← This Week in Startups
Loading summary
Transcript211 lines
- [00:00]
A
I'm a little worried about how we're going to tell people we're fusing neurons
- [00:04]
B
with computers, the world's first biological data center. When they compared it against their reinforcement learning systems, the neurons we had were 5,000 times more stable, efficient.
- [00:13]
A
Has anyone complained that you're tinkering a bit with the edges of humanity?
- [00:17]
B
The Vatican were worried about this. You do not want to create conscious systems because ethically a conscious system has the ability to suffer. And we do not want any suffering to come about from any technology.
- [00:28]
A
This week in Startups is brought to you by LinkedIn.
- [00:31]
B
Post.
- [00:31]
A
Post your job for free@LinkedIn.com twist, then promote it to get access to LinkedIn jobs. New AI assistant Quo, formerly OpenPhone, gives you a clean, modern way to handle every customer, call, text and thread all in one place. Try it free and get 20% off your first six months at quo.com twist and deal founders scale faster on deal. Set up payroll for any country in minutes. Hire anyone anywhere, get visas, handled fast and get back to building. Visit deal.com twist to learn more. Hello, and welcome back to Twist. Now, today we're talking to a company we spoke to in 2025 because I thought they were one of the most interesting startups in the entire world called Cortical Labs. They are trying to fuse silicon chips that we all know and love with human neurons, bringing the biologic and the synthetic together to create a new type of computer, a biological computer. I was so tickled by the idea I have to talk to them. But since. Since that first conversation, Cortical Labs has built out data centers of its biological computers, which means that we now have real robust capacity to kind of bring humans and computers together. So to tell us more about what's been going on over in the realm of Cortical Labs, please welcome back to the show. It's my dear friend Han. Han, how you doing?
- [01:46]
B
Good, thank you. It's great to chat again, Alex, and congratulations. I heard you had a new baby.
- [01:52]
A
Yes, that's why I've been extra tired these last four months. But we're powering through just the grace of coffee. All right, so. So, Han, last time we talked, you had just put out your CL1, which was the first kind of like, fully contained biological computer with neurons and chips, and you were selling them, I think, for something like $35,000 a piece. So before we get deep into the tech, just for the business folks out there, how has that product performed in market?
- [02:18]
B
We've kind of exhausted our entire stock of 30 units that we had capped. So that's good. And you can work out how much that ends up becoming.
- [02:27]
A
Roughly a million.
- [02:28]
B
Yeah. And we have actually. So I'm in the US right now partly because we're fundraising, but at the same time I've also, you know, CEO stands for chief everything Officer.
- [02:39]
A
Right.
- [02:39]
B
I'm also now the company courier. I just dropped off a unit at Johns Hopkins with some of the folks there. I just came from Boston, so Mass General Guard unit. Another one got dropped off at UCSF. So now there is about five US institutions with a CL1 device, the other one being Dartmouth. They got an early PD early last year.
- [02:59]
A
So tell folks what the CL1 is and maybe because you have one with you, show us a little bit for folks on the video version, what it looks like.
- [03:08]
B
Yeah, absolutely. So maybe before I show it, like to give a very brief summary. The CL1 is our attempt to build a computing platform that allows researchers and developers to get going of biological computing very easily. It saves you the effort of building your own hardware, writing your own software and you just program these things with Python and you know, they're recommendable 3U unit systems. I have one here to show, so excuse me if the video is a little bit jerky. And this is a CR1 actually maybe from the front here you can see what they look like from the.
- [03:45]
A
It looks like a very long space age toaster.
