The Case for Automating American Factories With Formic

That’s exactly what our guest in this episode, Saman Farid, CEO and founder of Formic, has spent years figuring out. 

In this episode, Saman talks to Chuck about the actual state of the robotics industry, why manufacturers are turning down orders not for lack of machines, but for lack of people to run them, and how Formic’s ​“robotics as a service” model is changing the calculus for manufacturers who can’t afford or simply don’t want to take a big bet on capital equipment.

They also dive into the surprisingly nuanced financial structure behind deploying a robot fleet, the AI of it all, and what the next decade of American manufacturing could actually look like if we get this right.

Chuck Templeton: Very excited to be here today with Saman Fareed, CEO and co-founder of Formic. Saman, welcome.

Saman Farid: Thank you so much. Thanks for having me on, Chuck. I’m excited to be talking about robotics. I think been obsessed with robots since I was a kid, and it’s cool to see the whole world has, over the last three months, it seems like everybody’s caught the bug

Chuck: Yeah.

Saman: Imagining a world where we have our kind of robot minions doing our bidding. So I’m excited to do our best to bring that to reality.

Chuck: Very cool. So you’ve built and sold companies, multiple of them actually. You’ve worked in my profession, venture. You’ve built factories in China. How did all this lead to founding of Formic? What was the insight that you had?

Saman: Yeah, like any good entrepreneurial story, it starts from frustration, maybe a slightly different kind of frustration than most which is I spent my childhood kind of in and out of factories, dreaming about robots. I spent even some time doing manual labor in factories when I was a teenager.

And then as my career evolved, I kept finding myself in situations where I was either on the kind of technology side where we were developing and inventing new capabilities and new technologies, or on the user side where we’re using technology to build things.

And in all of those environments, I just consistently was disappointed by the gap that I saw between what was possible. And what was actually in use. And I think it just creates a lot of anxiety in me to see there’s a better way. We could be doing this so much more efficiently.

And I think it’s the classic line, ​‘The future is here, but it’s not evenly distributed.’ And I feel like that’s my life’s work is to try to make sure that I’m not the greatest scientist in the world, I’m unlikely to invent brand new methods or techniques, but what I really find kind of in my calling is I just wanna make sure that what’s possible is actually in use in as many places as it can possibly benefit humanity.

Chuck: So what were some of these insights that you had that there’s this gap there that could be filled? Because there’s obviously a lot of companies that have struggled in robotics and physical goods and stuff like that but obviously there is a way to get into it that is successful. You’re proving that every day. What were some of those insights or misunderstandings that people had about robotics that people should know about?

Saman: Yeah I think that there’s two extremes when it comes to misunderstandings of robots. One of them is the misunderstanding that’s common in manufacturing, which is that they don’t work yet. That they’re so complicated, so hard, so expensive, so custom that I, as a factory owner, can’t use them.

And this is, I think, a common misconception. A lot of these factories have tried in the past to do it through custom engineering efforts and have failed. And to your point, less than 10% of American factories have even a single robot. So we see all these how it’s made videos of big automotive plants and things like that where robots are all over the place, but actually in the vast majority of American factories, there are no robots at all.

So that’s one set of misunderstandings, is people think this is unreachable and extremely hard.

Chuck: Yep.

Saman: The other common misunderstanding is the other extreme, which is I think equally bad, which is people think, hey, it’s all solved. Humanoids are gonna show up tomorrow and solve all my problems.

And AI is just gonna become so good that robots are gonna be everywhere tomorrow. That’s also, I think, a misunderstanding and actually very disempowering because if you go and look at real production tasks, there’s a lot of work to be done, particularly for these general-purpose robot hardware form factors.

And improvements in AI are definitely helping, and they’re one part of the solution, but they’re not enough to drive global adoption tomorrow. And so we’re actually somewhere in between these two extremes today, where robots work really well for some things.

They work really poorly for other things. And with the right infrastructure around them, they can actually be useful in a lot more environments than they are today.

Chuck: Yeah, and that’s, I think, one of the main things you brought to this is not trying to solve everything there was to solve for all these manufacturers, but to be able to be very choosy or specific about the areas that you could help them be successful at right now with all the capabilities and then figuring out how to grow into the capabilities or grow the footprint as additional capabilities come online.

Is that right?

Saman: Exactly. I think to put it in simpler terms, we are looking for the low-hanging fruit, right? We’re looking for the things where robots already can work well, can have a huge impact, and can work reliably. And there are a lot of them. And I think when we go to any factory, you can easily start to differentiate tasks that are being done.

And you can say, here are the set of things that robots today can already solve. Here are the set of things that, with some elbow grease, we can get them to do. And then here are some tasks that are with the current state of technology today, not solved. And there’s a gradient there. And I think one of the important factors driving this adoption of robotics is knowing where each task sits on that gradient and how you can best solve them.

Chuck: Very cool. Being in the food business, I’ve been around lots of food manufacturers and walked through some of their floors and facilities and whatnot and it strikes me that a lot of these machines are decades old. There’s one that I saw that I think was still running machines from the seventies for a product that has been around longer than that.

