Please welcome to the stage TechCrunch editor-in-chief and general manager Connie Loizos.
Hello. Good morning. Thank you so much for coming. We're so excited to see you. Um, as the voice of God told you, I am Connie Loizos, the editor-in-chief of TechCrunch. It's wonderful to see this room filling up again. Thank you all for coming back.
Um, you know, for more than 20 years, TechCrunch has covered technology through the people building it. Uh, often when their companies were a little more than an idea, uh, a few people and a very big ambition. That's what we have been about. Um, uh, and and really getting people in the same room, founders, investors, engineers, operators, and sometimes people who don't fit neatly into any of these uh categories.
It is of course a very interesting moment to be doing what you are doing right now. Uh, AI is changing how companies are built, how fast they scale, what they cost to start, and even what it means to be a founder. Uh, and some of the biggest opportunities are moving beyond software, which we're going to get into throughout the next three days, um, into, you know, robotics, biotech, engineering, uh, transportation, and the physical world.
So over the next three days we want to get past some of the hype and ask harder questions. Uh, what are people actually building? Where is the money going? And how do you build an enduring AI company in this kind of environment? You will hear from some extraordinary people on these these stages. But uh, this event is really not just about, you know, people with the microphones. It is the people that are sitting right next to you who could become a co-founder or an investor or just a friend uh to bounce ideas off of.
And I hope you will take advantage and introduce yourself over the next three days to the people uh at this event. Um, before we get started, one related announcement. Um, you will meet people here that are as deep into this world as you are, but we want this to become more of a year-round um possibility for for all of you.
So um, we are launching today something called the TechCrunch Innovation Board. It is a membership community for founders, operators, and venture capitalists. And it's built so that you can find each other year round, whether you need a a specialized service or somebody again who just truly sort of gets you and you can bounce ideas off of. Um, members can contribute pieces.
You can get special uh access to our events and hopefully eventually you'll be able to sort of also buy enterprise products uh that make your work easier. You can join the waitlist right now starting at techcrunch.com. Uh, that is techcrunch.com or just look for the QR codes. They're going to be all over the venue and you can scan one scan one with your phone and sign up for a spot.
But we will have much more in the coming months and we hope that you will join. Um, again, glad so glad you're here and with that I'm going to hand things over to the person who you're going to be seeing a lot over the next three days, my colleague Anthony Ha. Thanks guys.
We've been advancing energy technology for almost two centuries, powering innovation and shaping the way we live, work, and thrive. From the first fuse to powering the future of quantum computing, Schneider Electric is and always has been your energy technology partner. We electrify, automate, and digitalize every industry, business, and home, driving efficiency and sustainability for all.
From two brothers in a foundry to a global force for innovation, we have spent 190 years advancing energy technology for people everywhere. We electrify, automate, and digitalize industries, businesses, and homes. And we're already building what comes next as your energy technology partner. Join us as we celebrate 190 years of innovation.
Hey, I'm your dot. I'm new in ChatGPT. Ready to jump in when you need me. I can proactively keep things moving, like launching your small business. Okay, perfect. Getting your app rebuilt, tested, and ready for rollout, or even getting you prepped for your board meeting. Go ahead and update it with the latest numbers, please. Love it. Let's get to work.
Please welcome to the stage weekend editor and Battlefield host Anthony Ha.
Hello and thank you Connie. Thank you all for being here. Welcome to TechCrunch Disrupt and to the Disrupt stage. I am Anthony Ha, TechCrunch's weekend editor. And now uh, Connie covered the bigger picture, but I'm just going to do a little bit of housekeeping, a little bit of logistics so you understand how to do the next few days.
So we've got six stages running across Disrupt including the Builder stage, AI stage, Real World AI, Smart Money, and Smart Systems. So don't spend all three days sitting in this room. I mean you can spend a lot of them sitting in this room but there is a lot happening elsewhere.
You'll also find Startup Battlefield pitches, breakouts, roundtables, networking opportunities, and the expo hall throughout the building. The Disrupt app is your best friend for the latest schedule, room locations, and any updates. Now, if you're wondering about Wi-Fi or the TechCrunch events app, you can look on the back of your badge.
Also on your badge, there's a small device called Klik, K-L-I-K, that you can use to exchange contact information with other attendees via a simple click. Once you've clicked, the contact information will be uploaded to the app. Now, at the end of Disrupt, remember to recycle and return your badge before you leave because then you're entered to win a pair of Meta glasses, a Theragun, or a free Disrupt pass for 2027.
And there is one request from all of us at TechCrunch. Participate, ask questions during the Q&As, talk to our speakers afterwards, meet the founders exhibiting here, and introduce yourself to people you don't know. Um, obviously there's a lot going on on all the stages but the more you know the energy of Disrupt comes from the people here in this room. So the more you can kind of give I think the better experience it'll be for all of us.
Uh, we'll also remind you from this stage about different networking events and evening events as they get closer. So we've got three packed days ahead of us. So let's really get started. You're currently in exactly the right place. I mean, I'm a little biased, but I I really do think you're in the right place because it's time to start with the Startup Battlefield.
For more than a decade, Startup Battlefield has introduced the Disrupt audience to companies at the very beginning of their journeys, including some that went on to become household names. Today, a new group gets its shot. We're starting with founders building the next generation of space and physical AI technologies. Each company gets six minutes to pitch, followed by six minutes of questions from our judges. Our judges will choose five finalists who will then be competing in the finals on Wednesday.
And with that, let's bring out our judges. First, we have Anu Aiyengar, partner at General Catalyst. She was most recently president of Atlassian, a leading collaboration software company serving millions of teams worldwide. Over her decade at Atlassian, she helped scale revenue from $200 million to $5 billion and grow the customer base 10x to over 300,000.
Are we sending them out or... Okay, I'm just going to keep introducing them and things will happen behind me, I assume. Um, next we have Sarah Dash Pandey, general partner at Maven Ventures. Sarah is a uh general partner at Maven Ventures, a seed-stage venture firm investing in tech companies addressing emerging consumer behavior and trends. Sarah joined Maven as the first employee in 2014 and has been a VC for 10 years.
Next, we're going to bring out Ben Narasin, the founder and general partner at Tenacity Venture Capital as well as a prolific entrepreneur. He founded several consumer companies before launching his investing career, including fashionmall.com, one of the first e-commerce companies, which he launched in 1993 and led to a successful IPO in 1999.
Next, we have Atish Nayak, partner at Kleiner Perkins. He joined Kleiner Perkins as partner in May 2026 where he focuses on partnering with AI-native founders with deep vertical expertise across all domains. Prior to joining Kleiner Perkins, Atish was the first PM and VP of product at Harvey where he helped build the early product design, marketing, analytics, and support teams.
Next we have Mike Palihawadana, general partner at M13. Uh, he's been a general partner at M13 since its inception. At M13, Mike focuses on seed-stage consumer and enterprise startup investing across a number of sectors. Now, let's bring out all of our judges.
Okay, so we've got our judges obviously. Next, we need to bring out our startups. So, our very first Startup Battlefield company for 2026 is Longshot Space. Presenting for Longshot Space is co-founder and CEO Mike Grace.
Elon Musk gets to be a trillionaire. And here we are, the rest of us, kind of rolling around on the surface of a single planet trying to create unicorns and decacorns, scratching out a meager existence like peasants. I find this to be an unacceptable equilibrium. All right.
Um, however, the reason why Elon gets to command this kind of value is fundamentally because he controls effectively 99.999% of all the atoms that humanity has access to. And in a world where we fundamentally do not know where the value of software is going to ultimately occur, it turns out that owning everything is a pretty good business.
I also want to say that for the next 15 years, no one with a rocket is going to be within one order of magnitude of Elon's unit economics for putting stuff into space or his capacity for putting things into space. It's not going to happen unless you're the Chinese Communist Party because they will subsidize their own rocket creators to the tune of a hundred billion dollars. They will do that. So they might be able to catch up.
Jeff Bezos doesn't have the money. Nobody's going to give it to Peter Beck. It's not going to happen. So we are heading into a future where the solar system is much more open than it has been, which is great, but effectively the province of one person and the rest of us get to rent.
If we want this to go other than the way it is currently going, we have to do something that is fundamentally different.
My name is Mike Grace. I'm the CEO and co-founder of Longshot. I have the largest gun on the planet Earth inside of a shed across the water in Alameda right now. Uh, this is me, my co-founder, looking absolutely dashing astride the monstrosity over there. And this is a tiny demonstrator of what we are going to be building next year out in Barstow. We're going to have one of these operational in summer that's about half a kilometer long. There's a reason the company's called Longshot.
Big picture here, right? Elon has managed to bring the price of putting stuff into space down to a level where it begins to make sense to push things into orbit that were previously considered the exclusive purview of ground-based infrastructure. Right? $300, $500 a kilogram. Maybe it makes sense to put GPUs into space. That's a lot of not having to ask people for permits that you could do, right?
The case that I want to make to you is stretch a little bit further. If you could push the price lower of putting material into space, I think there's a threshold where it doesn't make sense to mine copper on the surface of the Earth anymore.
How much of the human economy belongs in space? I would argue all heavy industry. The Earth should be a garden. Just push everything that creates mess. The only thing that is unique about this planet that we live on today is that we evolved here. This is where life is. Everything else is sterile. Take advantage of it as such.
The way that our system works is it's a multi-injection gun. No magnets, no big centrifuge, none of that stuff. Instead of boom, my system goes bump, bump, bump, bump, bump, bump, bump down the length. Every time I double the length of my system, the maximum pressure, the maximum temperature, and the maximum force of acceleration being applied to the payload cuts in half.
As you make it longer, you're being more gentle both to the accelerator and to the payload, which is critical. I can build a 10 km long pipeline, but I can't point it up. The vehicle that you can see right over there, little art model of it. I actually rolled that myself out of stainless. Super fun project.
Um, this is a hypersonic lifting body. You build up momentum, you hit the atmosphere, and then you sacrifice some of your velocity for lift. This is actually technology developed for boost-glide weapon systems. But I'm not coming down, I'm going up.
Vehicle skips up over the atmosphere. Second stage solid rocket motor pushes you into orbit. The shell that protected you re-enters, pull it out of the ocean. The rest of it's disposable, at least initially.
We're awesome. I'm awesome. You can tell I'm amazing. Um, I was a paratrooper. Went to Afghanistan a couple of times. I worked at NASA for a bit. Um, I'm actually an economist and a molecular biologist by training. I'm all over the place, folks. My co-founder is the one with the true Silicon Valley bona fides. He's built rockets. He's built satellites. He was at Neuralink, etc.
We've been buddies for 15 years. He was building um satellites for Planet Labs in the garage of the house I was living in. And then um uh we became good buddies, known each other ever since. I managed to get a whole bunch of august investors in. Um, uh, there are a couple more coming in right now who I probably shouldn't mention on stage.
Right now, I've raised about 20 million bucks and uh there are a whole bunch of people who are panicking because they realize that Elon's going to eat his own dog food forever. He's actually not. He can make more money launching his own stuff. He's not going to launch other people's satellites. That's the way it's going to go. He's going to consume his capacity.
There's a whole bunch of people who are worried about how they're going to put stuff into space. And I have a lot of LOIs there. I've also got about $15 million in hypersonic technology development bookings with the Department of War.
I don't want to build something small. You might be mistaken in thinking that this is for cubesats or something like that. Put that aside. You must throw massive things to orbit or it doesn't work. Also, I want to build an accelerator that's long enough that G-forces come down to about 200. Your phone is designed to tolerate about 700. There will be a design burden on satellite makers to meet this system, but um it's a solvable problem at 200 Gs.
Um, I'm currently raising a Series A, so if you want to shower me with money, knock yourself out. I'm at your disposal. Um, hey, you know, I happen to have something down in Barstow I'd like to share with you guys right now, which is the world's only operational multi-injection gun. Can we switch over to that?
Uh, I'd like you guys to help me out with a bit of a countdown real quick. Okay. All right. We're going to go from five. All right. Five, four, three, two, one.
Lift off. Blast off. Oh, there she goes. We don't have any sound on this one. Uh, and I decided not to do anything inside of here as uh turning investors into paste is usually not how you get them to invest.
So, this is currently at our test facility down in Barstow, and we're going to have a 30-inch interior diameter gun operational um December. I'll stop right there.
All right, judges, questions.
Two quick ones. Um, there's been a lot of speculation over time about a space elevator. Why, if you want to get stuff into orbit cheaper, why wouldn't you pursue that? I would assume it's doable with graphene. Two, if you lose velocity by designing it the way you design it now, why wouldn't you put it on a material incline like a mountain and therefore already be pointed in the right direction?
Space elevator, if um you start seeing long carbon fiber in things like bulletproof vests or building materials, that's a prerequisite. It'll happen on the surface of the Earth in a whole bunch of applications, but basically the material science isn't there by a long shot.
Um, so that's why space elevator. Uh, Hitler built one of these in World War II. My goal is to steal technology from the past that has been proven and concatenate things together, not to do something new. You can do something hard or you can do something new. Don't try to do both. Right.
Um, with regards to your second question, putting it on an incline instead of a flat plane.
The advantage of being 10 kilometers long, flat in the ground. There's a lot of flat desert in the world, Atacama Desert, middle of Australia, Nevada, right? But the second I put it on a mountain, I'm limited to about 2 km long, and you multiply the expense of the project by a factor of like 100. It turns out building a pipeline flat on the ground in the desert is cheap and allows me to go to an enormous scale.
Um, so that's why, just build it long and let the vehicle do the lifting maneuver. Though you're correct, it does cost some velocity.
What do you envision using this for? I have to imagine the level of force um isn't going to work for some payloads that might be more sensitive to that type of pressure and speed. What would your customers want it for?
That's about 600 Gs. 600 Gs when you go from terminal velocity to stopped on the ground. So basically all electronics that are designed to be handled by human beings are designed to actually tolerate like hundreds of Gs out of necessity. So like semiconductors, solar panels, GPUs, all the things that are inside the anatomy of a telecommunications satellite or space-based compute, super tolerant, like upwards of like 600, 800 Gs.
Although you will have to re-engineer for the packaging. The packaging and the orientation of things then matters in that case. So I think uh large optical instruments, probably not, right? I think, and also the larger the satellite is, the more mass it has with which to crush itself. 20-ton geostationary sats, those probably belong on rockets.
