familiar with the BLE that you would see on your fitness device or your AirPods, that kind of thing. So Bluetooth has come a long way over the years. It's kind of gotten added onto, and it's picked up a lot of the great characteristics that many of these other standards had. So we can get a lot of range out of these trackers. And then we add a layer on top of that of security. How do we make sure that we preserve the privacy and the security of the tag that it's being applied to?

This specific one would last about three years. We have others that last six-plus years. We even have a really small form factor one. I've got these on my desk as well. And I don't know if you can see them. This is like a tracking label. So we're talking about a sticker.

Yeah, exactly. So these last about 45 days or so, so long enough for shipments to kind of go one way. And then these are disposable. So they don't have lithium-ion batteries in them, for example. So you can just peel them, stick them, track them, and then dispose of them.

Sure, yeah. This is a stack.

What's in it? It's hard to make out on camera, but there's basically some batteries, and then, of course, the Bluetooth chip, right, is running our firmware. And that is how it beacons up, essentially, its signal of where it is.

I would say for us, it's systems innovation. We did not build the battery. We don't make the silicon or the chip. But we work with partners to integrate all this together. And then we have the network, which is essentially—think about the millions of vehicles on the road, all the people running the Samsara stack. They form a community and relay signals for each other. And you've actually probably seen this on the consumer side with the Apple AirTag, kind of that concept of an ecosystem.

We applied basically the industrial-strength version of that.

Yeah, unfortunately, I don't have all the hardware that we make on my desk. But the vehicle gateway, think of it as a black box that goes on a truck or a piece of construction equipment or basically any kind of moving asset. That is a different type of collector, right? So it collects diagnostic information from the engine computers, and that's everything from how much fuel is it consuming to does it have fault codes to was the driver's foot on the accelerator or the brake?

That's all there on the diagnostic port. So we're able to ingest that information.

Yes. Well, we have to collect it and organize it, but it forms a long time series of many, many different signals. And even something that sounds as simple as a fault code, really, there's a lot of depth and richness to that too. Because if you read the fault codes very carefully, you can understand very specific dynamics about different kinds of engines and fuel types, air pressures, and so on. So we take all of that in. We have—

That's right. So I think you're familiar with dash cams. You've seen them in Ubers, right? Super valuable, useful for drivers because if something happens on the road, you can exonerate yourself very quickly. So we have a connected version of that. It records HD video, it has some storage, it's got the ability to send that to the cloud. And we run AI models at the edge. So we can do things like provide you feedback to increase your following distance, like I said, based on the weather condition.

Now, we're also seeing the driver side of that camera. So it's outward- and inward-facing. The driver side of the camera can do things like detect fatigue or mobile phone usage, provide real-time feedback to the driver. And the idea is they can self-correct, self-coach. And that is the aha: it breaks the cycle, or the bad feedback loop, of, "I'm going to look at my phone." If you get a sort of audio alert in the moment, and it happens many, many times, you tend to break the habit because it's sort of negative reinforcement.

That was a huge breakthrough for us six, seven years ago as we introduced AI at the edge. And now we're kind of going further with that concept, finding other forms of risk, and so on.

Exactly, because now that you've got the technology in the cab, what else could you do with it? Could you give the driver a briefing in the morning as they start their shift about where all they're going to go, traffic conditions, and weather? We have a little button. They can use that to call dispatch, for example, and say, "Hey, I'm going to be late because I need to go pick up some tools," or something like that. So that connectivity layer just got enhanced with the presence of all this technology.

That's right. Yeah.

And it's designed as an open system, by the way, because a lot of these newer assets, they have APIs, effectively. So a lot of newer trucks, for example, we can do a cloud-to-cloud connection. So you don't necessarily need the black box, but you want the data. And you want it organized and seamless with all of your other assets. Because typically in operations, you'll have Ford trucks and GM trucks and Caterpillar and Freightliner and all kinds of other equipment coming together. So we act as that orchestration layer.

So the hardware is very much part of the story, but we also have software interfaces coming into the system.

That's right. From an ingestion perspective, I think you got it. So, just massive amounts of data storage. This is all kind of sitting in modern hyperscaler clouds. So, it's not that we have, like, one big warehouse/data center, but a pretty large system. So, we ingest the data, we're storing it and organizing it, and then we basically have a bunch of processes that are automatically working on top of that data as well.