But they just keep punching it out, decade after decade, amortizing the CapEx across millions or even billions of units, over those years. And so how do you get into a situation like that where you actually can bring the cost down, you can make it safer? What are the benefits that you bring? Like what are some of the current solutions specifically that you have now?

Saman: So you’re a hundred percent right. There are a lot of manufacturing businesses in America that have been around for a long time, and they have this large existing install base of equipment that they can use.

One of our customers used to make parts for the Ford Model T. It’s still owned by the same family, and actually a lot of the machinery that they use is the same machinery they used back then to make parts for the Ford Model T. So when you walk into these factories, it’s sometimes shocking how you feel like you’re transported, 50 years, a hundred years into the past.

At the same time, that’s a huge asset for these businesses, right? They have this fully depreciated equipment, and they can use it to make more new modern products. Things like bending equipment or cutting equipment. Like it’s the same. A saw is a saw. A mill is a mill. A drill is drill.

And the problem that they have today is that they don’t have humans who can actually run those machines late into the night or early into the next morning. And so what we’re able to do for a lot of these factories is we offer what’s called a machine-tending robot. This is one of our product categories, and these machine-tending robots basically sit in front of a hundred-year-old press, or a hundred-year-old drill, or a hundred-year-old milling machine.

They load and unload it with parts. And so that allows the operator or the factory owner to basically drop off a bucket full of parts at night, at the end of the shift when they go home, and they come back in the morning to a completed batch of products. And that’s one of the things that we see allowing these existing American manufacturers to get superpowers.

Because when you have fully depreciated assets and you have fully depreciated buildings and fully depreciated forklifts, you can produce products for actually a very globally cost-competitive price. There’s no reason you can’t be cheaper than China or anywhere else. The only missing ingredient is access to labor, who can actually run that machine and who can run that part all day, every day.

And so we see this as very common in automotive and aerospace. There’s a slightly different dynamic in another industry where we work a lot in is consumer packaged goods and CPG and food and beverage manufacturing. Food and beverage manufacturing is a really interesting use case because it is an industry that’s dominated a lot by peaks and valleys, right?

Particularly when you connect it to the agricultural supply chain. Agricultural products generally mature on a relatively fixed schedule, and when they’re mature, you have to be able to process them. 

And so you end up with a lot of waste or product that’s stored for long periods of time until it can actually be effectively processed. And working in food processing facilities that are preparing products that we can then go and buy in a store generally requires a lot of automation and robotics as well.

Again, this is an area where labor is a massive shortage. These factories typically have 30 to 40% unfilled headcount. I was in one of our customers’ factories a few weeks ago. They make different kinds of food products that you might buy in a store. The entire factory has about 165 employees and in the calendar year, 2025, they had to hire 270 people just to keep those 160 headcount filled.

Chuck: Wow.

Saman: So it gives you an idea of the rate of turnover. These are really difficult jobs. People get injured, people get back injuries all the time. If they can find any other work, they typically do, right? Because these are really challenging roles to fill. And when we think about our supply chain resilience and our ability to produce the goods that we need, we’re absolutely reliant on our ability to fill up these factories with labor. There’s also the demographic crisis where a lot of workers in these factories are aging out of manufacturing and they’re not getting backfilled.

Chuck: Yeah, they’re taking a lot of knowledge with them too.

Saman: That’s right. Absolutely. Yeah. So a lot of young people are choosing to go into other professions and careers, often better-paying careers with different experiences.

But that leaves us with this massive shortfall. And I think that what a lot of people don’t appreciate is that this is absolutely an area that America has a chance to win in. People often compare to China. China has obviously an incredible manufacturing base, but the US has the world’s second-largest manufacturing base, right?

It’s nothing to sneeze at. And I don’t think we can let it atrophy. There are about 350,000 factories in America that produce all kinds of products. And they’re facing this constant risk of atrophy. And if we don’t build much easier ways for ​‘em to access robots and modernize their facilities, they’re really gonna struggle. To your point training is a big part of it. New tasks for humans to do more complexity, and then robots to do all the things that humans don’t want to do.

Chuck: So what’s the sell then? When you go in and the first thing you’re presenting or pitching to them, is it just pure labor replacement, or is there a different subtlety to it than that?

Saman: Yeah. No, absolutely. It’s generally mostly about production, throughput, and production capacity. So every American factory is behind schedule to some extent, right? Either they’re actually behind schedule, meaning like they have orders that are unfulfilled. We went to a customer recently that’s three, four months behind schedule on their deliveries.

So that’s one huge problem. Another big problem is that they are unable…

Chuck: And that’s not cause machine capacity or throughput capacity. That’s because of labor.

Saman: That’s right.

Chuck: Okay. Or maybe even parts. But primarily labor.