Um, I think that dense semiconductors though, which is where I think people are going to be making 99% of all money in space over the coming years, that's this.
What's going to be your lift capacity? Is it going to be a scaled approach? Will you do light lift, medium lift, heavy lift?
I currently define an MVP as the thing that can go toe-to-toe with Starship. That's what the MVP is for me. And in terms of unit economics for putting a ton into space, right? That's about a two-ton payload and a 15-ton vehicle. That's a 15-ton version of that right over there.
It's like a, by the way, it is like a bullet and an airplane had a baby. It's not really built the way you would build a typical flight article. It's massive. That thing's going to be stainless steel. There's going to be a lot of stainless in our system.
Not meant to be reusable, but I imagine a much lower cost to build out your...
Right. It's a bullet. It's relatively simple. No turbopumps, no nothing. But the shell is actually designed to be reusable. There will be an ablative phenolic heat shield that you have to reapply, but the structure and everything that's keeping it together, that actually can fall back in. And if you can survive the exit, it's much more violent and hot than the re-entry.
And are you building the second stage as well?
Um, I want to. I also think there are a lot of people working in solid rocketry right now that I'd be happy to buy from, but I do think a solid rocket second stage is where you start initially. Eventually, we have plans for a reusable upper stage, but it works with a solid rocket, and that's a very simple and easy way.
People have been shooting solid rockets out of guns actually since the 1970s. So, it's a no-go again, minimize novelty if you're doing something hard. Steal it from the Nazis. That is the recipe to success in all things aerospace engineering, right? Don't do new things.
One question for me. Um, what are like the key maybe science or physics challenges that you really want to solve in the next 6 to 12 months to have confidence in this approach, or is it just execution, regulatory, etc.?
We've tried to define the MVP such that there are no physics challenges. That like the heat shield that we need to build is the heat shield that was put on ballistic missile re-entry systems in the 1970s. It's going up instead of down.
I do think the majority of our sort of like the challenges that we have to solve aren't in the accelerator anymore. Like it works. It's in Alameda. It works. Um, the challenge is in you have to have that heat shield. It has to survive. It has to navigate.
And doing that at this, it's a two-stage-to-orbit system. You're not trying to get all the velocity off of the ground. You're getting about two-thirds. But uh, traveling through the lower, thicker layers of the atmosphere at um, you know, Mach 14 to 16, that's challenging.
Although again, when people are attempting to deliver atomic bombs to Moscow and Washington in the 1970s, that is exactly what they had to do. So it really is trying to pull together a bunch of weapons technologies, taking this sword and turning it into a plowshare. Go up, not down.
What part of the supply chain is the hardest to procure and where is it all sourced from?
The great thing about building what I build is that it's basically a civil engineering project. Like the thing that you saw me standing on top of is made of sewer pipe from Korea. Love it, right? I will never touch a drop of Inconel. I will never touch a single iota of rhenium. Forget that stuff.
Um, I need civil engineering supplies and I need um semiconductors just as everyone else does. I think that like the rest of everyone who builds anything, semiconductors, just the electronics to make the thing work are probably the only vulnerable part of it. The rest of it can be made if not in the United States, in North America.
All right, we are out of time. So, give it up for Longshot Space.
I think we're bringing out the card for our next startup. That startup is VeriRitoss. Presenting for VeriRitoss is founder and CEO Lucy Holland.
How would I know if this bag was real or from a sketchy sidewalk sale? Well, you can't until now with VeriRitoss. I'm Lucy Holland, the founder and CEO of VeriRitoss, and we've built secure microchips and software that let anyone verify a luxury product is real instantly.
Luxury brands lose over $460 billion each year to counterfeit goods. And currently, there's no hardware-level way to prove that an object is authentic. The current solutions provided include QR codes and NFC chips that can be easily copied, cloned, and relabeled to counterfeit goods. This causes brands and customers to lose trust in those brands.
As well, the time is now with the EU Digital Product Passport regulation coming into effect in 2027 that requires brands to link their products to digital identities. Currently, there's no secure solution that meets this bar until now.
We've created the hardware-level way to verify authenticity. We've designed a custom-built microchip to be as subtle as it is secure. Our microchip is flexible, washable, and waterproof, and it's easily embedded into garments through sewing in or injecting it into the product itself. This links directly to a digital identity that you can tap with your phone to verify instantly.
We've built this from a team of engineers and designers with experience across software and hardware companies ranging from Tesla, Apple, to more.
VeriRitoss works in three steps. We first embed the chip into the product at manufacture, working directly with the brand. Next, we link that chip to a digital identity, and that identity carries the certificate of authenticity and the brand record. And third, anyone can tap with their phone to verify instantly.
This creates a direct connection to that verified product owner and allows brands to create a custom experience that showcases the product history and care information, as well as that ownership and resale information, and even enabling marketing and post-purchase experiences. So now I'll go to a live demo. My teammate Justin will show you how it works.
This is a silk scarf embedded with VeriRitoss. And with one tap of your phone, it instantly pulls up the information.
On the phone screen, you can see the certificate of authenticity at the top, as well as more information about the specific product, including the materials involved with making the product and more about the brand and its legacy.
This is just some of the experiences and information you can enable with VeriRitoss. And now we'll go back to the slides to show you what the brands can see.
So you just got a view of what the customer can see on their side through the product. And now I'll tell you more about the brand intelligence dashboard. This is your brand's home base to see all the information about the product life cycle. You can see information about verification activity as well as ownership and resale information, receive security alerts, and also that product passport data, and even enabling marketing and event activations for your verified community.
This is a preview of the brand intelligence dashboard and where you're able to see an overview of your products, get insights into that data, and even create custom events for your verified community. The brand intelligence dashboard also lets you create a custom certificate to match your brand.
So where we are today, our iOS app is live and our current chips work across apparel and accessories. And our next-generation chips work for glass, jewelry, metal, and we're very excited to announce one of our partnerships with Alos to authenticate the Tiffany & Co. Championship crowns.
In a post-counterfeit world, brands are protected and customers can buy with confidence. So trust but VeriRitoss.
All right, judges. Who has a question?
One for me. Uh, could you talk about the target customers you're going after and what the go-to-market there looks like?
Yeah. So, we're starting with luxury where the counterfeit problem is the most prominent. So, working directly with luxury brands and brands who want to protect their products.
Um, and then go-to-market. So, focusing directly on working with those brands to integrate the chips into various products including garments, accessories, and apparel.
And then how do you charge?
Um, so we have our software and our hardware. So the software subscription and then the hardware cost.
Do you have any customers so far today? Um, and if your customers choose you, why are they choosing you versus their existing behavior or another alternative?
Yeah. So we're working with various brands across fashion and accessories and fine art. And one of the main reasons they're choosing us over competitors. So first is their hardware. So, we've designed a custom-built microchip to be as subtle as it is secure, including a few key security features.
One being our tamper wire. So, if the chip is removed or taken out of the item, it automatically notifies the brand. Um, and then second, our software dashboard that allows you to view all of the product information instantly. So, that's why brands are choosing us primarily for the hardware and software components.
And can you say more? Are you live with those customers? Like people have these in their bags and scarves today. These are out in the wild with consumers.
How thick is the unit that goes in the product and can it go like, I would think watches are an obvious area but it didn't look like you could get it in a watch.
Okay. And then so in the scarf example, they scanned a very specific place. How do I figure out where it is? The scarf's a large garment compared to the size of your unit. So am I just sitting there like...
Yeah, that's a great question. So, the product itself comes with a card that tells you where the chip is. Um, and that information card, we work directly with a brand to design that.
And sorry, you didn't hear an answer to the watch. Can you put it in a watch?
Yeah. So, our next-generation chip that we just highlighted scans on metal and through glass. And so, that's one that we're excited to bring to the watch industry as well.
When you started the company, were you looking to solve a counterfeit problem? Were you looking to give value to a consumer that might want to go deeper with the product? Was the value prop to the manufacturer, the brand that may want to give more information to the customer or deal with the counterfeit problem? Curious what motivated you to start the company? Was there one particular problem you were looking to solve out of the gate or you saw the biggest problem to be solved?
Yeah. So when I started the company, um, with my experience across software and hardware and also as a designer, I wanted to find a way to protect my own designs. So from that perspective, knowing that brands and designers want a way to protect their designs and create that certificate of authenticity was the core.
And then from there, with that same software and hardware, we're able to enable post-purchase experiences and sharing more about the brand and how they can connect with customers.
Do you sell this directly to the brands or is there a different distribution mechanism?
Yes, so we sell and work directly with brands. So partnering with them to embed the chips during the manufacturing process.
And is the typical competition something they natively build or are there other independent players in market doing this?
Um, so can you repeat the last part?
So is the competition typically something they natively try to produce? Uh, do brands try to produce the custom chips themselves or do you find other independent players competing with you?
Yeah, so the security aspect is something that we bring to brands is we're able to create and provide a secure microchip. Um, the traditional solutions like QR codes or NFC chips can be actually copied and relabeled to counterfeit goods. So a lot of brands are seeing if they work with something like that, it's not a truly secure solution.
So you mentioned both in your deck and just now the competitive landscape and I know there's several companies um that have had a pretty long history of doing this over the last five, six, seven years. Maybe speak to how they're doing. It seems like they're getting good traction and my concern there would be you're selling another flavor of ice cream, right? Maybe you're the best ice cream in the world, but if I already bought chocolate, I'm not buying vanilla.
So if Hermès has already bought, you know, RF chips, it's tough to get them to convert to you. So maybe just talk a little bit about where the competitors are today.
Yeah. So for other solutions that could be integrated with products, one thing again is the performance and security of our chip. And so with other options, they might not offer the same features in depth. For example, being washable, waterproof, or being able to have an extended scanning range and having a tamper wire as well. So in comparison to other solutions, um, from the hardware side, that's where we're different.
I was looking more for how much of the market they've penetrated. Like I know a friend of mine funded the one that I think is a QR code but it's mainly about imported product and detecting um counterfeits at the border when they're being brought through the system. Do you have an understanding of sort of how many brands have adopted those other competitive solutions so far?
Um, as far as the other brands that are looking to adopt this, I would say that especially with the EU Digital Product Passport coming into effect, brands will be looking for a data carrier. And so when brands look to a solution, they want to choose the best solution that provides the most security so they don't have to go back and change it after the fact. So I'd say it's definitely um something that brands are exploring and actively wanting to find a solution for right now.
Okay, we are out of time. One more round of applause for VeriRitoss.
I think it is going to take just a minute to reset the stage before our next startup. So, just sit tight for a second. Um, oh, actually, we're good. We're good. So, let's bring them out. Let's bring out Lelas Intelligence. Presenting for Lelas Intelligence is co-founder and CEO Howie Sun.
Good morning everyone. My name is Howie and I'm coming from Lelas Intelligence. Lelas is building the programmable world models that actively challenge physical AI. Nowadays, AI agents are moving from computers into the physical world. We're asking cars to drive themselves, asking robots to do manipulation and delivery. And before they enter the real world, we want to make sure they're ready for the infinite real-world complexities.
Let's take a very simple task. For example, if we have a robot dog here delivering coffee on stage to me, what's going to happen and what's potentially going wrong? Well, there might be a person stepping in front of it. There might be a chair blocking the path or there might be a huge step with hot coffee on your back. And this is something I can think of right now, but there's more things that we cannot think of beforehand.
And who is concerned about that? Well, everyone. Developers want to know what to train and test for their robots. The deployers want to make sure this is safe to be deployed in my setting. Insurance companies, regulation agencies, they want the safety evidence. But finding these failures in the real world is too hard, too slow, too costly, and too dangerous.
So we turn to computer, we turn to simulation. But it's not a silver bullet. Building simulation and designing tasks is actually slow nowadays as well. I'm coming from Mcity, which is the first AV test track in the world at University of Michigan.
About two, uh, three years ago, we started building what we call Mcity Digital Twin, which essentially means we want to build a 3D environment for our teeny tiny test track and let it be useful for every researcher in the autonomous driving field. It took us around 250, uh, 250k dollars and 15 person-months to have this built, and this is just the starting point.
Because on top of that, you also need experienced engineers to design, okay, what's the traffic light, what are the pedestrians, what are all the corner cases. Often times it will take hours per scenario. And this is just about autonomous driving. Structured environment for robots, this problem got even harder. Different places, different tasks, different people, and we just cannot design everything by hand by human beings.
So, Lelas is building world models plus AI agents. Our solution looks like a loop. We have a testing agent constantly proposing the next challenges for the physical AI agents, and the world model itself is constantly generating and editing the photorealistic environment according to the proposed challenges and so on and so forth. Our loop will get improved over time. Let's take a look at the demo.
Still we are talking about this robot delivery in this stage. You can see the AI agent is constantly building the 3D environment. All the chairs, all the uh people, the robot dog and robot coffee maker, and it's designing, okay, what potentially could go wrong in this situation. And after it figures out a couple of criteria, if we confirm it, we'll go to the next stage.
AI agents starting to design thousands of parallel potential challenges and run the simulation. Well, the robot dog might be blocked at a corner by someone or similarly just a lot of people running across the environment that block totally, like lose the way to the final destination. And also there are potentially some unexpected safety-critical events.
For example, the scenario four here. If we click it, you can see the robot dog is almost there. 95% is coming to me but it's too eager leaning towards me and flips over. This is very dangerous when you have hot coffee on your back. Right? Let's come back to the slides.
So what's the most difficult part of this problem is really about the two components. How can we constantly propose reasonable, realistic, but still challenging environments? And that is where our background and research come from. Herman and I come from the University of Michigan and over the past couple of years have been working on one single problem.
How to use AI to design reliable, realistic, and accurate and challenging environments to test physical AI. Our solutions accelerate autonomous testing by 10 to the power of 3 to 5 times and those were featured on the cover of Nature in 2023.
Currently our paying customers include the top automotive OEMs and autonomous driving companies like Plus, that's an autonomous truck company based in Texas, and we also won the US DOT RPI challenge. We are currently designing our first robotics solutions with our leading uh design partners, the robotics companies, and on top of that we also expand our solutions into general digital twin environments.
Here's a demo. We're working with Nvidia's Metropolis team and it was shown at this year's NVIDIA GTC 2026. We're integrating with Nvidia's VSS agent and using simulation to help the city of San Jose improve traffic efficiency and traffic safety, especially under this kind of corner cases we design.
