If you think about it, there's like a third shift between midnight and 8 a.m. roughly, right? That people tend not to work because they're sleeping. Imagine if operations like logistics could still run during that shift, right? Um, and then same thing, imagine you're a field service technician, you need a part. Like how amazing would it be if the part could just get delivered to you? Like that is something that's going to be a nice augment to operations. So it's interesting because typically when you see automation kick in, again volume increases, right? Because costs come down. There's way more demand out there than people realize because sometimes you'll say, 'Yeah, I could use that part, but I don't need to deliver it if it's going to cost 50 bucks for someone to drive it to me.' If it costs five bucks or no bucks, like how awesome would that be? So we kind of view it as like it will increase the speed that the world operates at.
In this episode, we talk with Sanjit Biswas, founder and CEO of Samsara. Sanjit formerly founded Moroi and has a legendary reputation amongst Sequoia's founders. So, I'm excited to welcome him today for a conversation about physical AI. Samsara is a $20 billion market cap public company with sensors deployed and streaming data from millions of vehicles capturing 90 billion miles annually. Sanjit shares his insights about the constraints of physical AI from running inference on 2 to 10 watt edge devices to why the messy diversity of real world data is both the biggest challenge and opportunity for embodied AI. If you're building in robotics or physical AI, this conversation offers a rare perspective from somebody who's actually scaled it. Enjoy the show. Sanjit, thank you so much for joining us today. You are a legendary Sequoia founder and it is a delight to have you back at Sequoia.
Thanks for having me. It's great to be back.
I want to start with your background. So you went from MIT's RoofNet project to co-founding Moroi through its 1.2 billion acquisition and you are now the founder and CEO of Samsara, a 23 billion market cap company with the best ticker on the public markets. What's the through line? And tell me about your personal passions and experiences and what the through line is between all of that.
Yeah, so I'm an engineer by background. So studied EE and CS. I went to undergrad out here at Stanford, went to MIT for grad school and that's where we worked on this project called RoofNet. So the through line for me has been about building cool products, cool technologies that have real world impact. And RoofNet, this is like over 20 years ago. The idea was could you build really big wireless networks because at, you know, kind of early 2000s Wi-Fi was not mainstream. It was a brand new technology. Internet access was just becoming mainstream and it was still pretty expensive. And so we saw this opportunity to take Wi-Fi chips and all that technology and use it to build really big networks. And so we kind of had this idea that internet access should be everywhere, right? It should be in the air. And how would you do that? We need to build a big network. So that was RoofNet. And then with Samsara, it's a bit of a different sort of focus. We focus on the world of physical operations. So think all the infrastructure companies whether it's energy utilities, construction, logistics, like all these real world physical industries. And the idea has been real world impact through things like risk reduction, improving efficiency, improving sustainability, just using all this data and now AI.
Physical AI feels like it's finally going through an inflection moment. You've been building Samsara for the better part of a decade now. What did you see at the time and like how has the field changed? Why now?
Yeah. So, if I rewind 10 years to when we were founding the company, we had a couple of sort of intuitive bets or guesses. And the why now for us at that point was connectivity. So, we'd been through the Moroi journey and we'd seen internet access go from being kind of rare and inexpensive to being ubiquitous. So, this is 2015 for reference. We saw basically the ability to process large amounts of data really coming online. So the cloud had matured. We were seeing the beginnings of the GPU wave. And if you remember back to 2015, Nvidia was a player and they were doing a lot of interesting like embedded GPUs. So if you picked up a Nintendo Switch back then, it had amazing graphics, but it fit in your hand. And so we saw compute was getting really good. And then we saw sensors really specifically cameras were getting really good because this is probably seven, eight years after the iPhone launch. So cameras had gotten extraordinary. And you combine all these three things together. You've got connectivity, you've got compute, and you've got sensors, cameras. And we said this is the sort of like makings for a total sea change when it comes to ability to process data in real world context.
Wonderful. Okay. I'm excited to nerd out more about technical questions on the frontier of physical AI. Before we get into it, maybe can you just say a word on Samsara for our listeners? I guess how much of the business is, you know, I think of it as very much having had roots in commercial trucking. How much of the business is that today and what do you see the ultimate vision of Samsara being?
Yeah, so we really focus on the broad world of physical operations. So think about all these different kinds of industries. Trucking is definitely one of them. It's about 20-25% of our business. So logistics and big trucks on the road. A lot of our business now is related to field service and construction. So other big kind of frontline industries. But we also are starting to do work in like public sector. So we work with local governments. We work with student transit. So we just signed like the largest yellow school bus operator in North America which is pretty cool. And we work in industries like aviation. So think about like labor intensive, asset heavy industries that really power the infrastructure of the planet.
Wonderful. Can I go back to your comment on so at the time there was a why now around bandwidth, compute, and cameras.
And it sounds like you may not have necessarily had a crystal ball on what was going to happen with AI.
But you kind of felt like you're on the right side of history and that with those raw ingredients, you'd be able to do increasingly sophisticated stuff over time. Yeah. So I guess the question is like since you kind of had a directional sense for where the world was going, how did that influence the way you built the business? Like was there anything specifically you did just kind of in anticipation of this inevitable direction that the world was going?