Saman: Typically, it’s almost always labor. It’s generally not raw materials. And there’s three ingredients that go into producing something, right? There are raw materials, labor, and facility, right? You generally have the facility. You generally have the raw materials. You typically don’t have the labor. And so one way of being behind schedule is not is literally being late on your orders.

Another way of being behind schedule is saying no to orders that you know you can’t take on, right? You’re at max capacity, you’re operating at 100% of what you can produce. So you say no to business all the time. And this is very common in American factories where they have to turn down business because they don’t have the throughput.

So what we do is once we put robots into these facilities, we see that they can run those factories for longer. They can produce more with the same footprint, and now all of a sudden they can take on more business, they can win new orders, or they can get up to speed on their committed deliveries.

So that’s a huge impact. The second one is it reduces worker turnover. So there is existing labor in most plants, but because these jobs are dull, dirty, and dangerous, people quit all the time. You can go to any Amazon warehouse, you can go to any factory, and you can see that people typically stay for a few months on the job.

I went to a cookie factory in Chicago near you recently, and they had big billboards up, all across the factory. I think it was a $5,000 referral bonus if you can refer somebody to work there, and another $5,000 bonus if you stay in the job for three months.

Chuck: Wow.

Saman: And basically, what they told us is they almost never pay out that second bonus because people come and they stay for a little bit and then they leave.

This headcount turnover is a huge issue when it comes to training, when it comes to upskilling, when it comes to improved expertise and experience. And what we find is once we put robots in for the worst jobs in that factory, actually, the workers end up taking on more responsibility.

They stay longer, they’re happier, they get way less injuries. There’s way less workers’ compensation claims that the businesses have to handle. And we end up with better quality throughput and product. And so for a lot of these businesses, this is the difference between life and death.

Because the typical American manufacturing business has pretty small margins. In CPG, it’s not uncommon to have something like five to 10% margins. 

These margin percentages easily flip to negative. If you have one month of unexpected situations, right? You have a few workers who leave, and you have to retrain them, or you have to spend a little bit more than you expected on health insurance or whatever it might be. These businesses suddenly go from being barely profitable to unprofitable.

And you can imagine as the families that are running these businesses, most of ​‘em are family owned, it becomes extremely difficult to plan for the future. And so with something like robots, they get kinda repetitive throughput. They get the baseline capacity handled, the workers can start upskilling and improving and going into management roles, and now all of a sudden these businesses are much more healthy in the long run.

So very value accretive.

Chuck: Very cool. And so is there a different business model that they’re used to, or how are you coming in to help them? ​‘Cause they’re not sitting on piles of cash. How do you help them move some of the stuff in, moving it in from a financial standpoint, but also then thinking about the, the training and other things that happen with that?

Saman: Yeah what we found is that the only way to drive adoption of robotics is to do what we call full service robotics. So full service, robotics means we take care of the end-to-end complexity. We walk through the plant, we help ​‘em choose what robots to deploy. 

Chuck: Cause you’re robot agnostic. Is that right?

Saman: We have a library of stuff that we work with. Theoretically, we’re agnostic. In practice, we do try to keep to a certain set of things. But we go and evaluate the site. We tell ​‘em, here are the tasks that you should automate. Here are the tasks that you should do later.

We choose the robots for those jobs. We foot the bill for all the deployment and installation expenses. We install these robots. Then we just charge a subscription on the usage of that robot. That subscription includes the hardware itself, it includes the software that runs that robot, and something that we’re constantly updating and improving that makes the robot more capable and more flexible, and it includes all of the service and maintenance.

So this is really important because you can’t leave a chocolate chip cookie factory owner with a very new complicated robot and expect them to know how to manage it and maintain it. So when we take all that complexity off the table, we take all the risk off the table, and when we take that risk off the table, all of a sudden they can start to imagine how they can expand robots.

So that’s how we get our foot in the door. We make it extremely easy. And this subscription-based model. Allows them to also flexibly scale up and scale down capacity. They have some product that they’re making today for Walmart, and maybe next month they’ll need to make a different product for Costco.

So if those product changes happen, we’ve developed this kind of model that allows them to easily flex in and out of different solutions. And that also ultimately translates to a much, much better. Operation on the factory floor because now they can work with a variety of equipment, variety of products, variety of SKUs.

They can comfortably go and say yes to new business instead of being locked into whatever million-dollar complicated custom equipment that they bought that doesn’t serve any other purpose.

Chuck: Yeah. You had mentioned in the past, I think you used the analogy of United versus Boeing, where Boeing builds the planes, but you’re United who runs all the safety, the security, the customer service, all that, right?

Yeah. Awesome. And, I want to get into the AI and technology a little bit. A lot of what our portfolio companies are seeing right now is certainly this AI coming into their worlds. Robotics is something that is an area that’s very interesting from a is there longer term defensibility, with it relative to the SaaS apocalypse that’s happening potentially in the markets and whatnot. Can you talk a little bit about how robotics is right now from a physical AI standpoint, where that’s at?