So this is us and uh if you are building physical AI agents, robotics, autonomous driving, robots, all things like that, please bring us your hardest deployment and let's find where it fails before the real world does. Thank you.
Um, could you talk a little bit about the differentiators for this versus other um kind of simulation platforms?
Yeah. So simulation is kind of like a long-term problem. Like we have been working on, every autonomous driving company has their own simulation problem. And our solution really differentiates coming from one conversation with my customer, a top automotive OEM. They're saying that, okay, I got a lot of 3D simulation environments, but my engineers are spending so much time designing all sorts of different corner cases.
They're good at how to test, but what to test still relies on human intelligence. And our solution is really bringing all these things together to basically make human out of the loop and design and explore this kind of unknown unsafe events beforehand.
So your customer is the OEM but really the end customer is their customer, maybe a robotic company selling into a warehouse company. Is the sales process the OEM and you jointly explaining the value proposition of why your world model is superior and why that makes the OEM's product superior?
Yeah. At the current stage we are uh providing our solution as a development software to the, you know, developers themselves. It's like a software-as-a-service solution, but we are currently working with a couple of uh regulation agencies and insurance companies to really bring this together because we think this can be shared evidence at that stage.
I think that's a good point to, you know, bring all of these things together to provide them trust like this can be deployed safely at my stage.
And is this a, sorry, is this embodiment agnostic?
Embodiment agnostic, like can you put humanoids, drones, dogs, cars, all of the above?
Uh yeah, I mean I think there's a couple of unsolved problems during the process, you know, simulation accuracy, efficiency, and AI agents whether it's proposing the correct thing.
Our solution is a combination of different, you know, generative AI approaches using data to model the real world. That's the first part. And the second part is really our, you know, our research is really about how testing agents can gather experiences from past testing processes using reinforced learning processes so they can be better and better about, you know, figuring out what could potentially go wrong in the real world.
You said the product's been in development for 15 months. Is that right?
Uh, between the beginning and now, what parts of the development have gotten exponentially better, you think?
Yeah, a lot of things, a lot of things happened. This happened, you know, two or three years ago. Back to that time we only had, you know, Unity or Unreal Engine and currently we have a lot of things like 3D reconstruction, like 3D Gaussian splatting. We actually have a comparison here using this generative approach.
Previously we were taking 250k and uh at current stage it's taking us $600 and 300 GPU hours essentially and it's decreasing rapidly. That's insane. I was totally impressed by it.
Amazing. There's a lot of activity in the space, huge, you know, potential opportunity, a number of competitors, and the technology like you mentioned is moving truly so fast. When you think about your success, what are the most important things that really have to go right for you to be successful?
Yeah, I think there's two things. First one is technical-wise, we need to reduce, there's a commonly known thing called sim-to-real gap. We need to reduce the sim-to-real gap as much as possible so that the testing in the mission can represent the uh real-world performance.
The second one I think is really about extending this shared evidence to a more broad ecosystem, like what insurance...
Companies and government agencies. If everyone is agreeing on one shared, you know, this evidence coming from submission, but I agree on it, it looks good to me, then that is another huge success point of our business. And your work today, is it primarily within the automotive space, both autonomous vehicles, OEMs, all in automotive?
And where do you first expect it will go next?
The next one is extending from autonomous vehicles to general, you know, mobility robots like AGV in warehouse, all things like that, because they are sharing similar techniques. Another one is really about every physical agents, you know, like a robot arm or a humanoid, all things like that. They are more manipulation context rich, but they're sharing similar fundamental problems, you know, the rare safety critical events. So that's going to be our next market.
Which market do you see as your biggest potential market relative to the ease of adoption?
Yeah, I think the early stage of beachhead market is auto driving because that is where the budget is already at. Every automotive OEM is working on that. But in the future, I think our market will expand with the physical AI deployment, like every single new robotics, like deploying new spaces will require additional like a quality QA process or, you know, like a gateway to get this approved or not. I think this will expand probably exponentially in the near future.
Who is your most formidable competitor and how do you win against them?
Yeah, we have a couple of commercial and, you know, open source competitors. One of the most important open source competitor is NVIDIA. They have their world model Cosmos open source, fully available. Everyone's using that. I think the biggest competitive edge is that our AI agent solution and our specific training process can really reduce the cost at the very beginning by another magnitude, like we are designing the most targeted corner cases. That's the first thing so that we can propose to our customer, okay, we cut your development and validation cost by like 10 times or something like that. And in the long-term future, I think with our deployment, we'll gather more and more what could potential physical agents fail in the real world. This kind of fear intelligence is our long-term content edges help us to compete with open source and closed source solutions.
All right, let's give it up for LLAS Intelligence.
So intelligence is going to be a big theme at this round of the startup battlefield because our next startup is Flock Intelligence. Presenting for Flock Intelligence is co-founder and CEO Sarah Nashkova.
That was the bear.
What you just heard is how animals have signaled danger, territory, and movement for millions of years. Their communication is so complex that they have names for each other and can even warn about a specific threat like a man in a yellow jacket. And we know this because today, for the first time in human history, we have started to decode that communication thanks to AI paired with wildlife science.
I'm Sarah. I've spent 13 years in transportation and autonomous vehicles. And I have huge passion for linguistics and wildlife. And this is where they intersect.
You see, the world is becoming autonomous. And I've seen firsthand that physical AI is brilliant, but only in the structured environments we engineered for it. But step outside and it's pretty much blind because nature is dynamic. It's complex. It's always alive and it never got an intelligence layer.
And nature exposes this gap through human wildlife conflicts every single day. A single bird strike can easily cost over a million dollars. A train collision 200,000. And a car hits a deer in the US every 20 seconds.
This adds up to $250 billion each year. And the budgets to fix it already exist. But today's tools are fences, science, or pyrotechnics. They just don't work.
We are Flux Intelligence and we are literally decoding animal communication and enabling machines to co-adapt our peers and speak back in their own language. And this is nature intelligence.
After five years of validations, in May we launched Flock Edge, the edge AI device that's built for the most remote areas, any weather, year round. And here's how it works. It observes the environment and detects the species, age, sex, behavior in real time. And then it decides in seconds on the device with no cloud and applies the right bioacoustic signals for the right animal and the right context.
And each encounter like this makes the whole fleet smarter.
So let me show you the live demo.
Bite me. Don't come to me.
Today we are deployed across three continents with some of the largest international airports like Denver International, with land managers like Parks Canada or USDA, with national highway operators in Portugal or armed forces in Sweden. The edge device is only the beginning though because the same intelligent software can ride on trains, cars, trucks, drones, robotic dogs, and beyond. In fact, it already does. For two years, we've been deployed in daily traffic on trains in Europe in partnership with the OEM giant Alstom. And now, we are preparing to go commercial together.
This is the train speaking with our software and the train's own hardware.
Our mode, our AI engineers working side by side our wildlife biologists with tiny site-specific models where generic AI fails and more than six years of animal communication and behavioral data sets that are extremely hard to replicate. We are operators who have built and scaled at MySpace, Uber or Site View with PhDs in wildlife science and bioacoustics with some of the world's leading labs like KTH, Cornell or Syracuse. And the words are starting to pile up. We are Flux Intelligence and we are creating a world where humans and nature coexist.
Thank you, judges.
Super cool. Like I have so many questions, so I'm going to try to limit it. Um, you were in development for five years. Worry. You sold Denver airport. How long did that take? And then from a technical standpoint, originally I thought you were detecting the sound of the animal and then responding to it. But it appears you're using computer vision to detect the animal and then using sound to say maybe don't come here. So time to market with Denver. What was that sales cycle like? How you detect Oh, and I sorry. Um, you said there was budget for this already. How much and where?
Yeah, so we founded a company in Stockholm, Sweden, and we've been validating with airports in Sweden pretty much from day one with paid pilots. And now we moved to the US about a year ago and we the sales cycle to get the first international airports was about two to three months. So it wasn't that bad. Um, and we've deploy we've been deployed at five international airports in the US now for a couple of months. Um, and for the technology, do you want to address that?
Yeah. So, what was the specific question?
Vision versus sound detection of the animal.
Yeah. So, it's vision. We actually have motion first to wake the system up, then computer vision.
And it says it's a bear, so I want to make a bear sound. And the last one was just where is the budget? How much is it?
Uh, so it depends. There are already budgets for airports. It's for pyrotechnics and lasers and so on, but also for the people that are there working with wildlife management. So a bigger international airport can have about 30 people full-time that are in charge of safety but also going around with vehicles and pirate techniques.
How is pilot success measured at an airport or on a train? What are the metrics that tell you and your customer that this works and it's
So we always have success KPIs and for transportation authorities it can be number of reduced collisions at that hotspot compared to historical data.
Is that a 50% drop, 10% drop? What's good?
What's good? So usually we do at least 80% reduction in the first year is great and that's we have surpassed that so far so that has been amazing and otherwise we do also work with for example farmers and so on and there is also about how much crop can we protect so it depends on the different use case.
This is fabulous, congratulations on developing it. Personally, I was in Antarctica deploying machines like this to basically track penguin rookery population and I was surprised at how effective low-tech it really just needed to produce some ultrasonic pulses and because of sensitivity of hearing for most animals, it was actually pretty darn easy. And we kind of replicated the same thing in Namibia for cheetahs. It worked across species as well. I was curious to hear how you think about margins because this is much more sophisticated than that. So to his question, I guess the success metric delta should probably justify this.
What exactly is the question? What are the margins on the device and why is that required? Why is that delta in the technology you just talked about required?
Yeah. So right now we have margins of about 50% on the hardware and then software is on top of that and then the more we scale because we have a manufacturing partner in the New York state so they are manufacturing for us and then that margin will go up. But also this technology it's quite advanced. So I mean if you work with animals maybe you can have basic sounds for a couple of days and it will work great but then they will start to habituate. So with us we really tap into those biological responses and we really go deeper into how they communicate about territory for different carnivores or distress calls, alarm calls, but then we go deeper because we know that different animals have different dialects and so on. So in order to be effective over season over season and year after year, you really need to keep adapting and learning with each interaction. Just like we saw with the bear.
What's the coverage area for each of the devices? How many of these devices do I need to deploy per square kilometer?
It depends on the hotspots. So with airports we started with about 10 to 50 just to see how we can produce the right behavioral change and then we usually go from there. So a bigger airport could have about 500 to 1,000 devices to actually protect the perimeter because you can think of gates or water drainage systems or fences where animals usually dig under or jump over. So you have a lot of those places that we that need protection.
It seems like the number of customers who need this level of sophistication is, I would think, kind of somewhat modest. I'm curious to hear we presented some really big potential, you know, TAM numbers. So how does this become a really big business? Like how many of those customers, to your point around how sophisticated it is, how many customers need that level of sophistication? And how do you envision this growing into a very big company?
I think all of them because if you talk to either airport managers or different infrastructure owners, they've tried everything and nothing works. So you can have radar, you can have lasers, you can have pyrotechnics, but there's nothing that keeps learning and works across different species all the time. So that's how that's why they team up with us. We started actually
Is airports alone a big enough business? Like if you had every airport and you were kind of the premier tech for is that is that a big enough company?
So the airport the TAM for airports is just $4 billion but the total TAM for us is about 110 if you think about all the shared landscapes right so agriculture, urban spaces because we know that animals are coming more and more closer to us, agriculture, forestry, we have customers across all of those but airports is what we attacked first.
You mentioned farms and my immediate thought was, well, every farmer hates wild boar because they eat their crops. So, if you can keep them from eating the crops, that alone is a pretty big TAM. I mean, anybody lives in Tahoe would like to keep the bears away. Just quick question. How many animals have you mapped and how long does it take to map a new animal? If you haven't mapped wild boars and I want wild boars not to eat my barley, how long is that going to take you to learn?
Answer quickly.
Wild boar. How long? We usually do like between three to four to six weeks of data. Depends on how much data you get, but it's pretty quick.
Okay, give it up one more time for Fox Intelligence. Also, give it up for the very first bear on the TechCrunch Disrupt stage.
Okay, we also have one final startup battlefield company for this block. That company is Satite. Presenting for Satellite is founder and CEO Rama Aulu.
They literally pay hunters to the board because they obliterate.
Hi everyone, my name is Rama Oangafulo, the founder and CEO of Satellite. Imagine if you could deploy and operate AI and software on existing spacecraft. I'm ex SpaceX, Google Cloud, and Tesla. And myself and my CTO Nelson Fenzen here believe that the next cloud is above us. Satellites have compute, but putting it to work is hard. There's limited downlinks, which means send the data down from a satellite down to a ground station on Earth is very difficult. There's repeated integration every single time you have a satellite mission in orbit. And there's unused computing power that is just waiting there to be tapped into. We've estimated at least $100,000 per satellite per year in estimated value loss every single year. So that means slower insights, high engineering costs, and underused capacity.
What if you had one software layer to rule them all across all compatible spacecraft? Imagine you're a satellite operator and you want to run applications on your onboard compute, but you also want to let software and AI developers bring applications to orbit in space. We've got the solution for you. We've got SAT AI where we can analyze and extract useful information on board your satellite. SATNET where we can deliver results across intermittent links in space to help them cross link and interconnect with each other. And SAC Cloud where we orchestrate all of this. Think of it as Kubernetes when we can deploy and manage serviceable workloads in orbit. Importantly, the operator retains all command authority in space. We don't get involved with flight controls. So, let's look at this sort of overall architecture where we start off with the satellite control plane which is the brains of the operation, which is our software directly within these satellites. This is interspersed within the satellite network which are a network of different satellite partners who beam this data down to ground stations on earth which is directly connected to our products sackcloud and sat dash within cloud service providers to create one uniform offering from space down to the ground. Over to Nelson for our product demo.
Yeah. So we provide a platform based on real spacecraft hardware, computer hardware like this where satellite operators and application developers can run their applications before being deployed to satellites, configure their compute resources and also their mission parameters and they are able to view a real time execution sequence of their applications.
We also run AI models directly on satellites for telemetry analysis, anomaly detection and fault diagnosis and mitigation. We also provide software for direct-to-device connectivity from the satellites.
And we are working with NASA on streaming 4K video directly from the moon to earth by using relay satellites on the lunar orbit and earth as hops in order to reduce latency. Back to slides.
Thank you. Back to the presentation please.