That's right. And there are many different personas. So you named a few of the key ones. The drivers in the frontline are very much just regular users of the system. You do have the dispatchers. You'll have other people who'd be, like, safety managers or, in certain cases, if they have to do paperwork, essentially compliance managers. But you also have people that do maintenance, for example. And so they want to know, what is the health of all these assets in the field?

Which truck will I need to maintain at the end of the day when it comes back to the yard? And then you do have the executives and all these other business-minded people who want to know, well, did we show up on time? Like, what is the efficiency of our fleet? And how do we do this at scale? Many of our customers, actually most of our customers, are large enterprises. So think operations with thousands, tens of thousands of people, and tens of thousands of assets, and so on.

Well, I think the great part about this technology is it's very tangible. And it's the kind of thing that, when you see it, you get it very quickly. So what we do is we tend to go on-site. We will demo the technology and do trials. So you can easily—these are plug-and-play—you can easily try it out in your environment, in your industry. And like I said, there's so many challenges in physical operations. It tends to never just be one thing.

So yes, we want to reduce the number of accidents we get into. But I think we're also leaving our trucks idling a lot because it's just a bad habit, or we're leaving tools behind at the job site, and we'd like to get those back because we spend millions of dollars replacing them. So we will often find multiple challenges like that, and then we demonstrate to them at small scale, like maybe a team or a region or something like that, that this works.

And when they see it, they get it immediately. These are people who are experts in their industry. So they would say, "I immediately see the value or the ROI." But they have to see it in that kind of tangible way. They're not just buying it because it's AI or big data or something like that. They're like, "No, if this solves problems for us in our construction business, great, let's do it."

Yeah, there are a few things. We see, whether it commoditizes or not, these models are incredible, right? And the amount of value they can unlock with their ability to ingest the data and reason is awesome. So we're very excited about what we're seeing on that front. The operational data that I was talking about, the physical-world kind of digitization side of things, is where we come in. These are not the tokens you're going to find online. You can't crawl Reddit and find out about what happened on a construction site, right?

Nor can you do test-time reasoning about it. You can simulate all kinds of environments, but really what our customers need to know is what was going on in that specific environment at that time, right? That requires this interplay of hardware and software, but also the change management. How do you get this out into the field? And the partnership. So that's a unique area for us that we focus on. This is what we've been doing for the last decade-plus.

And it takes a lot of work. I have to emphasize that, too. It's like we get out in the field with our customers, understand their business, and work backwards. So it's not something that can purely be just one-click deployed. It really requires a nuanced approach.

Very much so. On the dashcam side of things, the key insight there is, while these may be all different companies, we're all driving on the same roads, right? The Samsara system, on a given day, we're driving 99% of the U.S. roads, usually multiple times a day. So you can use that, of course, on the risk side. So we can understand where are the risky intersections or where are weather conditions bad, and how do we warn other drivers? So there's a network effect there.

But there are other sort of side effects, right? Because we drive all the roads and we have cameras, we can tell you where all the potholes are, right? And that is super useful for the city. So the city of Chicago, for example, they want to know which potholes happened after the winter weather season, what order should we go after them in, in terms of severity? You can use the camera data for that, the accelerometer data from the GPS tracker that I mentioned.

So every time you see that big bump in the road, you look at the video. And the cool part about that is not only do you know where the pothole is, but we can see what's happening to it over time. Is it getting bigger? Is it cracking? All that kind of stuff. So that is another sort of data network effect that we get. And then maybe the third, since we talked about asset trackers earlier, you have these millions of vehicles driving around.

A Bluetooth tracker on its own probably doesn't get picked up, right? Because think about a construction site. It could be acres and acres of land. But if you have one company delivering building materials, another company performing construction, the electrical contractor, one of those guys may pick it up. And that is another network effect that you get with millions of these vehicles and then tens of millions of handsets, like the mobile devices. It's an incredible kind of mesh network that forms.

We could spend an hour just talking about where what is going on. If I had to generalize, I would say at the edge, we're typically running inference and data collection. So the data collection, of course, gets us the training data. The inference is essentially running models where we have the weights and we send them down from the cloud, and they're running at many frames per second at the edge. And this is how we do the kind of real-time or low-latency detections and closed-loop alerting to the driver.