How do you think about training these systems and moving ​‘em in? ​‘Cause that polishing the friction off obviously if this is a factory that’s been around for 50, 60, 70 years, it’s maybe an older workforce there who’s not as leaning into what’s new given that they’ve had this same machine that’s lasted ​‘em, 50 years, and they’ve gotten a lot of use out of it.

How do you work with these companies and help them understand what’s possible now?

Saman: Yeah. One thing that’s really interesting about this kind of subscription-based robot model is that the customer actually now doesn’t care what equipment you bring, right? You say, I’m gonna solve this problem for you. I’m gonna pick five boxes per minute, or I’m going to package these things 10 at a time.

Whatever that performance commitment is, that’s the only thing that they care about. How we solve it is our problem. So if we wanna use a classical rule-based robot, if we wanna use computer vision, or we wanna use one of these modern vision language, action model-based robots, they’re agnostic. 

Chuck: And they get that. That’s not a foreign concept for them. They’re used to that kind of…

Saman: It is a very foreign concept. 

Chuck: Okay.

Saman: Because what they’re used to is huge asset purchases, right? So, if you buy the wrong forklift, you’re in trouble, right? That’s a big problem because you bought a forklift that’s rated for 10,000 pounds, and you have to pick up something that’s 15,000 pounds.

Now, you have this giant $150,000 investment that’s worthless. So that’s a huge problem. On the other hand, with this subscription model, it’s new, right? Because we say, look, we can actually abstract all the complexity for you. We’ll manage it, we’ll maintain it, we’re on the hook for the performance.

It allows them to imagine taking on much more complexity. And now they come to us all the time and they say, Hey, if you have anything new, bring it to us to test it. ​‘Cause we wanna automate as much as we can.’ So then is the other side of your question, which is the current state of robots.

So the current state of robotics is there are basically three camps, I would say, or three waves of robotic technology that are happening right now. There’s the first wave, which is the traditional rule-based robots. This is where you deterministically program a robot to do a specific thing. And traditional deterministic robots are highly reliable.

But highly inflexible, right? To make it do something slightly different takes a lot of work. And if your product has any variability to it, you’re screwed, right? Like it doesn’t know how to account for it. Good for highly repetitive tasks, not good for flexibility and adaptability, but this is the most mature style of robot today.

Style number two.

Chuck: And that’s Coca-Cola, General Motors….

Saman: Think about these robots that you see in Ford or General Motors, like they’re welding the same car over and over again all day, every day. Standard. The second bucket of robotics is what I think of as a hybrid approach, where you use some form of AI to understand the world, make a plan, and then you give the robot that robot a plan to execute.

So for example, a computer vision system, which identifies the orientation, size, shape, and placement of an object. And then it creates a plan. It says, Hey, robot, go to this position, pick the thing up, move it to that position, put it there. And here are the kind of sequences of things that you need to manage and monitor during that process.

So the robot is following a fixed sequence, but that sequence is live-generated in the moment using these AI tools. So that’s a more mature style of robot. It’s not widely deployed today, but it’s starting to become more deployed today. That’s a lot of what we’re deploying. And then there’s this third.

Chuck: And what sort of use cases are those?

Saman: So generally, pick and place use cases where you’re picking an item out of a bin that’s haphazardly placed there. 

Chuck: So if you were picking an item out of a bin and stacking it into a whatever, six-pack, case, or something like that?

Saman: Exactly. Or putting it into a big machine or that part may be a slightly different size, you might have some variability in that part. Even a lot of the welding tasks are now more in this style, where you don’t have to fixture those two pieces of metal to be exactly in this position.

They can be anywhere on the table, and it’ll still find the seam, and it’ll know how to weld it together. There are a lot of things like that where you can remove a lot of the complexity and remove a lot of the customization. Then there’s a third, latest kind of model of robotics, which is characterized, like most of them are called vision, language, action models.

Although there are some variations on this, which is essentially, think of it as like you teach ChatGPT how to control a robot, right? You give it a bunch of training, you give it camera feed, you give it some instructions around how to control the hardware, and with a lot of cycles, eventually, hopefully, it gets good.

Chuck: What is a lot of cycles? How many?

Saman: To train the baseline, you need like hundreds of thousands of cycles to at least give it a little bit of instruction. And then today, at least the state of the art vision, language, action models typically then require on a task-by-task basis, another 15 to 20 hours of training on that task. So you have a human, basically, remote-controlling this robot a hundred times. And then after you remote control it a hundred times, you save all those episodes, you train the model again, and now it’s able to do that task semi-autonomously. In that situation, the robot could follow any trajectory to get to that task, right, to pick up that item. You can place that item in any abstract way, or you can even, it could be a new item, it’s never seen before. You just describe it with text. You say, pick up the green pen and it’ll find the green pen, and it’ll pick it up. This third style of AI for robotics is I would call it today very much at the research phase.

There are no commercial deployments of this type of robot, but there are like trial test cases. They typically operate at 60 to 70% accuracy, which is good enough for a demo, but not good enough for a production environment, right? But it’s maturing very quickly.