Thank you so much Nelson for that awesome product demo. So Nelson's demonstrated that the value starts initially with one spacecraft and we charge satellite operators $4,200 per satellite per month to be able to use our software on their satellites and we also on the other side of this we bring software vendors into this by charging them $10,000 each plus usage fees for them to be lively satellites. If we look at the number of satellites in space right now, there's 20,000 at the moment. There'll be up to 100,000 within the next four to five years. If we only capture a 100 satellites, 100 paying spacecraft right now will already be at close to $15 million ARR and we turn this into virtual AI data centers in space. We started these single spacecraft move over to orchestrating across fleets and then turn them into this virtual AI data center network in orbit to make satellites as operationally efficient as possible and to give you access to these satellites in orbit. We're already doing this.
We've commercially contracted $500,000 plus from different satellite operators and software vendors. And we're in line to win a second tranche from NASA, which is a $1.3 million NASA STR phase 2, where in the phase one we were the only one of 47 companies in all of America who were able to win that contract. We already doing this in orbit. We uplinked our software into a rogue space satellite and reduced the telemetry and log data by 97% in orbit. Then we preloaded our software onto a momentous space satellite and we're able to run AI diagnostics with Google Gemma and also use it as a cyber security health monitoring device. Google DeepMind actually did a case study specifically on the work that we're doing. Next, a couple of weeks ago, myself and Nelson were actually out at the transporter 18 mission where take me to space satellite with one of our partners had our NASA research and our space situational awareness partner get launched up into space for us to prove the different business models. We also have a number of other planned missions including what you see at the bottom. And there's more as well that we are yet to be disclosed. We're also announcing today a phase 2 multi-node flight that we'll be attempting next year to cross-link across two different heterogeneous satellite systems. Finally, I myself mentioned XPace, Google and Tesla. Nelson here was one of the first people working with NASA on distributing space. Sharon here was a spacecraft engineer on the international space station. We partnered NASA, Google, Nvidia, AWS, and we recently announced an $8 million seed round that we closed a couple of weeks ago. Thank you so much. We are here to turn satellites into virtual AI data centers in space.
Judges. Um, one question for me. How did you land on the 4,200 pricing per month per spacecraft and how does that compare to what they pay for others?
Yeah. So the goal here is operational efficiency for the satellite systems themselves. We estimate between $100 and $300,000 worth of value loss. If you take the 4200 multiply by 12 gets to an even $50,000. And also the answer to all questions in the world is 42.
So your value prop to the operator is we help reduce your downlink spend, but really you're trying to aggregate this network of satellites in space to create this virtual orbital data center that you will then sell to other users. And is it incremental revenue to some of the customers who are paying you? Is it found revenue that you will pay them a revenue share based upon their share of available compute that they contribute?
Yeah. So it's all of the above, right? So we make them more operationally efficient by reducing the amount of data that they send down to the ground. Reduce the total latency, but what we also do is allow them to be able to sell resources directly to other satellite operators and also purchase resources of other software vendors directly in orbit. So, it's a sort of multi-dimensional marketplace in orbit.
I like the Hitchhiker's Guide reference. Thank you for that. I always love that in a presentation. So, a couple quick questions. How this is valid for a new satellite being launched, right? You're not trying to retrofit existing satellites in space.
No, we in fact our very first demonstration on that slide with the three was an uplink directly into an existing satellite where we reduce the data by the telemetry by 97%.
So you're accessing INC2 satellites and then changing the software on them just sort of rewrite to get this to work.
So we do both. We access institute satellites and we also preload onto satellites before launch. So of our three demonstrations, one was a direct uplink and two of them were preloads.
Got it. And what percentage of satellites right now are being put into orbit by space by Starlink? What percentage of satellites that are lifted are put up there by Starlink?
By put up by SpaceX, you mean? Right. So yeah, sorry. I'd say about 80% of all satellites right now. So 55% of all launches in the last year were by SpaceX. About 80% of all the satellites that we currently use are put up by
And you worked at SpaceX and Elon has a history of always wanting to bring everything in house no matter what it is. So it feels like, tell me if this is an unfair characterization, you will be Android to SpaceX's Apple.
Okay, fine. Big enough I guess. And how do you see competition from orbital data centers? How do you see that market developing?
So we actually run orthogonal to most of them, right? Because most folks are Carbers, Star Cloud, Google, SpaceX following the vertically integrated closed ecosystem similar to the Apple ecosystem that was just mentioned there. They all will also need to be able to be interoperable with other satellites in orbit. And we are actually building part of what we're doing with NASA is building this standardized comms protocol to make these satellites interoperable in space. So eventually they might not like us at the beginning but they will use us in the end.
Does your success depend on having kind of excess capacity and how much excess capacity really is there and will there be for you to utilize?
So no success is not completely dependent on that. In fact, our initial wedge is focused on operation efficiency of singular satellites, right? And then we start to scale out into heterogeneous cross connecting of different types of satellite systems. In terms of the excess compute, firstly, right now there's 20,000 satellites in orbit. All right. By the end of this decade, there'll be about 100,000. Most of that net,000 plus that will be getting added to the ecosystem. Folks are adding in satellite links. Folks are adding much larger GPUs as well to these satellites as well. So it's hard to quantify but we know right now that the orbital data center race is on and folks are putting more and more compute into space. So we're working with Nvidia and working with a bunch of other folks specifically on that.
I worked on a SAT cube project recently and I was actually surprised by the sheer lack of standards in software development there which is different than most enterprise software. Do you think you have a window of time where the standards is being developed that you have to do a land grab for you to have a shot at building a scalable business or is there a different way to think about it?
Yeah, just like a business model, we're playing both ends of this, right? So we're rushing to get into as many satellites as possible. Our target for this year is to be on 50 nodes, right? But simultaneously Nelson here for example was invited by the CCSDS to be part of the store and forward standardization for sending data from a ground station directly to cloud service providers right and at the same time of course we're working with NASA on the standards in orbit so we want to be a part of all of that sort of creating this TCP IP solution right but at the same time it's not completely dependent on that standardization will be successful regardless.
But if that standardization holds doesn't it change the character of the market?
It helps us right if there's more standardization that means we can cross link across many more birds in orbit.
But so can everybody else so to his question of the iOS and Android doesn't it help Starlink accelerate market capture even more?
Well that would assume that they actually want to play nice with others right and I think there's enough folks who do not want that association who are forced to have that association and we're giving them a healthy alternative.
What type of operators are you finding it easiest to sell to today? Is it Earth observation? Is it SASCOM? Is it government defense?
I mean, we've worked with quite a few different operators that want our software on their satellite. So, the Earth observation folks were what we thought were the most obvious because you know we can reduce any images with cloud cover for example, but actually literally we can make any type of satellite operationally efficient. So we're talking even with the folks who are doing orbital data centers as well who you think are our competitors who can't name right now but they want our software on their satellites also.
All right give it up for satellite.
And with that we are wrapping up session one of the startup battlefield. Here's what's going to happen next. The judges are going to follow me backstage. They're going to help us choose some finalists. Then we're just going to take a minute to reset the stage for our next speaker. Before any of that can happen, I need one more round of applause for our incredible judges.
We've been advancing energy technology for almost two centuries, powering innovation, and shaping the way we live, work, and thrive. From the first fuse to powering the future of quantum computing, Schneider Electric is and always has been your energy technology partner. We electrify, automate, and digitalize every industry, business, and home, driving efficiency and sustainability for all.
Hey, I'm your dot. I'm new in chat GBT, ready to jump in when you need me. I can proactively keep things moving, like launching your small business.
Okay, perfect.
Getting your app rebuilt, tested, and ready for rollout or even getting you prepped for your board meeting.
Go ahead and update it with the latest numbers, please. Love it. Let's get to work.
Please welcome back to the stage weekend editor and battlefield host Anthony Ha.
Yeah, it's me again. So obviously building a software company is hard, but building cars, building factories, supply chains, and an entirely new automotive company is I'm guessing I haven't done this myself, but I'm guessing it's much harder. Rivian founder and CEO R.J. Scarring has spent more than a decade doing exactly that through enormous expectations, brutal manufacturing challenges, and some very public ups and downs. So please welcome R.J. Scarring and Techrunch's Kirsten Cororusac.
Welcome back.
Yeah. You were here four years ago and in that time you've launched a micromobility startup called also which you were working on in secret and then a robotics startup called Mind Robotics that you were also working on secret. So what secret project are you working on now?
Oh boy. Right before he walked in I said so what are you going to try to get me to say that I'm not supposed to say today? So, we really started directly here. You know, ultimately there's lots of things we're working on within the business that we haven't shown yet. So, we've just launched it within Riven. We launched R2. Super excited about the vehicle. There's a bunch of products that come behind that both on the technology side and then the vehicle side. You know, within mind, it's incredibly exciting what's happening in robotics and what we're capable doing today. When we think about the progress that's happened on hardware modeling techniques bringing that together to develop robotics that can actually be deployed at scale and so we haven't shown any of that yet. That would have been cool to show it today but soon. And then as you said, within also we've got a bunch of products really around think of it as small form factor autonomy that we're quite excited about but
So what you're saying is you're not working yet. Are you saying that you're not working on another secret project?
No, no, no additional companies. Just those three.
Okay, you're done for a while. Okay.
One of the things that you have talked about a lot is that R2 is like the big project right now, the launch of it, and there's actually demo rides happening outside of here, so you should all take one. You've said in the past that it is an existential product for the company. Really that what happens with R2 will really be the difference of Rivian really truly scaling.
And I'm wondering are you encountering different problems with the scale up of R2 versus R1? Very different era of the company.
Yeah. So R1 was a flagship product we launched Driven with and you know this was we launched the car late 2021. We made a lot of the sourcing decisions and overall production plans for it starting in like the 2018 2019 time frame and so the level of maturity within Rivian was much lower. There were a lot of things we didn't even yet know that we needed to know. And we didn't expect COVID to of course happen and then the huge you know operational implications of that. So it was just a there's a lot of challenges associated with that launch. Fast forward to R2 the company is so much more mature. We've really built a lot of process around launching complex projects, managing you know supply chains where you have hundreds of suppliers across multiple tiers. And so there were a lot fewer surprises, but there's always you know in something like a car where you have you know tens of thousands of you know discrete components coming together into a single product, if any single one of those is late, you know that can delay the program, that can delay ramp up. And so there's always challenges of ramping supply chains but it's so much more control and the organization is so much more mature.
So is there anything right now that you are navigating that is particularly difficult that is just you maybe hadn't anticipated to your point you know there's always something that comes up.
I think the biggest challenge for us by far is around supply chain just in general and I think for anybody here that's building hardware you can connect to this the supply chain environment is just not particularly stable. We've got very dynamic geopolitics that are playing out. We have cost structures that have gone up dramatically around the physical world largely driven by the demand for a lot of those same resources for building data centers and so if you're building manufacturing plants there's constraints on resources. And so this is just this is sort of the nature of building physical products today I think any company that's doing it is going to encounter these types of challenges and so we knew that it's not as if that's a surprise coming into it we built a robust organization to handle you know supply chains that are more difficult, more dynamic tariff environments, and operating environments that I think are probably the most complex they've been in a long time.
Yeah. It's funny. It's a bit of deja vu, but a little bit different because there were supply chain problems with co. Yeah. This time around, one of the big issues is around memory chips, and I'm wondering how you're navigating that right now.
I mean I think everyone who has products that have you know any sort of electronics is dealing with this but there's really two ways. One is you fight like you know very very hard to get the amount of memory in every product down so minimize the amount of memory required and that's a tactical exercise and then the other is building lots of potential supply paths so multiple different suppliers providing memory and then at the end you're paying more. So you know in the case of a vehicle that has a lot of capability, you know you're adding many hundreds of dollars in cost with the price of memory going up so much, and that was not something 12 months ago, 18 months ago, you would have predicted.
One of the things though that I've seen with Rivian that's interesting is that you seem to be getting more and more vertically integrated. And I'm wondering if that's one of your strategies here because you're even going to build the silicone for future products.
Yeah. Yeah. Yeah. I often joke if I think for a lot of entrepreneurs that probably fall into this category. If you had infinite time, infinite resources, we wouldn't have any suppliers. We'd build everything. Our supply chain would be buying you know buying earth, buying pieces of land that you then process materials from. And so of course we're not buying land and making our own steel or making our own aluminum or making our own copper. So we buy materials from upstream suppliers. So we you know we leverage existing supply chains for raw materials, but there are some areas where we do go pretty deep. And so as you said, silicon is a really key one. In the case of our vehicles, we're really focused on delivering a much higher level of capability for self-driving. And so later this year, we launch a point-to-point capability where you can type the address into the car. It completely drives itself to its destination. Through the next two years, we're working towards a level four capability where the vehicle can drive itself empty. So that's you know critical for robo taxi applications. Imagine a vehicle that could drop your kids at school or you know drop you at the airport and drive home. And when you look at the cost to deliver that outside of the development cost, a big driver is the cost of compute, the cost of inference. And so we took the decision, this was now I guess four years ago, we took the decision to vertically integrate our in-house or our inference platform.
And you know it's many hundreds of millions of dollars to do a program like this but ultimately at the end you have a structural cost advantage where you've taken the inference cost in the vehicle down by you know many hundreds of dollars. And so that's an example where we invested a tremendous amount of money to then ultimately deliver this cost advantage later on. But that's true on you know motors, gearboxes, power electronics, all the electronics in the vehicle. All the software in the vehicle we do you know completely ourselves.
Well and now robots. So, Mind Robotics, this is a company that you launched in November of last year. You've raised more than a billion dollars, I think, so far. So, good for you on that. I'm sure a lot of founders would love to hear how you did that. But specifically, Mind Robotics is the approach is industrial automation. You plan on making Rivian the first customer. So, how exactly are you doing that? And are you dividing your time between Mind Robotics and running Rivian right now?