The reason we do it at the edge is practical, right? Sometimes you don't have a great cell signal. Many of our customers are operating in the middle of nowhere. And then also, the latency matters. If you can get feedback to the driver really very soon after something happened, it's much more likely they'll change that behavior. So again, very practical architecture for us. It's worked really well in terms of how robust it is and how it holds up.

But that being said, it's not fixed. So if that means we need to do some inference in the cloud, we are set up to do that. We have real-time tunnels that connect these devices.

There are lots of different model types. Convolutional neural networks are very much where we started. That's basically from the ImageNet era of, like, okay, can we detect a mobile phone? From there, these models have become more sophisticated. So we run basically a model backbone with many different classifiers and heads, or attention heads. So basically, once we see a device in somebody's hand, what is it? Is it a phone? Is it a vape pen? Is it a sandwich?

What is the activity that's going on with it? That's not a single-shot detection. It tends to be a little more nuanced than that. Same thing when we look outward: we have cameras that point out at the road. We have some cameras that point at the sides or to the back. And we're trying to do things like estimate depth, right? So am I likely to run into a lamppost or a mailbox or something like that? That's a different kind of model than what a CNN would be able to do.

We use generative in a few different ways. I would say, if we think about the overall class of models, yes, you can now reason about video. So what would have required a human-in-the-loop reviewer, and that's the kind of work that might have been done overseas in lower-cost geos or something like that, you can now do at much more volume in the cloud using these models. So, for example, if someone slams on the brakes while they're driving their truck, the naive thing to assume is, like, hey, the driver was distracted, and they kind of woke up.

The more nuanced thing is that driver might have been avoiding a deer or a dog or some kind of defensive event. If you can watch that as a video clip, you can now say, hey, we're actually going to give the driver some positive feedback because they did a really good thing. The VLMs are able to effectively do what I just said, right? Similarly, if you want to understand, did someone run a red light? These things happen. You need to have a pretty sophisticated model that understands the geometry of the road and all the conditions and so on.

So that would be like a JEPA-style model, for example. So we're able to use a few different model families. On the generative side of actually being able to create video, that's also very interesting because, from a coaching perspective, most of our customers are bottlenecked on the number of human-to-human interactions they can have. We could probably do that for a small fraction, but I can't do that for every driver. It's not practical. You may have seen AI-generated avatars, AI-generated people.

We can generate a coach, and that can resemble the VP of safety from that company, or a celebrity, or who knows, whatever the customer wants. But that can be a very effective way to deliver end-of-week coaching. So that's a form of generative video that, frankly, we couldn't have even dreamed of five years ago.

We are universally accepting of models in the sense that there's so much innovation happening. So yes, the frontier labs are doing great work, right? And we'll use multiple models from different labs simultaneously. The open-source models are pretty compelling. I think it's having its moment now, but we've been seeing—and this is really, I think, from the academic communities—open-source and really open-weights models really give you a lot of operational freedom. So we can do things like distill them, for example, to shrink them down to fit on a device.

So we use models like that. And then there's others that we train from scratch. Those might be smaller models, call it tens of millions of parameters, but very specific to something we need to do in the field.

Yeah. Well, a few years ago, we did introduce LLMs into our product. We called it the Samsara Assistant. So think of it as like a chatbot tied to your operational data. It became very popular. We saw customers asking all kinds of practical questions, like, who are my safest drivers, or which trucks need maintenance, things like that. Those tended to be single-turn or maybe like a few-turn interactions. Like, you just go back and forth with the chatbot window on the side.

The breakthrough that, of course, happened last year is these agents can operate over much longer time horizons. So instead of an AI responding to a question in a second, it can go do some work on its own, develop a plan, and go after it. We've seen the impact of that in the coding world, but there's also a lot of implication for the operational world, right? So one of the demos I did on stage is we have a warranty agent. When it sees a fault code, it can basically crack the service manual, look at your specific OEM-negotiated warranty agreements, and then correlate the two and say, yes, this specific issue, given the age or the number of miles that have been driven on this vehicle, is actually covered under warranty.