And the way I see it, I think we are on this gradient of maturity. And I’m sure this kind of vision language action model breakthrough is not gonna be the last. There will be additional breakthroughs that come over the next three to five years that will allow robot capabilities to increase.

So on the one hand, all of this is great news. It means robots are getting smarter and they can do more and more complex tasks without human support or human intervention. There’s the other side of the problem though, which is, this makes the problem that much harder when it comes to the deployment side because if you think about yourself as a factory owner or somebody who is running a chocolate chip cookie factory it’s not trivial to stay up to date with all these changes, know which ones to choose, know which ones to use, which ones to invest in versus not.

And so there’s a bunch of complexity that they’re faced and they’re gonna be asked to face a much larger, new set of complexity. It’s if I buy a piece of hardware that’s on the traditional robots, can I add a brain to it and allow it to be smarter? Can I add a sensor to it, or can I add a camera to it?

What are the things I need to add to it to allow it to have more capabilities, more flexibility, or is it fixed in its performance? What about safety? What about compliance? What about all these other factors? By managing all of this for them, they don’t have to worry about it.

Just when you and I go buy a flight ticket, we don’t have to worry is it a Boeing 747 or a 787? We know the ticket will get us from A to B. Or if we use Gmail, we don’t need to worry what server is it operating on?

Is it in the data center in Arizona or is it the data center in California? It doesn’t matter. Lemme just get the benefit, right? Lemme just use, get the part that I care about. And I think that is the world that we’re heading into with robotics is people need somebody to take on the responsibility, take on the complexity, take on the ownership, take on the management of these complex assets and allow people to get the benefit from ​‘em. And that’s how we drive adoption.

Chuck: Yeah. And that’s basically, again, being the United where you’re using other people’s parts and products. So you then layer your own software on top of all that. So you’ve got, off the shelf, maybe it’s too general of a description, but you get these not super unique hardware solutions that you then couple with the services that you bring to the quality of the software, whatever, additional safety you might add, what additional functionality you might add to systems itself.

Talk a little bit about that, and is AI used in that, or is that really just traditional software that you’ve been using for years?

Saman: Yeah, great question. So I think we essentially built all the scaffolding around the robot, right? So there are really two pieces of software that we’ve built that allow us to do that. One is what we call Formic Cortex. So Cortex is the software that allows us to automate all the configuration, fleet management, and orchestration of those robots. So we can evaluate the factory, evaluate the task, choose the robot, simulate the robot, validate the robot, and test the robot in simulation.

Chuck: So that’s like a digital twin type of thing? 

Saman: Yeah, and then it manages all the configuration of that robot, the physical robot. So it tests all the hardware components on that robot. Make sure everything works. It puts the code on the robot, it manages all the deployment process, and then all the fleet management, orchestration, and management of that robot once it’s deployed.

So all ongoing maintenance and monitoring and service, all of that kind of stuff. So that’s really critical infrastructure and applies to every robot that we build. It allows us to basically deploy these robots for ¹⁄₁₀ the price and ¹⁄₁₀ the time it takes a traditional way to do it.

The second software that we built is what we call Formic Core. And Formic Core is actually like the brain of the robot itself. And so Formic Core is what does some of the path planning. It does some of the organization, it does some of the computer vision tasks. It manages whatever model we use in the backend to actually process the camera feeds.

It also does all of the error handling. Any roboticist will tell you, getting a robot to do something is 10% of the work. And then error handling is 90% of the work.

Chuck: That’s funny.

Saman: Because all of the complexity comes down to managing all the different edge cases. Hey, the box was bigger than I expected.

The surface material was more porous than I expected. The conveyor moved faster than I expected. Like all these things that you hand-tune the first time around. In practice, in production, in reality, they’re always changing, right? There’s a beam of  ight that shone right on my sensor and screwed up the reading.

What do I do? And what you can’t have is a robot that just shuts down and gives up and says, sorry, I can’t do this thing for you. It has to observe the error. Create a response plan and then get back to production ready state, even though it was interrupted somehow. That is really a very complex set of asks.

And it also requires hardware management. So, for example, this joint, if I’m picking up this robot, my shoulder joint has a lot more friction than I expected. So one, maybe I have to change the way that I control that joint. Two is I need to flag a notification to my service team to come and fix it before the robot goes down.

We are very proud of the stat that we have a 99.96% uptime across our entire fleet of hundreds of robots over the last 12 months. And I can’t tell you how hard it is to keep that number at that level. Because you have to basically encounter every possible type of error and then recover gracefully from that error.

Chuck: Yeah. Yeah. The whole idea of I call it polishing the friction off the process. But the idea that you are not just giving them a piece of equipment and saying good luck, you’re actually the one going in and solving the problem that they have, which I think is such a powerful, and it’s not the first time that business model’s ever been used.

But it is such a powerful way to be able to really attack an industry that there’s lots of friction against doing this. Just because of, ah, it’s gonna be too much, my people aren’t gonna understand it, or it’s gonna disrupt my business, or those types of things.