Well, I think there's so many things to be excited about from a technology point of view that are happening. We're witnessing multiple inflection points and we've seen a number of those start to take off. So with electric vehicles, we're seeing that get to an inflection point where over the next decade we're going to see a really big transformation of the vehicle fleet that's on the road today. We of course have seen that with LLMs in the way we do our work, the way we run our businesses, the way we think about decisions in our own personal lives. And another really big inflection point which I believe is going to have perhaps the most profound impact on how our society operates is the role of robotics. And you know we've been dreaming about robotics you know that are capable of doing humanlike tasks or human skills for as long as we've told stories. So this has been topics in science fiction for a hundred years. It's been in movies for 50 years and there's many different versions of how robotics play into society and even today we still reference movies. We reference books when we talk about robotics. What's remarkable is that in just a few years in the next 5 to 10 years we'll stop referencing movies and we'll be referencing what happened. We'll be referencing what plays out over the next 5 to 10 years. And so with mind the objective was really to think about very hard how do we as we go through this transition how do we develop robotics that are capable of doing highly dextrous tasks capable of humanlike reasoning but develop those robotics so they can be incredibly helpful for humanity and our focus is on the industrial side of the opportunity. So really thinking about robots that can do manipulation for assembly of objects. That could be anything from a car all the way down to a shoe. And you know everything in between. So that's you know consumer electronics that's a dishwasher that's a you know that could be furniture. And so there's this enormous opportunity to really rethink our industrial supply chains in the context of robotics that are capable of doing these types of things. And so realizing how big of an opportunity is this is and just to put this to scale about a third of the world's GDP is industrial labor. And so this is an enormous need this is particularly profound in developed markets where cost of labor is quite high. So in places like the United States as we think about bringing different types of manufacturing back into the United States. And so realizing how big of an idea this was as you said last year I decided to start another company to focus on this and we're having an absolute blast. We've built an incredible team. Just announcing today we brought on Doug Field to really for me to partner with and building this organization. I've known Doug for years. I'm really excited about him joining to help build and drive the business. And to your question of like how am I splitting my time, it's there's seven days in the week and it gets split I'd say you know sometimes 50/50 sometimes you know a little more a little less but you know it's not like we're only splitting 40 hours. There's a lot of hours in a week to do a lot of things.
Yeah. We've had this conversation before where you're like well I take the plane here and then I go to this factory and then I work and now you've since then added multiple other startups. So is it fair to say that Doug is going to be sort of more steering the ship? And for those who don't know, this is former Tesla, former Ford, really reshaped their EV strategy, worked on or led the special car project, not so secret project over at Apple. Is he going to be the one really directing mind?
Well, you know, that's a great question and I've known Doug for years and there's people you meet and I feel like this is again something any builder goes through is you meet someone and you say, "Boy, I would love to work with that person someday." And so for years I've often said, "Boy, I'd love to you know work with somebody like Doug." And as I was thinking about building mine, as mine grows, the type of person that I would ideally have as a partner in building the business, I would describe it as I'm looking for someone a lot like Doug Field. And so ultimately, I called Doug and I said, "Hey Doug, I'm looking for someone like you." And it was you know sort of fun started the conversation. And him and I then really began digging into how important it is that we really think deeply and very intentionally about how we develop robotics and deploy robotics into our society into our economy. And we found such alignment in how we thought about this from the overall mission of the business. But even more important, we found really deep alignment just in how we both work and how we build teams. And so I was sort of on this journey of looking for someone like Doug Field and found Doug Field and so it worked out really well. And he's someone I've again as I said I've wanted to work with for years. So this is I couldn't be more excited.
Did you call him? I mean Mind came kind of out of stealth in November of last year. So were you was he your first call back in November?
He when I first set up mind the way I structured it is I think every startup benefits from being very flat in the beginning and not having traditional titles and not having sort of accepting that the org chart is going to be wildly fluid and I've said this many times in building mind I experienced this with Rivian experienced this with also and so by design mine started as a very flat organization stayed very flat the companies a few months ago we surpassed a 100 people and so as you start to get to that level of size you actually need to start designing what the organization looks like and advice I give all the time including to myself is don't get fixated on an org chart recognize that you know at what the way the org looks at 150 people is going to be different at 300 it'll change again at a thousand and so as I was thinking about that I was also thinking you know who is the leader? Who is a leader that I can work with and work alongside as we continue to iterate on what the right time appropriate or work chart looks like. And so you know that you know my first thought was someone like Doug and here we are.
Yeah. Great. One of the underlying foundations of this industrial automation, of course, is going to be foundational models, specifically AI. And this is another area that Rivian has really jumped into. And I'm wondering, you recently rolled out the AI assistant in the R2 and that is interesting because there's my understanding is that there's a model that was built internally, but then it will also work with other outside models like a Gemini right now. It almost sounds like you've built an agentic system. Is there a place for AI agents to be running around in the vehicle?
Well, there's a long context on this question. So I think first we have to recognize that the frontier models and the models that we're using, whether it's to do our work or to get advice on you know an exercise routine or how to make a meal, those agents we're using start to build a lot of context around how we're using them. And so this is a really important question for us of finding ways to integrate those agents into the vehicle and your experience, but then connecting them to the context that they have around you. And so the first version of our assistant doesn't have that yet, but that we see as a really important element that's coming. But the agent we have that runs in the vehicle, think of it as like a traffic light. So if you ask the vehicle something, it decides what model it's going to use. It can use a local model that's running on edge and vehicle. If it's a more complex question, it can go up to the cloud. And then even for that, we have different models that we use. And so there's a small model that we built that's essentially determining where to direct the traffic of questions. I think that will continue to be the case. I think what we'll then see play out over the next several months is the beginnings of your context from those different platforms moving into the vehicle and allowing you to select your own agent and being more deterministic on which agent you're using.
So is there a scenario in which there's the Rivian AI assistant which I recently tried out actually at your Palo Alto office just yesterday and where that is happening it can maybe in the future control aspects of it like the air conditioning and things like that but let's say I just love Claude will I also be able to do that?
Yeah and that's where it becomes magical is you're driving the car you say well I'd like to adjust the AC you know turn it's hot in here, turn the
temperature down, or you know, I want my cooled seats on, and then simultaneously say, 'We were talking yesterday about this meal I'm, you know, planning to have guests over on Friday. Can you help me think through what we're going to build, what we're going to make?' And like immediately know that that context is in that, you know, contact window from a conversation you had on your phone or, you know, elsewhere. And of course, that's what we want, but there's a lot of plumbing behind the scenes we have to do to make sure that that works. And it's not just within Rivian, that's also a lot of coordination with the labs to make sure that they enable that plumbing.
Well, okay. So, speaking of plumbing, I mean, we have—we were just talking about this this morning, which is there is efforts then on the security side to like rein in rogue, you know, AI agents. And once you have it in the vehicle, there is another new element there because someone is driving. So what are you doing on the security side to ensure that there are maybe not too many AI agents like running around within the vehicle?
Yeah. Well, the plumbing is really important. I would say that it maybe implies it's simple. It's quite complex because you have safety critical items like you wouldn't want an agent to determine, hey, it's time to apply the brakes or time to steer the vehicle. So anything that's vehicle level controls are not being handed to an agent. The things that we're allowing your AI counterpart to do is make adjustments as if you have somebody sitting in the passenger seat that's, you know, adjusting the radio, you know, or changing the station, changing the HVAC settings. You know, in the distant, distant future you can imagine talking to the car and having it completely actuate brake, steering, but that's, you know, that's more in the realm of self-driving, and there it's just 'take me to this address' and then the vehicle handles all the driving, and that's how we think about it. So there's not a world in which you'd say like, 'agent, please stop the car.' You tell the address.
Yeah. Well, so you have all this going on on the AI assistant side and also on the autonomy side. You have this joint venture right now, a pretty lucrative one, just under $6 billion I believe, with Volkswagen, and that doesn't involve AI or autonomy. It is just the electrical architecture and sort of software. But when I was in Palo Alto yesterday, I saw Volkswagen folks there, and I'm wondering, you know, they're seeing what you're working on. Are they interested in expanding that deal to cover AI or autonomy?
Uh, yeah. I mean, just for context, we at Rivian have built and designed all of its own electronics, all of its software, and it was built around a zonal architecture. So the idea of using a very small number of computers to do lots of tasks—which, you know, to tech folks that seems obvious—but in the car world historically the way that cars evolved is each computer had a discrete function or domain-based function. So we actually in the car world would call that domain-based controllers. So your electrically adjustable seat would have its own little ECU, a little electronic control unit, you know, small computer that runs the seat. Your cooling module would have its own computer. The right architecture clearly in the long term is to move to a very small number of computers doing all the tasks. Anyway, so we developed that—you know, outside of China there's really two companies that have developed that type of architecture: Rivian and Tesla. And then we decided to license that technology as part of a large joint venture. This was a $5.88 billion joint venture with Volkswagen, but it doesn't entail, as you said, any of the AI autonomy efforts. That's a huge focus for us as a business. So there's certainly opportunities to add to that relationship, but even as it stands, the scale of, you know, really working closely with the Volkswagen Group to rethink their network architectures, their approach to even supply chain and sourcing—you know, so this, you know, think historically you'd source a seat and the supplier wants to sell a smart seat. The seat has a little computer on it, and we absolutely do not want a smart seat. We want a seat with zero intelligence, and we'll do all the thinking and intelligence in a centralized way. And so there's a big shift that's happening within the automotive supply chain where historically a lot of suppliers have tried to put electronics or tried to put computers into their various components, and architecturally that's quite flawed. So you're seeing a push from, you know, tech-leading organizations—I said Rivian being one of those—to sort of scrape away all those extraneous small computers. Within Rivian that's very easy because, like, I control the company. So it's like, we're not going to buy seats with ECUs. Within, you know, large companies that have a lot of legacy and history, it takes a bit of work to make those changes.
Yeah. Volkswagen Group is something like 750,000 employees. So you have a lot of work to do. But do you expect it to expand to cover AI?
I think that relationship will continue to expand and grow. It's going really well. We have a great partnership with them. If you were at our Palo Alto offices, you probably saw lots of cool stuff. There's different vehicles, different brands, different things driving around. But we also see opportunities to sell this type of technology across our technology portfolio to many different brands. And so, you know, one of the unique things about automotive is it's such a, you know, it's a business that's been around for so long that the vast majority of car companies today were just born very different than Rivian. So they were born in a world where the cars didn't have any computers. So, you know, prior to the 1960s, cars didn't have, you know, any—like, there's no compute. It was 100% these were all mechanical. And so the skill sets, the topology of how the organizations are structured, even the geographies where a lot of the core operations are set up are quite different than what you would do if you were thinking about building a software and electronic-centric company today. And so by virtue of us being born much more recently and, you know, as a company focusing a huge part of our engineering and resources on software and electronics, that's our core strength and skill, and we see a real opportunity to help companies that just don't have that as a core strength or skill.
So what companies are you talking to now?
Well, we're all listening.
Um, one other company that you are doing a lot of partnership with in a separate area on autonomy is Uber. And this is a really interesting one for me because you were also internally working on, you know, essentially the equivalent of FSD, Tesla FSD, which is a point-to-point navigation that is a level two level of autonomy. And then separately here you have also struck a deal with Uber to do robotaxis launched, I believe, by 2028. Before we get into that deal, I'm wondering why not just progress to a point where you're selling personal vehicles to people that are autonomous vehicles and just forget about the robotaxi part?
Well, they're not mutually exclusive. Um, maybe it's worth me spending maybe a minute on this. So when we think about self-driving and autonomy—and this is of course a concept that's been around for a while and people have been working on this, you know, for close to 20 years now—and up until just a couple of years ago the way these systems were developed is they were what we would call rules-based systems. So you'd have perception platforms, so cameras and likely radars and maybe LIDARs, but, you know, call it two to three different modalities of perceiving the world. And that perception platform, as it's seeing the world, would then identify objects, classify those objects, associate vectors with those objects, and hand all those classified, vector-associated objects to a planner. And the planner was a rules-based environment, meaning like software developers would write the rules of the road, try to describe how to drive with code. And that was what was the state-of-the-art, the way it was done until just a few years ago. And the challenge with that, of course, is the systems are quite fragile. So if you change the vehicle, it requires a lot of updates. If you change the environment, you know, let's say go from San Francisco to, I don't know, Germany, it's a very different environment. What happened a few years ago is of course technology advanced dramatically and we came up with the idea of transformer-based encoding and an entirely different way of representing really complex bodies of human knowledge with foundation models and these large neural nets. Of course, we now know that so well because we experience it every day with LLMs. As that happened, the way that self-driving is developed has changed dramatically. Meaning there's very, very little link to how it was done prior to, let's say, 2021 or 2022. And today, everything is about building a large-scale model. And the model—it's actually helpful to think of it like this. The model is a large-parameter representation of how to drive. It's informed by billions of miles of driving, in our case from our fleet, and it's continuing to get better. And so as you upgrade the perception or as you upgrade the compute, the model doesn't forget what it learned before. So, you know, think of our own cells. We run a neural net in our brain. Imagine you learn to drive without glasses and suddenly I hand you a pair of glasses. You don't forget your previous knowledge of driving. You just suddenly can see details and nuances that you couldn't see before, and so you'll drive better. Now this is a little harder to imagine. Now imagine we increased your level of compute. We said we're going to make you 10 times smarter than you were previously. Again, you wouldn't forget your past knowledge, but suddenly you'd be able to identify and recognize patterns in ways that you hadn't before. And so I call that out to say that the way that we're now building autonomy is we have what we call LDM, a large driving model. And that model is the same for a level two++ system, something that is what we call point-to-point. So similar to Tesla's FSD, where the vehicle can drive itself to an address, you type it in, it just drives there. And something that has more compute, more sensors, and a further development of that model, which is what we're putting into our level four, but it's the same LDM. And so it's actually a great simplification for how we look at autonomy, where we no longer have to think about the workflows as being different between something that's for personal consumption where you still have to be in the driver's seat from something that can completely drive itself, whether you personally own it or it's a robotaxi. The technology remains the same.
So, in 2028 when Uber and Rivian launch—and we could have a whole separate discussion about that—how far away will the personally owned AV be? Is it about the same time?
Well, this is really interesting. So we think of robotaxi. It's super exciting. We're excited about it. We're building a product for that in partnership with Uber. We did a $1.25 billion deal with Uber on this. But, you know, 95% of the miles driven in the United States—so this is trillions of miles—95% of these miles are in things that we personally own. So like you own it or your household owns it. And so our view is it's unlikely, even with level four capabilities, that we're going to see a dramatic societal shift away from that. Meaning it may go from 95% to 90%, and it may go from 95% to 80%, but still the vast majority of the miles driven, for a lot of societal and psychological reasons, are likely to stay the same. Meaning you like your car seats in a car, you like it to be your own personal space, you have your gym bag in the trunk. And so our view is that you're going to still see the vast majority of miles be driven in things that we own or our household owns. And so we do really believe that personally owned autonomy is going to be a big part of our mobility ecosystem, but we're going to play both. So we're going to participate with robotaxi, but we'll also have a personal level four R2 and ultimately R3 and future products. But where it gets really interesting is then you think beyond R2 and R3 and start to imagine like the form factors that you want to build for a personally owned autonomous vehicle. You know, it opens up all kinds of design opportunities. It can create dramatically different experiences. I'm wildly excited about what the next 10 years look like in transportation because of what this enables us to dream and think about.