And then it can open a work order, put in the steps, and also tell you, hey, do any of the other trucks have that issue? That is what would have been like an hour or two of human labor that we've been able to automate down to under a minute. That is the huge kind of breakthrough unlock. And I just went very deep on warranties, but you can see how that would apply to reporting, how it would apply to briefing a driver at the beginning of the day.

You can use it in all kinds of creative ways.

Yeah, I think of it as we're just starting with the most practical areas we can have impact. The reality, by the way, is most of those warranty claims just are unfulfilled, right? Nobody has the time to go do all that work that I just mentioned and do the paperwork. So that's an area of tremendous interest for our customers, is, hey, I have all this extra work that I know would be useful, but I'm not able to get to.

So how do I do that? And then over time, I think the idea will be, how do we really autonomously run parts of the operation for the customer? If that's, like, replanning the route, for example, before every morning shift, we now have the technology to do that. And again, it's not super risky, but it requires a lot of business judgment of, that route actually is run by this person because they have a relationship. They've been seeing that customer for 10 years. You need to have all that context.

So in that sense, we are starting with things where we know we can have an impact, and then we're working with customers to figure out what else could we do.

I think agent reasoning was a huge unlock, like I said, this ability to build plans and work over long time horizons. But you do need to outline: what is it that I want the agent to do? And that is, in some lightweight sense, the workflow. It's also the guardrails. At what point do you say, "Hey, agent, you should ask me for some help"? Or, "Agent, I don't want you to go down that rabbit hole"? We have to kind of keep it on track.

So it's some combination of operational context, which comes through workflow and guardrails, with also this now kind of new agentic reasoning ability. I don't think either really works well in sort of isolation. And the workflow side of things, by the way, we had elements of that in our product. I'll give you a very simple example. In most commercial industries, there's a walk-around inspection that you do at the beginning of your shift and end of your shift. We see about 300, 350 million of those a year.

That has historically just been a workflow on a mobile device, right? You're kind of going step by step, taking some pictures, saying something's safe. If you combine that with the diagnostic information, the location information, who last did the check, what was in the picture, that is a huge unlock. So that's kind of what we mean by combining these two things.

Oh, boy. That ceiling question changes, I feel like, every week. And there's some nuance to this. So, for example, these new models, like the kind of Gemini 2.5 class and GPT-4o, it's hard to figure out where the practical ceiling is, but sometimes you do see them get distracted or lost in a loop somewhere, right? So I think there is some aspect of, like, they may find the answer eventually, but can they find the answer in 10 seconds or one minute or even one hour, right?

So that's one area where I think there is still a practical ceiling. And my guess is, as these models become more and more powerful and more sophisticated, that will shrink. And then these algorithms are getting more efficient. So maybe the compute combined with the algorithm, combined with just smarter model architectures, will make what would have been a one-day task a one-hour task or even faster than that.

I think a lot of our ability to predict what happens next year or two is by looking at what is barely possible now. And then what will the cost curves look like or capability curves look like? And something that we talked about at our Beyond conference last month was this idea of a driver ride-along. So in operations, it's quite common to basically have a manager sit with you over the course of your day, drive around with you for eight hours. And what they're doing is they're not looking for how fast you're going.

They're looking for your habits of, do you check your mirrors? Are you alert and are you aware? That kind of thing. That's basically a massive amount of video computation, right? So you can run a tokenizer. It just turns into a lot of compute. You can do that today. It's pretty expensive and costly, but it works and it's doable. We're pretty optimistic that the cost per million tokens is dropping so fast and the capabilities are rising that we can deliver that at better and better cost over time to our customers.

So I think in a year or two that will be possible. And I have to say, just over last weekend, I was playing with Cerebras, which is one of those big wafer-scale chip companies.

They have an inference model that you can run on their chip, which is basically Gemma 3, but hyper-accelerated. That's a great example of that is nonlinear in terms of jump, right? What you get out of these big models today, if you run Gemma 3 on your GPU, you might get like 100 tokens per second if you have a fast card. If you run it in their cloud, you get anywhere from 800 to 1,500 tokens per second. So call it 10x faster.

That is the kind of thing where it unlocks these new use cases that we couldn't get to because it would have been either expensive or too slow.

Awesome.