- [03:48]
B
Yes. So they're actually very similar to rack rack. Multiple sleds that some of the GPUs come in the same format, form factor. So this is 3U. So in a server rack cabinet, this will take up about three rows. In a 45U system, we can pack 20 of these units and we have about 120. So they're in six racks. In our lab in Australia that we're calling the world's first biological data center, somebody came by and they were like, hey, don't you sell biological computers? And I was like yeah. And they're like, well, what do you call a place with a lot of computers? I don't know, data center. And they're like, well now you've got a biological data center. I was like, huh, that's really nifty. But I'm just going to show you what it actually looks like. So we open up the top. This is the neural chamber. So this is where we load up the compute unit. So neurons go into this chip. There's a life support system that is connected to it. There's a heating element underneath here that keeps it at a nice 37 degrees Celsius at around 100 Fahrenheit, we close this latch and think of this as our neural link. Right? So this is the neural interface. There's a compute unit at the back and, and then all of this stuff here is life support. So we build mechanisms to keep the system flowing, to give nutrients to the brain, to remove the waste. So we have pumps here like the heart. We have a feeding and a waste reservoir. So think of it like a stomach and a bladder. These are filtration units like the kidneys. We have a gas mixer like the lungs. And also the really interesting thing is you, you also have your traditional and non traditional IO units. So we look at the back here. This is USB C, USB A Ethernet. But you also have gas inputs, so filtered room air, CO2, nitrogen. And we have a waste gas outlet here to relieve the pressure. We jokingly call it the fart valve.
- [05:53]
A
I'm just laughing because we're talking about, you know, rack mounting and how much space it takes up in a server rack. And then you're like, and here are the lungs and here's where it expels waste. It's such an interesting combination of things that I understand but never see together. Now, how many neurons can you pack into a CL1? And then in compute friendly terms, how much power is that that you bring to bear?
- [06:16]
B
Yeah, so in a CL1 you can go all the way up to a million, 2 million neurons if you so wish. People grow organoids on these things and they have several million neurons for the cortical cloud and the offerings that we have, we go down to about 200,000 neurons. A, because that's a number that we can grow pretty easily and keep them alive quite well. Makes it commercially viable. But also we found that you can actually get some learning and training with that.
- [06:46]
A
So I want to make an analogy here because I think that everyone's very familiar with the idea of parameters in an LLM, you know, like 105B would be 105 billion parameters. If it's mixture of experts, yes, fewer than 105 billion would be activated, blah, blah, blah. But we kind of get that number. So 200,000 neurons. A million neurons, Same number to me, I have no idea if that's a lot, not very many. So how many neurons do I have upstairs and how many do you need to actually have the CL1 function as a biological computer?
- [07:15]
B
Yeah, I can't remember exactly the number, but I think it's like 100 billion neurons that you have in your. In your brain. So, you know, we have that and then a ton, more like trillions of synapses. So the closest thing that, you know, you can analogize what we have here is maybe a cockroach or, or a, you know, a fly kind of thing.
- [07:36]
A
So that actually doesn't tell me why I'm wrong here, hon. But cockroach is not famous for their intelligence, famous for their durability. But I mean, not, not for being super smart. But when you combine that number of neurons with a chip, a sil. What's like the multiplication factor that we get from bringing those two things together?
- [07:57]
B
Yeah, actually. So cockroaches are actually pretty smart, along with apologies, with bees and flies. You're not wrong. In some cases, they're not intelligent in the way we view human intelligence. They're not going to solve calculus. Right. But what they do solve really well is have you ever tried killing a fly? They're really hard to kill.
- [08:18]
A
So hard.
- [08:19]
B
Yeah.
- [08:19]
A
So quick.
- [08:20]
B
They're so quick and they're so agile. They almost like predict your actions ahead of time. And so this whole thing, and I think the industry needs to get this terminology right, we have accomplished super intelligence, like GPT, whatever, 5.5 is super intelligent. Is it generally intelligent?
- [08:38]
A
No.
- [08:39]
B
No. Right. Because what is it? Steve Wozniak has the best Tesla AGI. Can you walk into a stranger's kitchen and make yourself a cup of coffee? Everything is different, everything is new. You'll have to experiment with it. Right. You've never seen it before? We don't have that yet. So I would say that biology, even very simple organisms like a fly, has generalized intelligence, something that none of our machines have. So we're hoping to exploit those properties and even very simple systems, and maybe we can get a lot of stuff done there without having to go into the realms of, you know, human intelligence and all the baggage that comes with it, like consciousness and so forth.