So that's your secret project then?
Yeah. Well, we better have something beyond what we've shown.
Yeah. I mean, I can tell you that the thing after R—so we did R1, then we did R2, then we did R3. The thing after R3 is going to be called R4.
We just haven't shown what it looks like.
Yes. Okay. And it could be potentially future. It will be cool. All right.
Great.
Well, thank you so much for coming today and sharing what comes next for Rivian.
Thank you for having me.
Almost got me.
All right, I can tell by the movement that everyone knows what time it is. But first of all, thank you R.J. and Kristen. Thank you to the Startup Battlefield companies for getting things started here on the Disrupt stage. As you can tell, we are about to break for lunch, but I'm just going to remind you of a few things that are going to be happening this afternoon right here. Startup Battlefield returns at 1:00 p.m. with founders tackling energy and manufacturing. Then we're going to get into the future of AI compute with Cerebras Systems CEO Andrew Feldman. And then we're also going to look at the future of the human body with Science Corp founder Max Hodak. But of course, you can look beyond this stage. On the Builder stage, if you're raising your first round, Pun Agarwal and Austin Clement are talking about winning pre-seed funding before you even have a product. Later, Google's Robbie Stein gets into how product decisions change when you're scaling from an MVP to billions of users. And Gusto CEO Josh Reeves tackles hiring when AI is a co-founder. And the Industry stage switches gears this afternoon to smart systems, the infrastructure underneath the AI boom. We'll get into the data center power crisis, the next era of software infrastructure with Twilio co-founder Jeff Lawson, and why clean energy projects keep failing to connect to the grid. So, grab lunch, check the app, and build your afternoon around whatever problem you're trying to solve. And if it's energy or manufacturing or AI compute or the future of human-machine interfaces, we'll see you back here at 1:00.
Please take your seats. The program will begin in 10 minutes.
Please take your seats. The program will begin in 5 minutes.
I'm going to make a website. I'm going to make a website.
Obsessed with your idea? Put it online with Hostinger. Just tell AI what you want and watch it get built. Then make it search-ready with SEO Agent, your built-in SEO consultant. It researches relevant keywords and spots SEO opportunities. Make your website more visible with AI keyword research and content improvements. Hostinger, the agentic growth platform to build and grow your business.
Welcome to Build Mode, the TechCrunch podcast for founders. We talk to investors, exited founders, and early-stage founders who are building the future today. Real advice, no gatekeeping, no sugarcoating, straight from the mouths of people who've gone from zero to one and beyond. We knew we wanted to be a public company. Nvidia has stamped its tech approval to say this is going to be transformative to your industry. I think there's always this balance in the market between people's appetite for first-time founders and people's appetite for repeat founders.
This is Build Mode. Subscribe wherever you get your podcasts.
Please welcome back to the stage weekend editor and Battlefield host Anthony Ha.
Hello and welcome back from lunch. I am Anthony Ha from TechCrunch and we are going to have another round of the Startup Battlefield. This time we're looking at companies rethinking how we power, build, and manufacture the physical world. For those of you who were here before, it's going to be the exact same format: six-minute presentations with live demos and then six minutes of Q&A. One thing I failed to tell you last time is that if you have a favorite company or a least favorite company, you can go into the TechCrunch events app and you can vote for your favorites, and then you'll get to see who everyone else's favorites were at the end of the conference. Also, I mentioned before that the finals were on Wednesday. That was a lie. Forget I ever said that, except that I'm mentioning it now. The finals are on Thursday. I was confused because of the slightly different schedule this time. So, I mean, you should come from every single session, but definitely be back here on Thursday for the finals. Of course, in order to choose our finalists, we're going to need our expert judges. So, please welcome our fabulous judges.
First up, we have Caleb Appleton, a partner at Bison Ventures where he invests in physical AI, a biomedical engineer by training. Caleb's career spans early science work in the field of engineered gene and cell therapies, frontier tech investing, and most recently operating a scaled technology business.
Next we have Fesa Haskar Raman, partner at Felicis, where she focuses on AI-native enterprise security, developer infrastructure, and the platforms powering an AI-first future. Before joining Felicis, Fesa was an investor at Menlo Ventures. Her earlier career spans engineering and strategy. She built products at Fitbit and Analog Devices and advised Fortune 500 technology companies at McKinsey.
Next we have Show Leia Ho, partner at General Catalyst. Show invests in early-stage founders across sectors and serves as a board observer at Together AI. She joined General Catalyst in 2024 as the firm's youngest partner.
Last but not least, we have Leah Sullivan, general partner at Precedent VC. Leah Sullivan is the founder of TaskRabbit. She's a technology entrepreneur, investor, and board leader who has spent more than two decades building and scaling products that improve how people live and work. She is the founder and managing director of Precedent VC, an early-stage venture capital fund backing category-defining technology companies, and continues to serve as a general partner at Fuel Capital. So, give it up to our awesome judges.
Now let's bring out the first startup in Battlefield round two: energy and manufacturing. That startup is LNK Energies. Presenting for LNK Energies is co-founder and CEO Cheyen Biswas.
Hydrogen is the energy of the future. We've all heard this. It is this awesome molecule that has the ability to power heavy-duty industries. Think ships, trains, trucks, microgrids. We make a lot of it in the US, but we don't really see it around. And that's because of one single bottleneck: hydrogen is very expensive to move around. It ships as a compressed gas or a super-cold cryogenic liquid, requiring purpose-built, messy, and complex supply chain. Over 95% of what you pay at a gas station is just the cost of moving this fuel around. LNK does it differently. We move hydrogen just like diesel. We use a non-flammable, room-temperature liquid that can store and release hydrogen. The best part? It's compatible with your existing diesel delivery infrastructure. So use your same tanker trailers, pipelines, and gas stations, but now you have clean hydrogen molecules. We bring down that 95% delivery cost all the way down to 6%.
To add on to the benefits, it's less flammable than diesel. So hydrogen done in a way that's safer than conventional fuel. We have long-term energy storage sorted out. It's produced in the US. To say, LOHC basically makes hydrogen transportation simple, safe, and scalable. We've already shown how to use LOHCs to actually generate power and power all your different devices. So if you think about it, we've shown this in lab with four years of research at MIT. We've built a system that takes this hydrogenated molecule, this LOHC, and creates power through it. Now, we'll have Alexis, who is going to come and show us the demo of the hydrogen release unit that we've built. So, this is a scaled-up model of our hydrogen release unit. As you saw in the photo, what we have is the LOHC fuel will go in from the bottom. It will interact with the catalyst, as you can see by these pellets, and then they will be heated by these heating tubes, and then the hydrogen will be released from the LOHC fuel. We separate the hydrogen, and then the hydrogen can be sent to your combustion engine or a fuel cell, which we've already demonstrated in our prototype, to provide power wherever and whenever you need it.
Thank you for that awesome demo, Alexis. We've done over 50 hours of testing with it. We've already shown this. This is a patented technology from MIT. And to really appreciate what we have been able to put together, let's just talk about what systems before used to look like. They were as large as shipping containers. So that's the stage that we are standing on right now. We've managed to shrink it to something that's as small as a mini fridge. And this is exactly what's opened up the possibilities of all the different applications that we are talking about. The improved power density comes from improved thermal efficiencies, high-pressure hydrogen release, allowing us to provide hydrogen at a price that's 26% cheaper than diesel. So the sustainable choice is also your cheapest choice.
This was a part of my PhD at MIT. I'm Shyep, co-founder and CEO of LNK Energies. Alexis, our CTO, she's a mechE by training. She has designed, built, and raced race cars. Also our CSO is Professor William H. Green, who is the director of the MIT Energy Initiative and a leading researcher in this LOHC space. We are getting to market with the LA drayage trucking segment. So drayage trucks take containers from ports to regional warehouses. In this space there is already a move towards decarbonization and zero-emission fueling. Given the high fleet utilization rates and the difficult drive cycles, battery electrics don't work very well. Fleets are already using hydrogen. What we are able to do for them is come in and provide an option that is increasing their profit margins by 8x. Now this is game-changing. This is revolutionary. This is why we have partners and customers on the operational flexibility. We have a solution that is equivalent to diesel from refueling times to miles to payload whilst being zero emissions.
Now let's see what we are working on to effectively get this on a truck and drive that. You have a single hydrogen release unit. That's what we have showed here. You put more of them in parallel to create a stack with a higher power that sits right behind your truck. It's a small system that sits without taking any more space. You have a standard diesel tank containing this fuel that goes to your hydrogen release unit. The hydrogen comes out. It gets separated. It gets purified. It goes to your fuel cell. That fuel cell makes electric power, goes to your wheels, and off goes your truck.
We are bringing this with our partners. We are not doing everything ourselves. We have partners everywhere from hydrogen production to end-user logistics. We have support from local government. We also have research partnerships with UC Riverside. The entire idea is we can focus on our core technology, and as long as we deliver that, we already have the other bases covered. One thing we would really like to highlight is our partnership with Symbio. Symbio is a leading fuel cell manufacturer who makes fuel cell electric vehicles. We are working with them to take our technology and get a commercial truck that's going to be running in the next 18 months in California.
Our business model is simple. We sell the hardware kit once. We sell it to our partners like Symbio and others. We are platform-agnostic, and then we make margin on the fuel supply for the rest of the utilization of our hardware. We get the fuel again—it's an aggregator model where we aggregate the fuel demand, and we have partners to be able to provide that to our customers.
We are starting with the $2 billion drayage market industry, but our technology is a platform tech. It can go into ships, rails, locomotives, microgrids, AI data centers, with the ability to touch a TAM that's over $500 billion. We are here to make diesel obsolete. If you want to join us, please come see us at our booth. And we are doing a seed round. So for investors and partners, please connect with us. Thank you.
Judges. I'll start. Great presentation. And I'm curious, with this breakthrough technology—faster, cheaper, more efficient, smaller footprint—can you talk about how the incentives across the supply chain are going to be changed and where you think there could be some misalignment in the rollout from the business model standpoint?
Yeah. So for us, the entire idea is not to monopolize the value creation, if you may. There are already margins that people like, you know, the hydrogen producers have a margin that they want to be able to charge so that they're able to sell their hydrogen. All the costing that we presented here, the 26% cheaper, includes the prices that they're already comfortable making from their initial investments.
So all stages have already added what percentages that they would want. So what I'm presenting is what is left after the piece of the pie already takes care of everybody there.
Okay. Great. And another thing to point out is, you know, this is a commercially available molecule. It's sold by Eastman Chemicals. So it's something that people already know how to handle their safety data sheets for. And we have been talking to a lot of these government agencies to handle fuel distribution, you know, the Federal Motor Carrier for vehicle specifications and all of that. And they haven't flagged anything really with our new technology and the fuel because of how safe it is and how it's already commercially available.
Great.
Yeah. Can you double-click on the compliance requirements and any sort of rocks that you need to turn to make it widely available?
Yeah. As Alexis was mentioning, we've already engaged all of the agencies starting from, you know, FMVSS, NHTSA, things with vehicle level as well as moving the fuel around. One of the best parts of us is because of how similar we are to conventional fuels, being even safer than them, nothing has been flagged so far. We are in the process of getting permitted for it. There's already permits that exist in different countries where this actually moves through pipelines just fine. So what I would say is I don't have the permits today, but the conversations have not yielded any red flags. In fact, we have kind of gotten a pat on the back on how we're making this technology inherently safe such that, you know, even a mis-spilling event or anything like that doesn't pose any risk to the urban setting that it's going to be deployed in.
I'd love to learn a little bit more about your guys' story. How did you guys meet? Why did you guys decide to work on this idea?
Yeah. Yeah. So this stemmed initially from Shine Deep's PhD thesis. They were working on the reactor itself, which is, you know, what you see in the demo, and looking at how do we improve the existing LOHC systems, you know, from that shipping-container-size unit to something that can actually fit on a vehicle. When I joined the project as a researcher, it was to work on the powertrain side, actually building up the engine. So for us, you know, our work together has been a good mix of chemical engineering, mechanical engineering, integrated. And ever since then, we've been going through all the accelerator programs through MIT, and now we're ready to spin out and start our seed round and commercialize.
Yeah, very impressive what you guys have been able to show thus far. I think you said 50 hours or something in the lab running. Obviously much harder to be running on the back of a truck 24/7 in probably somewhat hostile environments. Talk to us a little bit about the main challenges from moving from the lab to the actual deployments here.
So for us actually, the 50 hours of operation that were done in lab do reflect realistic conditions. So they weren't as much like babyed along, if you may. The best part of our tech is when it comes to scaling up, it's putting multiple of these systems in parallel. So we already know how each of them behave. If you think about it, AAA batteries, you just put multiple of them together. You expect everything to function properly.
So it definitely is an easier task than, let's say, completely changing the scale and thought there. There's obviously some certain aspects to figure out. We have technical milestones six months and a year along the lines. The biggest one coming with UC Riverside providing us third-party validations. We already have those things in place as a part of our developmental program. So they'll be doing a lot of the certification as well as getting us the CARB licensing, permitting things along those lines. So the way I would say is what's remaining at this point is the engineering and the manufacturing aspects to figure out. The core science, the fact that we can deliver on it, that's already de-risked. But, you know, like shaving off that last pound on all the systems, improving the manufacturability, that's kind of what we are working on right now to get that commercial product out.
Yeah. What worries you most?
For us, it would be mainly, you know, the manufacturability side of it, because especially when we are scaling up and have 10 orders to go, there are certain manufacturing techniques that remain inaccessible because they only make sense when you have a production line and things like that. We are working with a lot of partners in the manufacturing space to see how to not just come up with a design that works when, hey, there's 10,000 orders here, but how do you start with 100 orders and you still have a reactor that's basically not much pricier. So getting to that commercial price point with the fewest amount of quantities and doing that as quickly as possible.