Yeah. A couple of thoughts there. The first is we actually do spend a lot of time on the frontline with drivers and other frontline workers. So it's very important for us to get their perspective because they're the primary users and really beneficiaries of the system. Something people don't often think of is, if you have a dash camera, what is it used for? The majority use case is actually exoneration. So what I mean by that is helping explain what happened if there was an accident because, for example, if you are The Home Depot, you're a very well-known brand.

They're a customer of ours. Lots of claims, auto claims, are placed against you because they'll say, hey, a Home Depot truck backed into my car on this highway, right? And that is something that really upsets a driver because they'll say, look, I was doing my job great. I didn't run into that guy. Now you can basically produce HD video evidence of where you were and, if there was an accident, who caused it, all that stuff. That eliminates all the ambiguity, right?

Now you can just resolve it. And look, if there was an accident, the company may choose to just settle it out and pay it out. But if there wasn't, which is a very common case, now you can really fight it and say, look, we know exactly what happened. Drivers love that because, I have to say, 90% of the time, they're doing a great job, and nobody's seeing it, right? And so that has been a huge unlock, is this positive reinforcement of, we're analyzing the whole drive.

We're seeing all these good behaviors, defensive driving or exonerations, things like that. That, I think, is sort of the counterweight to the, hey, what is all this for? How is it being used? If you have that in your culture, if you are kind of doing the equivalent of employee of the month but showcasing really great work, I think people get really excited about this. Also, from a safety perspective, we should remember, in physical industries, the risk of injury is on the person, right?

In other words, we want everyone to go home the same way they came to work, right? That is an important concept that I think people don't think about. If you're working construction, you're working oil field services or something like that, you take a lot of risk when you do your job every day. So it's actually in the what's-in-it-for-me: we are trying to keep you safe. If you have that and you do it in a transparent, thoughtful, respectful, privacy-conscious way, it goes a very long way with the frontline.

Yeah, very important. I think transparency, again, is the key word here. And by the way, our cameras are not hidden cameras. They're quite visible. So it's not like a secret sort of recording device. And the whole idea is to bring the frontline along. So we call this change management, right? Like, hey, we're introducing these things. What do they do? What are they for? How can you use them? How can they be useful to you?

If you have that conversation early in a transparent way, it tends to be quite constructive because I mentioned Home Depot earlier. They saw like a 65% reduction in their claims, like auto claims. That was a huge win for that organization, both at the sort of executive level, but more importantly, at the sort of regional level. Those kinds of wins are what we want to help create. Now, could you use this in a bad-guy kind of way? You could, but that would be like, who is sitting there watching each driver?

It's super boring, by the way, to sit and watch drivers, right? So when you are transparent about what the system's doing and what it's not doing, and then how the data is used, that is how you get the buy-in and you earn the trust of that entire organization.

Yeah, if we imagine five, 10 years out, where does this go? I do think, yes, we expect there to be a lot more robots involved in physical operations. You see this, actually, if you go into a warehouse today. So if you go into either manufacturing or a fulfillment center, there's actually a lot of automation and robotics going on. And the cool part about that is it reduces risk of injury to a lot of the human workers. Lifting injuries are very common.

10, 20 years ago, they were way more common. They're way less common these days because, quite literally, the robot's doing the heavy lifting. Now, there's still humans working there because there's kind of all the handoffs and there's some nuance to the operation. But we expect something similar to happen out in the field, right? So think about a construction site or maybe a company building a roadway or modernizing the grid. There's a lot of kind of repetitive work that has to happen. Imagine you're grading a site, like you're making it level.

Could that happen during the third shift between midnight and 8:00 AM, right? That could be a really cool way to do productive work on the side of the road where you're just going for like five miles, making it flat, right? We see robots being able to do that in the next five years. Now, all the rest of it, though, it's still pretty messy, and construction has like exception after exception. Like, you're solving problems constantly. That's where I think the humans offer a lot of experience and judgment of, well, how should this work?

And I'm waiting on this building material. Well, I can perform this other thing. Meanwhile, the robot's making the road flat, right? So that's kind of what we see in the next few years. The same thing applies, I believe, to logistics and supply chains. There's a lot of kind of last-mile complication that happens, and you have to physically pick up, deliver a package. Some of our customers, they stock the shelves in the grocery store. Maybe we get there with humanoids, and never say never.