- [09:18]
A
So when I. When I think about biological computing, I. I presume we're talking about a lot of use cases in drug testing, drug discovery, you know, biosimilars, blah, blah, blah. All the stuff that's kind of flesh and blood. But do you foresee a future in which the. Probably not. The CL1, maybe the CL3, manages to bring silicon and neurons together in a way that creates a computer that is better than today's GPUs and TPUs and so forth. At certain types of calculations that we use in the technology world versus the
- [09:48]
C
biology world, founders scale faster on deal. That's the deal. You can grow your company without borders and you can set up payroll for any country in minutes. Hire anyone anywhere like a modern startup or large company does. And Deal is going to get all the visas handled fast so you can get back to building. There's a great talent war that's going on right now and and you need people with superpowers for your startup to be competitive, to beat your competitors, to get your products to market. But anytime you try to grow your team with overseas hires, oh my lord, you've got to reinvent the wheel. And you gotta navigate a tangled web of international laws, regulations and you can't get these things wrong folks. You wanna onboard new staffers in other countries, you wanna get em set up on your network nice and secure IT access, all that good stuff. You wanna manage their benefits. Trust me, this is all a nightmare. Unless you partner with Deal. They are the people stack for startups. They're going to take care of all the onboarding payroll, hr it benefits, everything you need quickly in one place done perfectly. So visit deal.com twist that'S-E-E-L.com twist yeah,
- [10:52]
B
actually there is one that is already like proving to be better than CPUs or GPUs and that's in reinforcement learning. So this is some very novel work. It's still getting written up, so I don't really want to jinx it, but we're hoping to get it published later this year at Neurips, which is also happening down in my part of the world in Sydney. And what we've discovered is doing some work with a really strong research partner. The question was, can these neurons exhibit goal seeking behavior or pathfinding? Right? And the answer is yes. But not only was the answer yes, which shocked us, the secondary thing that they discovered was when they compared it against their reinforcement learning systems, their benchmarks, their GPU based systems, the neurons we had were 5,000 times more sample efficient than their GPU based systems. What that means is that for every step a biological system is doing, it takes 5,000 steps more to do on a GPU system. The saving grace with GPU reinforcement learning is that you can just accelerate time. So you just run time 5,000 times faster than the real world. The caveat with that is you can't accelerate time if you're a robot. You're operating at the same speed like everyone else in the physical world.
- [12:16]
A
So if biological computing is good at reinforcement learning, which as I think everyone listening to this knows is an enormous part of improving AI Models today.
- [12:26]
B
Yeah.
- [12:27]
A
You could end up pretty far outside of the realm of the biology side of industry. That's very interesting, but I feel like I've taken us down the wrong path. Let's back up and talk about what you guys have built. So you've built a biological data center. You have 120 units in Melbourne and you're going to do one in Singapore next, which I believe can get even bigger.
- [12:47]
B
Yes, exactly. So we were working with a data center company called Day One. You know, they've partnered with us because A, they like the technology for several reasons. A, because it is an alternative way to do compute that doesn't affect their energy budget. So this way they can say, you know, we have a data center operating at a very tight window specified by the Singapore government of 200 megawatts. They're going to provide the same chips like everyone else, but on top of that they're going to provide our compute, which is not affecting any of the energy budget because they don't have to do any special cooling and one unit only uses about 30 watts of energy.
- [13:28]
A
So a rack, which is basically zero for this compensation.
- [13:31]
B
Correct. So they're getting more for not much in terms of costs. So that is one of the reasons why they've partnered with this, but secondarily as well. And I think this is something that is a little bit mindbreaking as well. They've built not just the space that has a thousand CR1 units, but next to it, a laboratory for us to grow the cells for the compute on site.