Yeah. And I think another thing to really emphasize is that our system is essentially just a different fuel delivery system. It's a lot simpler. You don't have to change any of the powertrain side. We can work with a fuel cell or an engine. The core powertrain, which is really kind of the hardest part of the vehicle, that we don't touch. We just replace the compressed gas tanks with our HRU and the LOHC stainless steel diesel tanks, and that's basically it.
All right, we are pretty much out of time. So, let's give it up for LNK Energies. Thank you.
All right, it's just going to take just a second to reset our stage. As you can see, we bring out some pretty cool stuff onto the stage for Startup Battlefield. By the way, was anyone here for the bear earlier? That was a new one for me. Judges, how are you doing? Is everything good? You feeling ready for the next company?
How did we like LNK? We don't have to give everything away, but—
I am an MIT alum, so I have a relationship.
I can't believe we didn't remove you for a conflict of interest. Jeez.
Okay. Well, we are ready for the next startup. That startup is Industrial Next. Presenting for Industrial Next is co-founder and CEO Lucas Pancow.
Hello, I'm Lucas Pankal, co-founder of Industrial Next. We're building robots for humanity's greatest ambitions. Over six years at Tesla, I've seen how much engineering effort goes into making robots truly useful. And with Industrial Next, we want to eliminate much of that work.
Imagine it's 2018 and Tesla has just launched the Model 3 and is ramping up production. And we were very proud of hiring the best system integrators to set up automated production lines to build as many cars as we can. The only thing that we didn't consider is that traditional robots are set to do pre-programmed things and do them very well, and they're very good at it. But we always had the belief that we should offer new innovations to customers as soon as we have them. And so we were pushing 300, 400 changes to the car every week, and the robots just couldn't pick up. And so towards the end of 2018, we set up a whole new production line in a parking lot behind the main factory, five weeks from a lot with potholes to the first car rolling off. And the Industrial Next founding team set up the first half of that line.
Fast forward to 2026, and it turns out that factories still require hands. They still require humans that are very adaptable and very intelligent about the kind of work that they're doing. Just in the US alone, there's almost 300,000 manufacturing facilities that do some form of assembly and manipulation. They're handling different parts that have very detailed requirements on how to contact them and how to place them accurately. And so this is the target market that we're going after first. And this is just the US. This does not include any facilities in much more manufacturing-heavy countries. And ultimately, in these factories, changing the product means changing the tooling, and where automated changes the programming and the validation of that programming.
Our product is really the AI that lets robots do human-level work. It's really the AI for dexterous robot, dexterous factory manipulation. The main part is our AI software, and we combine that with a compute and sensing stack, and then install it on a robot or work cell.
We're already deployed in consumer electronics factories throughout the world, and most recently at a factory making one of the world's most popular smartphones. In fact, to date, our machines have produced over 7 million of those smartphones already. So there's a good chance someone in the audience has a phone already built with our equipment. But there's a little story here. And it turns out about two months ago, we noticed that one of our machines was offline. And that was surprising to us because the machine was just performing flawlessly. It was building phones without really any errors whatsoever. It was fast to reprogram. And so we did a little digging. Turns out that the contract manufacturer decommissioned one of our stations, placed it in a secret portion of the facility building a phone that was not announced yet, to which we didn't have any access and our machine has never seen before, and just turned it on, and 60 seconds later it was just producing phones. And to quote someone I greatly respect, it made me feel very delicious, because that's exactly why we built Industrial Next—is to make this kind of work possible.
Factories ultimately buy a solution to solve their problems. And for us, if the factory already has robots installed, we can just source the AI software along with compute and sensing. And if they don't have any robots, then we source the software and also a work cell that we know is compatible with it. Really the recurring part is the software licensing itself, and we start with one task and then expand where the financials make sense. Our team is comprised of both factory and robot learning experts. I led a lot of the autopilot architecture for Tesla, along with many factory projects, and then the sensing and compute stack development at Waymo for the self-driving cars. My co-founder Allan was hired by Elon to work on his unstoppable alien dreadnought, and then before then developed self-driving cars at GM. And Gil, our CTO, worked on sim-to-real and dexterous robot manipulation at Meta, with stints at MIT, CMU, and University of Washington. Gil will show us what our AI does. So let's go to demo.
Hello everyone. My name is Gil Lee. I'm the CTO of Industrial Next, and today I'm going to show you one of our latest demos that is built with our own foundation model.
So here the robot has two hands, and then the task is simply to pick and place a diverse set of objects that are on a moving conveyor belt. And the nice—
The cool thing about this demo is that it has been trained largely with simulation data and then post-trained only with about two hours of real-world data. And during simulation, a lot of diversification was added, including object shape, belt speed, lighting condition, which really makes this type of task in the wild very robust. So at Industrial Next, we strive to bring robots to the real world. Thank you. Back to you, Lucas.
And back to slides. This was filmed in our real-to-sim-to-real arena just down the street from here, where we capture the task, reconstruct and train in simulation, and then test and improve in the real world. But where are we taking this? We're announcing our first partnership with BMW Group, launching with a few tasks in the arena to evaluate our dexterous manipulation technology for manufacturing. We're Industrial Next. We're building robots for humanity's greatest ambitions. We want everyone to build without limit.
All right, round of applause. Judges.
My first question is, you know, there are a lot of companies that are building foundational models for robotics right now. Did you build your own foundational model, and if so, why take that route versus post-training an existing model?
It's a very good question. We've tried post-training many models, and we benchmark every model we can get our hands on, including from our partners. And one thing that is a limitation here is that for most of the data that's available for most foundation model training in the industry, it's not specific to manufacturing, or it's not specific enough to manufacturing where you can get that repeatable, high-precision performance at speed. Right, because factories get paid for the number of units they put out, not just including AI in there. So it needs to work fast, but it also needs to be precise. And the only way we found to do this is by training our own foundation model. And the way to scale this really is our proprietary simulation approach. So what you saw in the demo was about 90% trained in simulation.
Very impressive. I think the smartphone example was a great one showing the flexibility of your approach. I would be curious, with dexterous manipulation operating in industrial facilities, you know, 87% accuracy or 50% as fast as a human is just not good enough. And so what does it take to go from use case one to use case two, and how quickly can you approach 100% or human speed, whatever the right KPI is?
Yeah, so one of the challenges with this kind of technology is that not every task is similar enough. But almost everything in manufacturing is based on some form of pick and place, which is where we're starting our work right now. And then everything else, like screwdriver, is an extension of pick and place, right? And the real challenge there for factories, especially the ones that are more traditional, is that if your equipment is not working at least better than the rest of the line, they're just going to yank it because it's something new. It's a risk, right? And the same goes to speed. If it can't keep up with other units, it's going out. And so we spent a lot of time optimizing for performance. For example, those seven million phones that I mentioned, they're running at about three-second cycle time, and that's running our AI stack.
Just kind of diving into that a bit more. Can you talk about the killer use cases here? Your background is in transportation and car assembly. You've talked about the phones. So what are the killer use cases, and how do you think about that go-to-market rollout?
Yeah, so for manufacturing, we have deep expertise in consumer electronics and automotive, and so we're focusing on that. And our partnership that we've launched today is kind of driving that home. But fundamentally, one thing that we've realized is that a lot of manufacturing types are actually kind of the same. If you go to a cell phone factory or a laptop factory or even a car factory where you're building something like a self-driving car computer, you can't really tell them apart even in many ways. It's the same with, you know, many modern phones. If you open them up, they're almost all the same. And so being able to perform a subset of those tasks, even something like pick and place where you can cover 70% or so of this, that already gives you something you can use in adjacent industries too. So there's something like farm equipment, for instance, like John Deere would be a good example, right? There's something like medical devices, although the challenge of medical devices, of course, is certification. How do you certify a system that is based on AI, right? That's also something we haven't quite figured out yet. But there's a lot of these adjacent areas and adjacent technologies that are already accessible for this. We just aren't starting there because we don't have personal experience in it.
Got it. Thank you. Congrats on the partnership with BMW. I would love to hear a bit more about your significant learnings from your customers so far.
I'm sorry, what was the second part?
Your significant learnings from your partnership with BMW or other customers?
So the biggest learnings, honestly, what we found is that the things that we worry about in the beginning are never the things that really matter. Fundamentally, it's, you know, we as engineers, we always spend so much time on the technology, right? We want to get the technology right. We want to get this working. We want to get this novel thing to actually do the task. But really, what they care about is, can we build what they need in the time that they need it, in the quality that they need it, right? And so it forces you to change a little bit your outlook of what it is that's actually important.
We have a super quick question. All right, if not, let's have one last round of applause for Industrial Next.
Time for the third startup in this session of the Startup Battlefield. That startup is Helix Carbon. Presenting for Helix Carbon is co-founder David Brown.
The steel industry built our modern society, and it's going to continue to build our future society. But that comes with some risk because steel production is the world's largest industrial emitter, responsible for 11% of global CO2 emissions. But 90% of those emissions come from steel that's been made in a blast furnace. And to understand why, we just need to do a little bit of chemistry. The problem is metallurgical coking coal, known as met coke, and it's layered into a blast furnace with iron ore. On their way down at incredibly high temperatures, they're struck with a hot air blast, reforming that met coke into carbon monoxide. And this carbon monoxide is the key reagent for converting that iron ore into the liquid pig iron that later becomes steel. But unfortunately, the carbon monoxide that's used becomes carbon dioxide. And there's no way around that. It's part of the chemical fundamentals of the furnace. Even if you had unlimited green energy, you couldn't solve this problem because it is a matter of chemical necessity. But the real issue here is that this blast furnace in reality looks like this. That is a gigantic piece of infrastructure. And more to the point, it is a multibillion-dollar asset. No steelmaker is going to tear that down and replace it while it still has usable life left. This blast furnace, commissioned two years ago, it could live to be over a hundred years old. If we are going to build a future based on green steel, we have to solve the met coke problem at plants just like this one. And that is what Helix Carbon is here to do. We developed a new electrochemical system that can be containerized and placed on site at a blast furnace. We can take in the waste gas without carbon capture and without gas separation, and we can react the carbon dioxide right out of it, creating the exact same carbon monoxide mixture that's already being used in the blast furnace. By injecting it back into the furnace, we create a carbon loop that not only eliminates emissions, but can actually offset fossil fuels. That is how we build the future of green steel. Cut the coke, keep the furnace. And I'd love to show you how this works. So we're going to live to CA, Helix Carbon's HQ in Cambridge, Mass, just outside of MIT.
And what you're seeing here is a half-kilowatt system, which is great for a proof of concept at the bench scale, and it could even be piloted at a smaller-scale furnace. Off to the left, what we have is a number of gas cylinders which allow us to mimic blast furnace gas exactly in accordance with our steelmaking partners. That gas flows through these tubes right into our system, where our unique catalyst is able to react the CO2 right out of it, creating that carbon monoxide-rich mixture you see here. That would be injected into the furnace. But in the lab, what we're able to do is prove that it's the exact same gas that blast furnaces are already using. So we know how valuable it is already. Thank you, Mike. Let's go back to slides. So the ability to react with CO2 right out of dilute and dirty waste streams is a major technical unlock, but the real advance here is that by replacing significant amounts of met coke, we can actually save the ironmaker money. About $14 a ton without green premiums and without any incentives. And in an industry that has extremely volatile margins that are also incredibly tight, an additional $14 a ton regardless of market makes a massive difference. Even over the last year, that would have been a 68% margin expansion to the tune of about $35 million. But even with these economics, the steel industry is incredibly conservative when it comes to new technology, which is why our go-to-market strategy is built around mitigating that risk. We want to sell modular units that can start small at a furnace with low risk and low capex, prove that it works, and expand that unit and expand it again, and continue to add larger units, each one providing a capex opportunity and additional recurring revenue for maintenance and servicing. In this way, a single furnace could grow from a pilot to a quarter-of-a-billion-dollar opportunity. And there's over 1,500 of these furnaces worldwide. That opens the door to a nearly $200 billion TAM with an additional $23 billion in annual recurring revenue. But in order to get there, you need that first pilot. Which is why I'd like to announce on this stage for the first time that Helix Carbon will be deploying an industrial pilot with a research furnace in Minnesota early next year. And 12 months after that, we'll be following it up with a commercial unit which is already under purchase order with a Fortune 500 company in the United States. This team is rapidly approaching commercialization. Evan Hos is our CEO. He worked with the Biden administration on their policy on their iron and steel decarbon policy. He knows what these steelmakers' pain points are. I myself am a former US Army Apache helicopter pilot and operations officer. I know what it takes to lead a team through complex, high-stakes operations and come through on the other side. Add to that three extremely technical PhDs and a slate of industry expert engineers, and we have a team that knows not only how to develop this technology, but to scale it and commercialize it and to get it into the hands of the people who need it the most. So if you want to be a part of a future built on green steel, come find me at booth CE6 and let's close the carbon cycle together. Thank you.
Congrats, David, on the pilot announcement. I'd be curious, my understanding is a decent amount of electricity goes into the process to create the syngas. Are there certain kind of electricity costs where this does or doesn't work for your partners?
Absolutely. So at the electricity cost that we have modeled, that's five cents per kilowatt hour, which is in line with what our partners have told us is what they're paying for on PPAs. Once you start to hit seven cents, the economics start to look a bit less favorable. But it's really about the arbitrage between the cost of electricity and the cost of that fossil fuel. So as the price of coke rises, a higher price of electricity is okay. We also are able to operate intermittently. So we can turn the system off as the price spikes, and as it starts to fall, we can turn the system on again. So we're always ensuring the lowest cost of delivered gas.
David, I love the tagline, cut the coke and keep the furnace.
Yeah. It looks like you have this incredible team that you've brought together, lots of different skill sets, etc. How did you guys land on this particular problem to solve, and what other team members do you think that you'll need to add over time?
We landed on this specifically because of Evan's experience. So we started with ultra-high-efficiency CO2 utilization. And that seemed, we took that into the market, and a lot of industries said, you know, I don't care how efficient you can make it. Carbon capture is not going to work here. It's too big. It's too expensive. And we said, well, what if we could do carbon capture without, or what if we could do carbon utilization without capture? And Evan said, I was working with the Biden administration on their steel decarb policy. This is exactly the problem that iron and steel has. Let's go find, let's go talk to them. And the results have been, results have been phenomenal so far.
I'd like to double-click on your initial pilot. How did it come together, and sort of what else constitutes your ideal customer profile?