But really changing that model from selling just pure products to pure solutions, I think is such a smart way to build it in this industry. 

Saman: Yeah. Actually, last time we talked, you mentioned this idea of polishing the product and polishing the sales process. Something I’ve used a lot internally with our team now. ​‘Cause that’s exactly right. There’s always little sharp edges that things can get caught on.

And constantly thinking about what are ways to streamline this just a little bit more, make it a little bit easier, make the deployment experience better, make the sales process better, make the management of these assets just a little bit better. That, that polishing I think is often underappreciated by technology entrepreneurs. Even myself, when I started this company, I thought, oh, there’s a couple of big breakthroughs and flash in the pan moments, and now all of a sudden we have a great product.

And to your point, I think the moat of every business ultimately is the number of those little edges that they’re able to polish off. ​’cause there’s just tiny little pieces of experience that come with every failure and every snag.

Chuck: That’s awesome. Good. One of the things that has been interesting is just training data, you hear a lot about it. These large language models have been training on years and years of text, years and years of code. But there’s not a lot of training data for physical AI out there.

You talked about 20,000, repetitions or hundreds of thousands of repetitions. These things need to go through. How do you skip the line and jump ahead in that and do that in a competitive way.

Saman: That’s a great question. I think we look at an analog to learn from, and really the analog is the difference between two companies, Waymo versus Tesla. Waymo took this approach to generating data, which was, let’s go spend a billion to two billion per city to map every square inch.

Use that to create a training set that we can use to train our cars, and then hopefully our cars can eventually generate enough revenue to recoup that cost. And hopefully some of that is transferable to the next city, and the next city, and the next city. In reality, every new city has a bunch of new training data that they need to generate.

It does get less over time, but it’s a model that’s well-studied. Waymo now has cars operating on the road, and they’re doing a good job. But it cost a lot of money to get there and many years. The other example I think is more something to study is Tesla. And Tesla said, Hey, look, we are gonna just sell you cars that you wanna buy, not because of the AI, but just because they’re good cars.

And, millions of people bought Tesla’s cars, and people used those cars to generate training data. So instead of paying a couple of billion dollars to get training data, Tesla got paid a few billion dollars to get training data. And so that reversal of capital flows really allowed Tesla to get a much larger order of magnitude of data, and that allowed ​‘em to train better robot better AI models, which translated to better cars, which now translated to full self-driving in their cars.

I think a similar approach is going to happen in robotics, where there are companies today that are spending ungodly amounts of money to try to observe humans and synthetically go and create large data sets. I think that’s a very expensive proposition. There’s the other alternative, which is I think what we’re endeavoring to do, which is we are deploying a variety of different kinds of robots.

Some of them are deterministic robots, some of them are that kind of level two that I was talking about. Some of them are level three. But our priority is to get as big of a fleet deployed into the real world as possible. And in the real world, we will on the one hand, generate revenue from those robots.

On the other hand, we will create a training set that includes real-world edge cases, failures, and unexpected situations. And that allows us to build a much, much more robust set of training data that we can use to improve the robotics AI models. I don’t think there’s gonna be a single winner in this space.

There’s gonna be many winners, and I think our approach is one of many, I’m sure. But we, fundamentally, I believe, to build a sustainable business, it is a prerequisite to actually building a large and meaningful data set.

Chuck: Very cool. And switching gears a little bit financing strategy. So you’re one of our portfolio companies. Tell us a little bit about robotics companies and what they’re doing these days, and how did you know you finance what you’re doing? The capital you raised from us, like how is that different than what might be happening in the broader industry right now?

Saman: Yeah, I think the big mindset shift that a lot of people need to take on is to not treat robots as some novel gimmick, but to actually treat it as …actually Travis Kalanick recently wrote an article last week ​’cause he started a robotics company recently, and he uses this term that I really like, which is a gainfully employed robot, right? Which is the robot has to generate revenue for it to be useful. And so if you think of each of these robot assets as a gainfully employed robot, as a robot that has an ongoing revenue stream attached to it, you can actually start to think of it as an asset that you can build and improve.

And Formic really takes this mindset of as a asset management business in a sense, like we’re managing this asset of robots, what are the levers that we have at our disposal to allow scale of adoption? Some of them we already talked about. Technology’s a big one, right? We can use technology to make it cheaper to build these assets.

We can use technology to make it cheaper to manage and maintain these assets. We can use technology to make these assets more applicable and more use cases, so we can increase the utilization rate of these assets. But another really important factor is how do you actually build a capital stack to finance this new class of asset?

And there are different levers you can optimize for, as with every capital stack, right? Equity is generally the most expensive and the most flexible form of capital, right? You can raise it, and then you can use it to do all kinds of things. You can do RD, you can hire people. Then you have increasing levels of structure that you can build in terms of the capital availability that come with their own set of requirements, right?