Well, actually the historical context is important. So our first company Moroi, which was funded by Sequoia, we were domain experts. So we knew a ton about networking because that's what we'd been working on in terms of our PhD. With Samsara, it was kind of the opposite. We knew nothing about this domain. Like we'd never driven a commercial truck before, never worked in a warehouse. And so we were sort of eyes open about it. What we did have was that intuitive sense of the compounding rate of those underlying technologies. So we said, okay, there's this really interesting like problem space, this world of physical operations. It's kind of like overlooked, especially 10 years ago. No one was really talking about infrastructure the way they are now. But there are, you know, things are changing very quickly behind the scenes in terms of tooling. So that intuition is exactly sort of what we were powered by and we said even if it's not mainstream yet or it's not ready yet certainly in 5 to 10 years which is about now it will be possible to do this stuff. So I think for a lot of the current founders it's kind of like if you look at AI model capabilities even when like you know the chat GPT moment happened these models weren't perfect they've gotten a lot better in the last 2-3 years and they're going to get even better in the next two to three years. I think technical people understand that in a way that consumers and customers often may not see yet.
So I think you have an embedded systems background and you know you're one of the unique people that's operated at the intersection of the hardware and software worlds. I'm curious what are the things that make building AI in the physical world different than running AI in big data centers?
A couple of things. Well, it actually is a lot of fun. So the physical world is very diverse. You know, you see a lot of companies now working on physical intelligence and world models and it's because the training data set is really broad and vast. So if you think about our products, we have products like dash cams that end up on the roads on millions of vehicles. They see like 99% of the US roads. It's just an incredible data set. You've got urban, you've got rural, you've got residential, you've got weather. And so we see all these like interesting exceptional cases. So the training data is really interesting. And then what we can apply all the inference and basically pattern matching to is also interesting. So I think that's the most fun part. The most challenging part though is how messy it is and how distributed it is. So for our products, it's not practical for us to just stream all the data to the cloud. It would be like a crazy bandwidth bill. You need pretty massive data centers if you think about millions of video streams like constantly running inference. And so we have a much more distributed architecture where we actually run in the cameras themselves. And that changes your compute and power footprint. You know, we're talking about 2 to 10 watts, not like kilowatts, right? But you can do a lot more because you've got millions of them.
And so what does it like, how do you run, I'm thinking some of these large LLMs, even the image models are very large right now that people are working with. Are you running just very bespoke small models on 2 to 10 watts? That doesn't give you much.
It doesn't give you a lot of room and that's a fun engineering problem. So if you think about it, these state-of-the-art models, they are very large. So you're talking about like hundreds of millions of parameters or billions of parameters that is simply not possible. So our footprint is much more similar to what you can run on like your mobile phone, right? So it's not tiny. It's not a microcontroller. It runs Linux. It's got like hundreds of megs of memory, maybe gigs. But it's not like a big data center, right? So what we tend to do is we will train models in the cloud, we'll basically distill them down or use teacher models. So we'll use a big model to basically instruct a small model that's really designed for our use case because we don't need to be able to answer, you know, what the capital of France is, right? Like that's not something the dash cam has to encounter. But we do need to be able to understand what is the risk profile on the road. So we train it with the data that's relevant for the task.
How much of the data, you see 99% of US highways or US roads. How much of that data can you make use of? How much of that data do you make use of?
Yeah. We can make use of a lot of it and we basically have the ability to train over like this entire data set. There is a very practical question of like okay you run a tokenizer at the edge, you send all these to the cloud, what do you do with it? And what's cool about that is what we do with it this year is so much more interesting than what we could do with it two, three years ago. So two to three years ago and these products really started around this idea of reducing risk. So if you think about the problem we're trying to solve, it's that you know our operations customers they operate on these roads every day. It's actually the riskiest thing that they do is more so than construction or working in oil and gas. Driving on the highway getting to and from the job site is where they incur like most of their like fatality or kind of high severity risk. So the question is how do you go take all these images and tokens and turn it into a risk signal, right? A couple years ago we said you know the biggest risk we are seeing right now is mobile phone usage, right? Like people are on their mobile device while driving a big truck and that's super risky. So we built a detector for that. You do that and you say okay we can solve this problem, we can detect mobile phones, what else drives risk? Now we're seeing things like weather, right? And weather's always been a risk factor, it's not a brand new one but it's now something we can detect using these pretty sophisticated models. Training a weather detector using like old school convolutional networks like an AlexNet style model, you would have gotten a lot of things wrong. You couldn't tell the road conditions. Once you use more sophisticated models like the ones we have today, you can really figure it out. So that's the cool thing is there are these unlocks that happen every couple years as model capabilities increase and our data set increases. So these two things really work in our favor.
Is there an upcoming unlock that you are most looking forward to?
In terms of our products, or just in general?
A new capability that's going to unlock some new use case or some new feature for your product?