- [13:57]
A
Oh, so you don't have to ship in your, your stem cell based. I point out, we're not killing people here to steal their brains. Stem cell based neurons. You can just make them. Make them, grow them.
- [14:07]
B
No, raise them, raise them, make them, grow them. Same thing. Yeah.
- [14:12]
A
Okay, sorry, it's my usual words don't work as well when we translate them over to here. I'm sure you've already gone through this, but for me it's still novel.
- [14:18]
B
Yeah.
- [14:19]
A
So what does that save in terms of like operating costs for you guys to not have to stick them in a cooler and fly them?
- [14:24]
B
Tremendous amount. But it also means that there is no supply chain constraint. And it decouples your data center from not just being a place where you just buy from a central vendor and you wait for it to come in and you, you deploy your chips to one where you're also manufacturing your chips on site. So you kind of decentralize the entire model where every data center is self sufficient. Every data center technically is not reliant on one vendor from one country.
- [14:54]
A
Yeah. Remind me how long the neurons live before they need to be replaced.
- [15:01]
B
So neurons can actually live a very long time if you keep them in a, you know, well kept, so to speak. The thing that does require replenishing are the tube sets. And the main. The main culprit for that is, are these filtration units?
- [15:17]
A
Here, let me see if you're on the audio version. He's pointing at.
- [15:21]
B
I'm pointing at these cartridges here. You think of them like kidneys. And what happens is that over time, these filtration units clog up with large protein growth factors and they kind of result in a bit like a kidney failure situation. So, yeah, we just swap them out and then we get another four to six months.
- [15:38]
A
Okay, so. So really then it's, well, what do you feed the neurons? I was about to say sugar water as a joke, but I think it might actually be a little bit wrong.
- [15:46]
B
No, it actually is pretty much sugar water.
- [15:48]
A
Okay, so basically you've made a biological system that keeps a small number, compared to our brains, of neurons alive that are connected to these chips that do a lot of cool things.
- [15:59]
B
Yeah.
- [15:59]
A
Here's my question, though, as time goes on. Do you think that as chips get smarter, we're going to need more neurons to interface with them? Or is this amount of biological compute on top of a chip enough? And we'll get gains just from improving the chip component of this. I'm trying to kind of figure out, like, putting some dots on the chart for where we're going.
- [16:20]
B
Yeah. So it's always a case of like push and pull. Right. So for instance, let's say we referred back to GPUs, right?
- [16:28]
A
Sure.
- [16:28]
B
We had way more GPUs than what we knew what to do with them before we got to the LLMs. And the breakthrough was the algorithm was a breakthrough success. And we were like, oh, we need more processing power. And so there's always this push and pull tension between. Are the algorithms there yet or is it a hardware limitation? Right. Now we see this as an algorithm's limitation because the bottleneck is how do you, how do you best represent digital information that is all on the Internet and so forth with analog systems. Right. Which is you and I. So that I think is still being worked on. You know, you have really smart people in neuralink and synchron working on that. For the BCI side, we're kind of working on it as well. But we have an additional challenge which we have to write information directly into the neurons. So, you know, to answer your question, we don't think we've saturated the system yet until we have enough capacity with our current setup. But who knows, maybe somebody cracks another new algorithm and then we're like, okay, we gotta get more cells going on the system.
- [17:31]
A
I'm just really excited about what you're working on because when we think about the systems we're using to make artificial intelligence smarter, we're throwing the equivalent of bodies at it, we're throwing more GPUs. The bitter lesson, Jevons Paradox, blah, blah, blah, blah, blah. But we're working on systems that are so fundamentally much more power hungry and less efficient and less attuned to the problems trying to solve than our brains, the things we already have with us. And so to me it just seems very logical that as the chips get better and as we make get better algorithms to make the neurons function as we need them to, we should be able to have two different intelligence curves working in synchronized fashion and we should get much faster gains.
- [18:13]
B
Yeah.
- [18:14]
A
And this is when religion comes into it. So as background hon, I was raised in a very conservative Christian church.