There’s venture debt, and you can use it for certain things. Then there’s asset debt, which you can use for certain things. Then there’s asset securitization that you can use for certain things. Then there’s third-party flow or selling off asset securitization and future revenue. There’s working capital. So you actually have to build a really sophisticated financial infrastructure to be able to efficiently build and deploy these assets at the scale that makes sense. We are a startup. We’ve raised a certain amount of equity. It’s not an unlimited amount of equity, as I sometimes have to remind my team. But the equity that we have is mostly best used on the high-risk, high-return activities, right? Scaling the sales team, scaling the R&D efforts, things like that. It’s not best used on investing in an asset with a highly predictable return profile. Like you don’t take a bunch of venture dollars and then go buy bonds with it, right? You, for things that have a highly predictable future revenue stream associated with it, there are forms of capital that are designed just for that. And so that’s really one of the things that we were proud to partner with S2G on, and we were able to build this asset securitization facility. But this ultimately allows us to have more flexible capital that allows us to go and fund these assets.

It’s designed specifically to look at the expected future revenue from these assets rather than looking at the limited equity that we have as a company to underwrite against. And it’s a much better thing to underwrite against because you have these robots that are worth, let’s say, just making up numbers, like a hundred thousand dollars each.

But then each of these a hundred thousand robots actually has five or $600,000 of expected revenue coming from it. And so when you put that package together, you can say this is actually a decent package to underwrite against, and it gives us the funding to go and buy and build and deploy and scale that fleet very efficiently.

Chuck: Yeah, it’s a cohort analysis that a lot of firms do for software and whatnot, but really looking at it from a hardware standpoint and what is the lifetime value that product versus what did it cost you? And then, how to go maximize the gap between those two.

Saman: Absolutely. Yeah. Yeah.

Chuck: And so pulling back even a little bit more, you at the beginning of the conversation talked a little bit about you, the US having the second largest install base of infrastructure from a manufacturing standpoint. But you’ve also mentioned this idea of American factories needing their AWS moment.

What do you mean by that? And what do you think gets us to that AWS moment?

Saman: Yeah. So if we look at AWS, AWS I guess it first started as excess capacity for Amazon, right? They had all these servers, they didn’t know what to do with it. Then they realized, actually, there’s a bunch of little companies out there who don’t wanna go and spend a bunch of money building infrastructure.

So I’ll just provision AWS as infrastructure on a case-by-case basis. So it started out with people who couldn’t build their own infrastructure. But what happened was once that flywheel started to turn, and more and more people started to use AWS, it became cheaper and cheaper for AWS to go and build and scale their data centers, their servers, they could build microservices on top of it.

They could make it easier and easier for you to use to the point where eventually it got so good that even a company as big as Netflix said, Hey, I don’t want to go and build my own data center. I’m just gonna use AWS. And I think that’s really the world that we are in robotics today.

The very large companies, Ford and GM, for example, do have robots deployed, but they do it the same way that JP Morgan used to have their own data center, right? But then, all the little manufacturers across the country have not had access to automation. So what we’ve built for them is a way for us to take on that complexity on their behalf as we’ve scaled our business by investing in the common solutions that everybody needs.

We can invest in new software, we can invest in new infrastructure, we can improve the robots. That allows us to get more and more efficient at deploying these systems. To the point now where we actually have a lot of very big companies that have chosen to work with us. That have large in-house engineering teams, but they just know that they can’t deploy, build, and deploy robots as efficiently as we can do it. And so there’s a lot of benefits to the economies of scale that come from solving this problem again and again for hundreds of different manufacturers. And I think this will be a big unlock for American manufacturing once we get it right.

Chuck: Yeah. And what causes that AWS moment? Is it a moment or is it on a scale? And it just, one day we wake up and look around, and suddenly we’re in an AWS world from a robotics standpoint?

Saman: I think we’re already well on the way. We have hundreds of factories now that are using our robots, so I think, at least for them, the AWS moment has arrived. And I think it’s our job to make sure that goes from hundreds to tens of thousands to hundreds of thousands of factories.

Chuck: Yeah. So when we first started the conversation, you talked a lot about sort of labor, and there’s conversations about robots taking jobs and AI taking jobs. What’s your perspective on that? Is that indeed the case? 

Saman: Yeah, I think I can say pretty definitively that it’s not the case. And it’s funny ​’cause this question of robots taking jobs has only come up to me, from like reporters and things like that. In manufacturing, it’s blindly obvious to every factory that we talk to that there’s just not enough people available to work.

And we’ve deployed our robots in hundreds of factories at this point. I don’t think a single one of them has ever fired people because they added a bunch of robots. If anything, they’ve increased. And there’s actually data to back this up from a bunch of economic research that’s been done over the last 10 years.

Factories that deploy robots are much more likely to retain their employees, are actually much more likely to hire more people over time, and are actually more likely to be profitable in the long run. And factories that don’t often are on this kind of slow decline towards obscurity.

I think as robots go into more places, as robots get better, it is something that as a society we need to continue to think about. But I could speak from personal experience. I’ve deployed a bunch of AI agents for myself and for our team at our company. And so these are like AI robots, right?