So that first industrial pilot came together after a pitch just like this one. There was the right person in the room who specifically was like, this is what we've been looking for. And we've got the research furnace, let's make something happen. And it moved very quickly after that. I would also like to go, you asked how many people we needed to add. We do need, you may notice that we don't have any iron and steel insiders. Yeah, we are adding a board of advisors to go with our seed round, and we're going to be bringing on industry professionals for that as well. Sorry, I just remembered I had to finish.
One question on my side is, what's the business model? Are you selling the product, the hardware, or are you doing a rev share? What's the strategy?
Yes, we are planning to sell it as a capital expenditure to the iron and steel industry because that is how they're used to operating. They like to own assets. Frankly, they're really good at it. Their cost of capital is significantly lower than mine. And so we'd love to get that into their hands. And then these units will need yearly revenue, sorry, yearly maintenance. And so we're able to charge additional for that yearly upkeep.
What are the core kind of challenges between now and delivering on that pilot? I'm not clear exactly what scale you've been able to produce at today.
The pilot that is going to be delivered in Q1 next year needs to be, we simply need to build the plates and get the membranes and cut them. Technically, the risk on that is put to bed at a pilot of that size. That's going to be about a three-kilowatt system. The commercial unit will be about a 150-kilowatt system. So that is significantly bigger. But the good news about electrolyzers is that you can, I heard someone else talk about it as well, you can scale them up in parallel. And so it's not just about going up in the XY, although that is something we will have to do, and there is some risk associated with that. But you can add, once you get a plate working and a stack working, you can add additional plates, and then that's how you can grow the system. To go back to the team members as well, one of the people that we added recently is one of the, has built multi-kilowatt stacks before during his previous work. So that's a great add, but we're going to be adding additional team members who can own that process and build that out.
Quick final question.
Do you own the IP on the process? Really impressive.
We do. So there are a few pieces of IP that are trade secret right now, internal to the electrolyzer itself. But the loop, the carbon loop utilizing CO2 right out of the waste stream with no carbon capture and injecting in the furnace, that is Helix Carbon IP. That is under patent.
All right, give it up for Helix Carbon.
Thank you so much, David.
All right, now you may have noticed a little bit of secretive activity happening in the shadows there. One thing I didn't mention about Startup Battlefield is we generally require startups to do some form of live demo. And as Startup Battlefield has expanded, and we're not just doing smartphone apps anymore, that has become increasingly challenging. You've seen some of the companies have done that by zooming into their offices, zooming into their facilities. Sometimes we just bring some really cool stuff on stage, which is what's happening now. We have an incredible backstage crew that gets all of this stuff on and off stage without anything getting mixed up or anything like that. So I think we may be ready to go now. Okay, let's bring out our next startup, Nobia AI. Presenting for Nobia AI is co-founder and CEO Harine Carr.
Hello everybody. How are you doing? Hi, I'm Harvin. I'm the co-founder and CTO of Nobia AI. We are building self-driving factories where Nobia is your AI process engineer that helps you find and fix defects in hours and not weeks. China is able to scale a new manufacturing technology to market in less than two years. And in US, that takes us more than five. A big reason for that is the troubleshooting speed. China has a deep bench of process engineers that are expert and that are able to find failures faster and get the production lines back on track quickly. In US, we have the expertise, but we just don't have enough process engineers to do this at the same scale. And that shortage becomes a bottleneck every time something goes wrong on the line. A tiny defect on the line can lead to a product scrap or sometimes even a safety recall. And whenever a defect happens, a process engineer has to manually run the same loop: detect the defect, find the root cause, and then fix it. And just diagnosing root cause alone can sometimes take weeks to months. And while that's happening, your line is operating below your, below the mature yield. So our solution to that is to automate the process engineer's entire workflow. So we are building LabOS, where we do this by fine-tuning our open-weight models, that we fine-tune them on manufacturing and then connect them to the equipment directly. So LabOS understands your entire manufacturing plant's ontology and it connects information across different processes and is capable of multi-step reasoning. And Edge, our other offering, which directly plugs into your equipment, and these are industrialized equipments that understand that process exactly and they are able to diagnose the defects and troubleshoot the defects in real time. And we tested this AI at an injection molding plant where one defect, a specific defect that would take that plant two weeks to troubleshoot, with Nobia we were able to resolve that in three hours and got the line back up running. So let me show you how we do this. So let's go to the demo.
Okay. So this is a fluidized bed reactor that we have here. It is a benchtop version of one of the industry's most commonly used equipments. And we will be controlling this with our Nobia platform today. So let's turn on the roaster. Okay. So in the screen here, you can see that the beans are now fluidizing, which we just controlled through Nobia. And you can see in the screen here, we have live temperature reading coming from the roaster, which is that green curve here. So this is a recipe that's running at the moment. But how does Nobia figure out which recipe to run? That's what the experiments tab is for, where we have a Nobia agent that is trained on your manufacturing process. I just go to the agent and tell it the product spec that I need, and it figures out the best experiments that you need to do and make that product. And once the data comes in, the agent analyzes all the recipe data for you and gives you the best recipe, and that's the recipe we are running today. Now let's see if we inject any defect, what happens, how does Nobia catch it. I think we... Okay, sorry, technical difficulties. Let's go back to it quickly. So here, what we did was we injected a defect that happened, and within seconds, Nobia was able to detect the defect and the agent has started investigating. And the defect here was that we just cut off the heater, and we can see that the temperature started dropping. And once the agent figures out, we will see quickly an approve button where the agent comes back with the recommended actions. And once that comes in, once we click approve, we go back and it applies the fix autonomously. Okay, let's go back to the slides.
We started about a year ago, and now we are a team of five with four PhDs and one ex-co-founder with combined 50 years of deep tech experience amongst us. My co-founder and I have both lived these manufacturing challenges ourselves, and we have seen this pain. So exactly, we are building the product that we needed ourselves in our past lives. The way we price Nobia is 2% of the line capex of any production line that makes it completely autonomous. We are starting with thin film coatings, which is an $8 billion market opportunity. Scaling it to advanced manufacturing expands that and brings us closer to $200 billion, and then all of manufacturing, which is an $800 billion opportunity. We currently have 32 companies in our pipeline, three active pilots, and four LOIs signed. We have $65K logged revenue and $600K plus requalified revenue in the pipeline, which we aim to hit by next year. And quickly, a word about our technical roadmap. So most plants today operate at level zero autonomy. Nobia today is at level two, where our AI can autonomously detect defects and recommend fixes, but a human approval is needed. From there, we aim to be fully autonomous for coatings specifically by mid-next year, and then we go towards full autonomy, which is fully self-driving factories, and that is a north star. Thank you. With that, I come to the end, and come meet us at booth BW5 if you want to learn more. Thank you.
We have a rogue coffee roaster. Can you talk a little bit about how companies are actively looking for solutions in here, and what the competitive landscape looks like?
Yeah, so during our customer discovery interviews, a big challenge that companies are facing is that it takes weeks to months to troubleshoot defects on the line, and these are like advanced manufacturing companies where finding defects is really hard. There are multiple root causes that can lead to the same problem. So they are looking for a solution that can detect the defects faster so that they can ramp up quickly, and then they want the scrap actually getting reduced. So a lot of our competition in this field is they are working on similar things, but what they stop as is just finding the defect, but nobody's actually fixing the defect on the line. So we go a step further by integrating into the equipment directly.
Cool.
What are the common characteristics of the first design partnerships that you have, or the customers? Like, what did you optimize for in your ideal customer profile?
Yeah, so our early customers, one of the companies has a very advanced thin film coating line, and they are actually an equipment manufacturer. So a big problem for them is that they are unable to currently correlate their inline process parameters to their final products. So we are helping them build predictive models very quickly so that they can understand this and better improve their recipes.
I actually roast my own coffee at home, so good to see the demo here. Very relevant. What led you to focusing on thin film coatings as your first kind of customer segment here?
Yeah, so that's a very good question. Thin film coatings alone is very critical across semiconductors, batteries, medical devices, and it's so wide that that alone gives us an $8 billion market opportunity under this pricing model. And both my co-founder and I have deep experience with thin film coating processes, where we have scaled them from lab to high volume.
I'd love to understand what it means to have a fully self-driving factory, if you could just expand on your definition of that.
Yeah, so our vision is the equipment are able to run themselves. So what we do is we just give them a product spec, and they are able to figure out the right experiments that they need to do to be able to give you that recipe. And that happens throughout the factory, where every equipment is thinking for itself, and it's also self-learning, where any fault that happened, it remembers it and makes sure that it doesn't happen again, and optimizes your process along the way.
How far away are we from that vision?
So we aim to be fully autonomous for coatings by mid-next year, and then full self-driving factory is like 2028 is what our pipeline is.
Can you also share a little bit more about the implementation? What does that look like? Do you need to integrate into their source of truth systems, and then sensors-wise, like, what do you need to tap into?
Yeah, so one of the big focuses for us is to be able to talk to the factory hardware. So in that case, we are directly talking to the PLCs in the industrial equipments, and most of the industrial equipments have a standard communication protocol. So we can just directly integrate to that. And along with that, manufacturing is in general very fragmented across various data sources. So we have built connectors to apply all of those. So your MES, your LIMS systems, so we are connecting all of that.
You talked a little bit about fine-tuning open source models. Can you walk us through that tech stack a little bit more?
Yeah, so a big reason why we chose open source models is because all of our customers that we talk to, their equipments are not connected to the internet, and they don't want their data leaving their facility, and especially to the frontier labs for their training. So that's why one of the offerings is Nobia Edge, which is where we just host the models that we fine-tune and all our architecture in an edge device that just plugs into your existing equipments.
And what is the kind of length of process for going from, you know, DeepSeek or Kimi, Qwen model to something that's deployable at the customer?
Yeah, so for Lab Edge, we are specifically using small language models, which we can fine-tune very easily because the job for those models is to just understand one specific process. So we don't need to train on the entire line or in general like the entire world. We just need to train them on a coating process, on a specific process physics, and from there we can just work with small language models.
Final quick question.
All right, one more round of applause for Nobia AI.
Okay, as we get this coffee roaster off the stage, I have a quick question for our judge. Now, don't say a company, but I'm just curious. Do you have a favorite yet? Is this going to be a tough discussion or an easy one?
It'll be a fun discussion. Yeah.
That's a good answer. That's a good answer. And as a reminder, of course, you can all vote for your favorites. Just go to the TechCrunch events app. And, you know, I think there's going to be a number of favorites that you can submit both from this session, from this morning's session. And of course, there's going to be two more sessions tomorrow, and then the finals on Thursday. But of course, like, don't pick your favorite yet because we have one more startup presenting in this block. That startup is Li-Cube. Presenting for Li-Cube is co-founder and CEO Yohi De Kaguchi.
Like you, we make lithium for the future of humanity. AI will need 4,000 terawatt hours of electricity by 2040, which is the same amount the United States uses today. We need renewables, fossil fuels, nuclear fission, all in, but we still need a game changer: fusion, a sun on Earth, clean, safe, unlimited energy. That's why billions in private capital have gone into the fusion sector in the last few years. Lithium is the fuel for fusion energy. Fusion is a reaction of deuterium and tritium. It's called the DT reaction, using Einstein's E=mc² to produce enormous energy out of it. Deuteriums we can easily extract from seawater. Tritium doesn't exist naturally. So lithium-6 is the one to produce it. Sorry. Nuclear fusion needs nuclear-grade lithium. Natural lithium has only 7% of lithium-6 isotope, which needs to be enriched to a much higher level for nuclear fusion. We call it nuclear grade. Nuclear grade lithium has only been produced in very small amounts for nuclear weapons. Lithium has three grades. Battery grade, we use lithium for the batteries. Do you know what's the purity of it? This is 99.5%. Above that, space grade, 99.995% for advanced materials. At the top, nuclear grade, we stop and reach. We, Li-Cube, make the space grade lithium today and nuclear grade next. We founded December 2024, headquartered in Houston and lab in Japan. We have raised more than 50 million USD in equity and grant within two years of history. We have 30 patent filings over 20 years of research from the national university. One pass through our system purifies the lithium to space grade, and multiple passes enrich lithium to nuclear grade. One core technology can produce dual products of space grade and nuclear grade. We do not sell technology. We sell the lithium powder. So lithium feedstock comes into our black box plant, and we sell the white powder. That's how we keep our technology and know-how in-house. The market size for battery grade is a hundred-billion-dollar commodity market. Space grade is a niche premium market of $3 billion, which is a very important and profitable market for us because it requires only one pass through our system and is available right now. Nuclear grade will be a $50 billion market by 2040 when fusion comes online, and then we expect the first shipment to the customer is around 2029. What if fusion is late? That's the most frequent question investors ask me. We can still sell the space grade to aerospace alloys, AI and chip optical crystals, solid-state batteries, lithium metal anodes, psychiatric medicine, and more. All fast-growing sectors. So we can still make a multi-billion-dollar company if fusion is a bit late. And we are already generating revenues. Our pilot production, we have started, and for the first 50-plus customers, and we already have $100 million offtake interest for our commercial production. So the market is there. Our question is, like, when can we deliver actual products? So let's go to the live demo. Here's Professor Sodeyama, society founder and inventor of the technology. Hello. So this is our electrolysis system. Please show it. So this is our electrolysis system, like, running 24/7 for our sample production. Don't worry, we only show one system for you. And then this is a three-generation outdated. We never show you the latest model to the public. And let's go to the powdering section. Yeah, behind that, there's a powdering section. So lithium is a liquid solution we're producing and making into the white powder. And what he's holding is actual white powder we're selling as samples to the customers. So this is a pilot production. We're making a 100-times-larger demonstration plant in Houston right now, which is under confidentiality. We're not going to disclose it. Thank you so much, Professor Sodeyama. So let's go back to my presentation.
Excited about lithium. If you are still not sure about us, let's talk about the team. Myself, I did a PhD from the University of Cambridge, UK, and I founded the company which became one of the largest EV charging networks. And then I brought the company to the public and I hit the $1 billion market cap. I retired myself to Monaco, south of France, at the age of 32, and then I realized, oh, I'm too young to retire. So I put myself back on the stage, trying to address the most important problem in the age of AI, which is fusion fuel. With me, there are 30-plus people, mostly electrochemists, plant engineers, and serial entrepreneurs who have built a billion-dollar company together with me, backed by the US Tier 1 VCs, everybody knows the name, we haven't disclosed it yet, and strategic partners and a government grant. Li-Cube, making lithium for the future of humanity. Thank you so much.