- [18:24]
B
Yeah.
- [18:24]
A
And so growing up in the 90s, I heard a lot about stem cells and cloning and a lot of just what I would call fear mongering.
- [18:32]
B
Yeah.
- [18:32]
A
I'm now a non religious science fiction nerd, so I'm pretty much your biggest fan. But I'm a little worried about how we're going to tell people that we're fusing neurons with computers. Because I think you've now taken this from proof of concept when you guys played Pong to early commercial with the CL1 to playing Doom recently, that was cool. To now building out data centers in an international format. So this is coming to market now we can talk about this sort of thing. Has anyone, the Pope or similar, complained that you're tinkering a bit with the edges of humanity? And is anyone worried?
- [19:09]
B
Actually the Vatican were worried about this. But fortunately Brett, my CSO has done an excellent job engaging with bioethicists and actually being at the forefront of this. Right. And so I think because of the space that we're working in and the fact that there's a lot of ethics that need to go into it, even just doing research work with any biomedical aspect to it requires a ethics board. We're very attuned to these kinds of potential criticisms. Right. So we, we try to engage them proactively.
- [19:45]
C
If you're growing a startup or a small Business. You can't sleep on incoming calls, even if they come in after hours or on the weekends. That's some of your best customers calling you when they need you. That's why today's episode is brought to you by Quo. Keep Q U O the smarter way to run your business communications. Quo is the number one top rated business phone system on G2, and it's trusted by more than 90,000 businesses. They're going to bring all of your calls, texts and contacts together in a shared collaborative space. That means your entire team is going to use one shared number if that's what you choose. And there's no more missed calls and no more disconnected conversations. I've been using this product for well over five years across three different business units. It's perfect, it's affordable, it's feature rich. And the team is so sharp, they keep releasing new features that I didn't even know I needed. Money is on the line. Always say hello with Quo. Try quo for free plus get 20% off your first six months when you go to quo.com twist. That's quo.com twist.
- [20:49]
B
And I think, you know, firstly, the most important thing that we're trying to work on right now is to get everybody in this space, plus the consciousness space, to come together and actually agree on a shared nomenclature. Right? Because if you cannot agree on the thing that we're all studying on that we have no chance of actually making any progress and all we get is just a lot of fear with not much understanding. So we're trying to do that. And also, yeah, I think actually Brett was in discussion with the Vatican who wanted to find out what's going on here and were there any religious and ethical issues. And, you know, fortunately they actually agreed that what we were doing was all right and it was actually, you know, fine. Yeah, because. Because ultimately, the main question that we all we need to have is, you know, is there a harmony part of the plan? And also, you know, the principle of double doctrine, is there actually a net good that can come from this technology versus a net negative? And so that's the reason why, you know, you started out with the traditional drug discovery, disease modeling and stuff, because that's still a very important part of the work that is the primary use case for all these labs purchasing these devices, which is, you know, so, you know, in Mass General, it's a Alzheimer's dementia researcher at Hopkins, they're looking at toxicology and alternatives to animal testing. At ucsf, they're looking at, I think they're looking at movement disorders and a whole bunch of other. Like. Really?
- [22:13]
A
Oh, that would. Nice.
- [22:14]
B
Yeah.
- [22:15]
A
These are all. By the way, this is why this technology, even in its current form, is freaking awesome.
- [22:19]
B
It's.
- [22:19]
A
He just listed three major medical centers in the US that are using this technology now to make our lives better. So, like, this is. The dangers that are hypothetical down the road that we're touching on are not meant to undercut the current usefulness and commercial application and research application of this product. I'm just having fun.