They’re not physical robots. And of course, I think everybody expected that when you deploy a bunch of AI agents, you’re gonna fire a bunch of people. The reality is that all of us are much more busy as a result of having this giant, AI workforce available to us. But we’re doing much higher value work, and we now, as an organization of 70 people, are doing what used to take maybe two, 300.

And I think that ratio is gonna continue to increase. I expect us to be very close to a one-to-ten ratio soon, in terms of each of us should be doing at least the work of 10 people a few years ago. And I think this idea of robots taking jobs or AI taking jobs, really discounts the human ingenuity and our ability to invent new tasks to do.

If you ask any business owner, what would you do if you had 10,000 more people available on your team? All of them probably have ideas after ideas of things that they would wanna do. I have a to-do list. I was just joking with my team today. The to-do list only gets longer, right?

Like my wishlist of projects only gets longer. And I think it’s the same for every factory. There’s more that they wanna do, and having these kinds of robot minions available to us is only gonna allow us to accomplish more.

Chuck: Yeah, and I have some agents running around doing a bunch of work for me as well. And it’s changed what I’ve done, but it’s also I think, made me more busy than I’ve been in a long time, actually. I’m actually having a lot of fun experimenting. Some agents don’t quite cooperate, some people don’t quite cooperate. 

Saman: You always have the disobedient agents.

Chuck: Yeah, exactly. Good. So final question where do you see manufacturing, and maybe specifically manufacturing in America, in the next five to 10 years? Where’s it going? Will see different business models? You guys have a robotics as a service sort of RAAS model.

What do you see changing over the next five to 10 years in manufacturing? Because  roles and responsibilities, I think that people will be just as busy at companies, but the roles and things they do will definitely change based on the new tools and capabilities that are out there.

What’s your vision, and what do you see happening in manufacturing?

Saman: That’s a great question. I think that in America, we have some catching up to do first. That catching up is really more in the form of building the basic building blocks. Where America has atrophied the most is in building simple products. So nuts and bolts and screws and windshield wipers and buttons and all of these things.

And what I think a lot of people don’t appreciate is that you can’t build complex products, advanced products without being able to build basic products. So if you look at China, China went through these multiple five-year plans. Every year they release a five-year plan, and in each of those five-year plans, they built a certain set of industries.

So it started in the seventies with let’s go and build basic steel processing, right? Steel mills and mining and things like that. Then they work their way up to some level of processing, some level of machining, making simple assemblies, making toys, making garments, making simple, secondhand parts for components for automotive and things like that.

But then in the last two decades or two it started to really significantly accelerate. Right now, they are able to produce chips. Now they’re able to produce automotive EVs. They’re able to make solar panels, they’re able to make aerospace parts. And at a scale that now in the US, we don’t have the ability to produce.

And so the manufacturing industry is this pyramid where you start at the bottom and you work your way up. And I think in the US we had built this pyramid. And then we focus all of our energy on the top of this pyramid, but we let the bottom of the pyramid atrophy and fall away. And what happened is as we lost capability and infrastructure in that bottom of the pyramid, the top of the pyramid got weaker. Right now for us to, we can make EVs in America, but it costs triple the price of BYD. We could make solar panels in America, but it cost double the price of somewhere else.

We can make batteries, we can do all these things. We have the kind of technical know-how here, but we don’t have the infrastructure to support it. So I think the next five to 10 years are really a game of reshoring the foundation, right? We will use robots to rebuild the ability to make basic things, and once we build those basic things, we will improve the kind of entirety of that pyramid.

Then the really exciting stuff starts to happen, right? I think then in the 10 years after that, is when we can use advanced robots to really push the boundaries of our capabilities because scientific advancement is happening quickly. People are inventing new products, new chips, new sensors, new robots, new drones, new cars, things for going to space, all kinds of machinery for mining.

And really, I think we’re at this kind of cusp of a renaissance where we can drastically increase the availability of daily products for the whole world. If we wanna at the minimum, share that standard of living with the rest of the world and hopefully improve the standard of living for everybody beyond what it is in 2026 America. We actually need to produce much more. So we wanna create this world of abundance. And that abundance comes from many more robots building many more things.

But the prerequisite is we need to at least spend the next five years reshoring the foundation of what we suck at, to be honest.

Chuck: Awesome. That’s a great answer. 

Saman, thank you so much for today. Learned a couple of things. One is I probably don’t eat enough chocolate chip cookies, ​’cause that’s you’ve been in a lot of those factories, it sounds like. And then two, just the, where this is going. And super excited. Thanks for your time today, and I really appreciate everything you had to say.

Saman: Thank you so much for having me on a really stimulating conversation. I had a lot of fun.

Chuck: Awesome. 

Tonya: Thank you for tuning into the S2G podcast. If you enjoyed this episode, please subscribe, leave us a review, or share it with a friend. For more information or to connect with us, check out s2gin​vest​ments​.com.