You know I feel like we are seeing just such incredible foundational model capabilities that are making it possible to just inference over huge amounts of data. So historically what we did is we understood like what was happening in the moment, right? So like I said mobile phone detection or not wearing a seat belt or following distance. Now we can start to really look over the course of a trip and we're not only detecting like negative like risky downside events, but we can actually detect good behaviors too. And I'm really excited about that because frontline workers 80-90% of the time they're doing a great job. No one's able to recognize it because no one sees it. So what's awesome is we can now see that someone's doing awesome and give them a high five or like some kind of recognition or kudos that is like making people's day and it's a cool like silver lining side effect of having all this stuff running. So anyway, it's kind of an unexpected upside sort of thing. Yeah.
And do you think it'll be video reasoning models that sort of empower that? I know you can't run a ton, you can't run giant models at the edge, but are you doing stuff server side that takes advantage of LLM?
Yeah, I should have mentioned that. So the model's connected. We have a ton of inference running at the edge. It's running continuously because when you're driving there's like continuous risk and then we're taking those tokens we're streaming them up. And additionally we have images, we have video, we have other kinds of telemetry, and then we can go and run all kinds of sophisticated things in the cloud. So if we need to understand when an accident happened, what really happened, we can run a full video language model, like a reasoning model essentially in the cloud and that can say oh this was actually defensive driving and this guy got cut off or these were the conditions. So that is really cool. We couldn't have done that 5 years ago.
Do you believe in world models? Loaded question.
I do. I'm cautiously optimistic about them, but I think you need a tremendous amount of data.
Yeah. Are you guys training your own world models?
We are not building our own world model. And I think that requires a very specific kind of focus, but in the same way we don't train our own base foundation models, but we are looking forward to using them at some point.
Yeah. And I imagine you have an incredibly rich data set that might be useful.
We do. Yeah. We see about 90 billion miles on our system every year. So, it's a lot of driving.
Yeah. It seems like the sensor footprint you've built out is like a tech nerd's dream, right? Most people dream of a connected world and, you know, you should be able to have so much telemetry on all these different attributes of the physical world. But as far as I can tell, you're one of the only companies that's really gone out and, you know, put sensors on the physical world in a really meaningful way. Why do you think that is? And what's the key to actually being able to make that dream happen versus have it just be a, you know, tech nerd's dream?
Yeah, first of all, it takes a village to actually get the stuff out there. And I think that's maybe one other big difference between just pure software and physical world. Is we have to get the products installed, so they're installed in millions of vehicles. We have to train frontline workforces on like what the stuff is and what it's doing. And then we have to provide value to all these customers kind of from day one, right? Like they have to get something out of it. You combine all this together and you get this big footprint. But it's been hard because you need thousands of people at our scale now to do this and to do the change management, like the installs and all that kind of stuff. You know there are a few companies that have data sets of this scale but it's like Tesla and probably us, right? And then Waymo, there's thousands of Waymos but not millions and maybe it will be millions in the future but we're not there yet. So that gives you a sense of how much effort is just like sheer willpower is required to get this stuff out there.
Speaking of which, I think there are a lot of founders right now who are technical founders like yourself who've built something cool and are now encountering this crazy supercharged race to scale that the AI wave seems to have brought. And so I guess the question is you are a technical founder. I think both Samsara and Moroi have been known for go-to-market execution. And so maybe the question is like how important has go-to-market execution been to your success and as a technical founder was it obvious to you at the beginning that it was going to be that important or kind of what was your journey like in appreciating the importance of go-to-market execution if that makes sense.
Yeah. I'm replaying like 20 years in my head really fast. So when we started Moroi at that point in time like I had never sold anything in my life. In fact like as an engineering nerd I avoided any situation where there was like you know those like fundraiser you have to sell candy bars at school. I was like does anyone need a website for this thing? Like you just try to find some way out of it. So I really was not like a salesperson in terms of background and no one in my family had done sales. So it was very foreign. The thing that turned me on to it was this idea of this is what it takes to get the product out there. And if the product's not out there, it's not having impact. So if you're driven by impact, if that's what motivates you, it's fun to see people using it, right? And then this is what makes us sustainable. So with Moroi, we were growing the company between 2006, it was acquired in 2012. In the middle of that was a great financial crisis, right? There wasn't a lot of funding at the time. Like risk capital was just like turned off. So we basically had to make the company operate at break even, right? Or thereabouts. And that's what really convinced us like we have to figure out how to have sustainable sales execution and a model that's highly predictable. And as engineers, we're like, hey, this is actually a big engineering problem, right? And then that stuck with us with Samsara. We're talking about impact at scale. We raised capital along the way, but actually we reinvested way more just from the revenue of the company, the gross margin. So if you look at our numbers, we're public now. So you can kind of go back through the balance sheet. You can see we've invested probably close to three billion dollars just in getting the stuff out there, right? R&D, customer success, all that stuff. That is only possible with a lot of sales, right? So once you understand the why, you can kind of buy into it and say, I'm going to figure this out. It was not natural for us, but it was a pivot that ended up being something we had to do, and I'm really glad we figured it out and have been getting better at it each year.