- [22:36]
B
Keep going. Yeah. Thank you. And so you know, that is. It's all happening right now. The compute side of it is by all means the smallest, but the newest, the one that's also most forward risk because you never know what people are going to build. So this is something that we're trying to keep a hold of. And I think internally at the company, it's really important to understand this whole. This discussion of consciousness because I think we've drawn a red line for us, which you do not cross. The poem is, where is this line? And we got to figure out where it is, is consciousness. You do not want to create conscious systems because ethically a conscious system has the ability to suffer, and we do not want any suffering to come about from any technology. So that is the stance that we take. We try to do as much as we can when we have the cloud and we can monitor things, but when these things go out of the. Into the real world, we don't really know what's going on. But fortunately, most of our users, actually 90% of all purchases, have all come from academic R and D institutions that we trust that they do the right thing.
- [23:36]
A
But the cloud, you guys have built the data center in Melbourne or Melbourne or whatever, and then Singapore, we're going have. If it's 200,000 neurons apiece and we're going to have, I think it's up to a thousand in Singapore. You can do the math. It's quite a lot. The reason why it doesn't scare me is it's all kind of like cut up into little pieces. Like each one has X number of neurons. But there is the science fiction question of if you have a data center full of these and they can interact, does that change the math? But the good news is that I think we're pretty far away from having so many biological computers in the world all plugged in together that we have to think about that. So getting back to practical applications here, why go the cloud route as opposed to just selling the devices? Because clearly, if you sold out your first run, plenty of demand. Why? Yeah, so cloud. I'm just curious about why that was the right commercial approach to the market.
- [24:26]
B
Ultimately, what we really want to do at Coracle Labs is to, you know, I hate the word democratize, but it is democratize the technology by reducing the accessibility barrier to biological computing. Right. Because, you know, ultimately, you know, I. I always go back to Nvidia. Right? Because ultimately the AI that we have today is actually an accident. It was completely a fluke that had happened. Right? Because, you know, everyone's like, yeah, Jensen was so smart, he's so brilliant about this. I was like, yeah, if he was so brilliant, why didn't it take seven years from when CUDA was made to when Alex Kraszewski made Alexnet? Right. It was a sheer serendipitous moment where somebody was Jeff Hinton's grad student who was solving for image recognition technology, who had a GPU and you had a program, cuda. So I like to think we would
- [25:12]
A
have had some other serendipitous meeting point along the way, but that is how it happened. It does feel a little tenuous, a little fragile when you look back.
- [25:21]
B
Exactly right. And so what Jensen did, I think brilliantly at the start, was he made CUDA free and made it so that any gpu, even the crappiest gaming gpu, could also run cuda. And so the accessibility barrier was really low, so anyone could get into it. So that's what we're trying to do here with the cloud system. Because unfortunately, if you have a CR1, you really can't do anything with this unless you have a lab. You can grow cells and you can keep them alive.
- [25:49]
A
Right. So you need to have a person who's doing the kidney replacement and the feeding and the waste extraction and so forth.
- [25:54]
B
Exactly, exactly. But what if you don't have any of that, but you have a great idea you want to experiment with? So the idea of the client was born when we said, let's just give it to people with ideas so they can muck around with it. And that's actually the story of Doom. We actually didn't build Doom. Doom was actually a student developer who participated in a hackathon at Stanford. No biology background, used our APN SDK and he built something that was really cool. And so we decided to help him tell the story. And word got around and I think he's now been accepted into a very prestigious incubator program I shall not name.
- [26:35]
A
Which could it be called? Kai Bombinander.
- [26:40]
B
Or rhyme with that can either confirm nor deny.
- [26:44]
A
Got it, Got it. So last time we talked Han, we were talking about the first application of your technology to a video game, which was Pong.
- [26:53]
B
Yeah.
- [26:54]
A
And for the young people, Pong is a game when you have two paddles on either side of the screen. They go up and down and knock a ball back and forth. Yes. We used to think that was fun. It still is, frankly. But you told me that when you, you built the system, you had to kind of like create a reward or punishment mechanism, I forget which. To give it the incentives to learn how to play correctly. Right?
- [27:14]
B
Correct. Yep.