Like, this stuff always is evolving. But in the meantime, could you automate the long-haul segment between Dallas and Phoenix, right, of all of those beverage cans coming in or something like that? So we see this as exciting in terms of what the future looks like. And what we've seen in operations is very diverse: lots of different types of equipment, lots of different types of labor. So it's going to be different makes and models and different makes and models of different kinds of robots.

is my guess.

Well, practically speaking, most of our customers would tell you they're supply-limited in terms of labor, right? So these are labor-intensive, asset-heavy industries. So they welcome this idea of, like, could I automate some of this labor so we can basically perform more? So that's kind of like what our customers are going through. They have a lot of work to do. What we want to do is provide the single pane of glass so they can orchestrate the whole operation, trigger the workflow of, okay, that truck is arriving from Dallas.

Let's get queued up so the warehouse system is ready to go and accept it. Let's make sure that we are notifying our end customer, and so on. And then I showed the tracking label earlier. You could put that on the pallet of goods so you can track it end to end as it changes hands through the supply chain, as it makes its way to a job site, as it's installed. This is all very opaque today. If you think about your shipments that you receive as a consumer, you might get five updates, right?

Left the facility, out on the road, out for delivery, et cetera. We see hundreds and hundreds of pings. That's going to be really important when things are moving on their own. Whether that's aerospace assembly, right? We have customers who want to know where this really expensive asset is that they need to perform their job. Right now, that's sort of untracked.

I think on the robotaxi side, it's going to happen a bit faster because the operations are much more regional. They're much more similar, right? Like, the way that you and I ride in a taxi across town is going to be quite similar. And so that's, I think, where you're going to see the biggest sort of visible impact of autonomous vehicles. On the trucking side, it's also important to realize there's long-haul trucking and kind of moving stuff from point A to point B.

That tends to be a minority fraction of what the commercial vehicles on the road are doing. Most of the commercial vehicles are in industries like field service, right? So they're HVAC technicians or plumbers, electricians, people performing some work. And they're either doing something like that or they're in industries like construction, where they're building the road. That tends to be where current-day autonomy doesn't work so well. It's like the really messy long tail. So for that reason, we think the adoption might be a bit slower, but it's not like a no.

It might take 10, 20 years. And these are industries where, again, the equipment's highly specialized. Like, if you look at cement mixers or garbage trucks, these are custom-built. So for the autonomy systems to make their way out to that edge, it's just going to be a longer diffusion curve than for a sedan or a van or something like that, which is very much the same.

Yeah, very much so. I think our customers—it's very clear—they're busier than they've ever been before. So from an American economy perspective, we're seeing a lot of intensity. I was just in the field last week with a large energy utility, and they've been involved in grid modernization. And they shared with me a really interesting stat. They said, over the last 125 years, we built a certain amount of grid capacity in megawatts or gigawatts, really. In the next five years, we're going to triple that.

They, as a company, are going to 3x the amount of power they deliver. And that's in five years versus 125 years. So that requires a tremendous amount of infrastructure build, even with new technologies. They can't work fast enough. We are seeing that across so many different kinds of industries. And in that case, they also shared that 90% of that demand is data center-related. So as the data center demand continues to skyrocket, the energy needs—there are all these different bottlenecks that have been appearing.

So many of our physical operations companies, customer companies, are involved in that directly.

Yeah, I will warn the lawyers thinking about becoming plumbers: it's a pretty messy job, right?

It's pretty hard stuff. So yes, we've been continuing to see that there's basically a labor shortage, labor bottleneck, in a number of different trades, but also things like long-haul trucking, commercial driver's license holders, things like that. So, in general, I think there's a lot of growing demand for these professions. And this data center boom is a great example. There are just not enough electricians out there right now. So you see companies like Meta doing initiatives to reskill people, train them on how to become a good electrician.

And then how can you take the people who are trained and make their jobs as efficient as possible? So you don't want them waiting on materials at a job site. You want to put them to work to perform wherever their skills are needed, kind of thing. So, very much kind of a bottleneck, but the trades are in incredible demand. And I also think that their jobs are getting more modernized as well. Because if you think about it, as an electrician, a lot of it is getting to the job site and having the materials and knowing what you're going to do.

If an AI can kind of help you with that, it takes a lot of the mental load off, and you can focus on the really unique trade value you have.

Thank you. It's been fun.