- [27:16]
A
In Pong, not a lot of variables. In Doom, many more. So do you know how they managed to set up the reward or punishment mechanisms to actually interact with a game of that complexity? Because I'm curious if it was hard or not.
- [27:30]
B
Yeah. So it pretty much is also the same. They're using audit stimulus versus disordered stimulus as the reward and punishment signal. There were a few more variables. And you know, honestly it's kind of funny because I think they didn't, they didn't disincentivize or punish it for wasting ammo because there was no ammo restrictions. Restrictions. And so essentially what it did was it figured out that if it just spammed the shoot button and just spin around in a circle, it would just win. So it's kind of funny because that's what happens in actual reinforcement laying systems as well. So when we changed actually was I guess punished for wasting ammo and all that stuff. And I started to learn and actually started to have some interesting gameplay from that.
- [28:18]
A
I just realized why it's very important that you don't create conscious systems because they can feel pain. Because if you're using disordered inputs as a punishment mechanisms, a little bit harsh, everybody working on the terms. But you wouldn't want to do that to a conscious system, correct?
- [28:33]
B
Exactly.
- [28:34]
A
Yeah. Okay, that makes that I've connected that and that makes a lot of sense. But here's the takeaway. I mean, okay, look for Doom's great. If you haven't played Doom 3, do it. You owe yourself to have that fun. But if a kid and I say that with love can build this with your API, to me it shows that the technology is not that hard to bring to real world application, hence the cloud. And that brings me to my last question, which is how many people can you customers can you serve with 120 CL1s in a data center? 1 10,000. I literally have no idea what the range is.
- [29:12]
C
Adding a new member to your team is a crucial decision, and you don't want to rush into a hiring situation that you will regret. But if you're a busy founder, you also don't want to spend a ton of time in the trenches looking for that perfect candidate. Instead, you need a partner and you need a trusted partner. And LinkedIn Hiring Pro is that partner. How do I know this? Because I use LinkedIn Hiring Pro. You want a real world testimony. All right, here you go. We recently hired a new customer success manager. They handle all the advertising accounts here on this Week in Startups and they ensure that we're keeping all of our wonderful partners happy. LinkedIn helped us connect with an amazing candidate right here in Austin, and he had exactly the combination of experience we were looking for. It's magic, it's alchemy, it's everything for me to find a great team member. So hire right the first time. Get started by posting your job for free@LinkedIn.com twist. Terms and conditions apply.
- [30:08]
B
That's LinkedIn.com twist so we have 120 that we've deployed. Unfortunately, there has been some teething with the biology. So while we do have 120, only about 20 or 30 users are on the cloud at the moment. Actually, 20 users. We have another 10 more that we have to bring online. But that's also because there's a whole conveyor belt system of getting sales growing and all that stuff. And it's one of those things where unfortunately, because it takes some time for it to grow, what we see now is the result of decisions made two months ago kind of thing. So they're only starting to come online. We've had a slow sort of start, but yeah, a lot of. So no, a lot of people have started to come on board and the waiting queue is getting smaller. So, yeah, we have that. We actually have a couple of corporates and partners coming on board to do some experimental work. But yeah, yeah, but so when.
- [31:06]
A
When doom is running on your biological computers. Is that one Cl one? Is that five? Like, how much does it take?
- [31:14]
B
One? It was just one. So we haven't come up with the. With the chaining of the systems yet. That's coming in, I think two.
- [31:19]
A
Okay, so each one's still a discrete box. Okay, that then my question was actually relatively silly. Then you could serve 120 customers at once.
- [31:25]
B
Yeah, exactly. I mean, sorry, that's fine. Because, you know, this is where we. What we're doing at the moment. There has been some Internal research work about, you know, where we, we're doing PDMs, microfluidic wells. So we can actually segment out the, the surface of the chip and you know, rather you get all 60, maybe you only need 16. And that way, I don't know if you're familiar with like VMware, maybe you can get a hyper, like a virtual machine with multiple people running the same thing. So that way we can increase yield and stuff.
- [31:55]
A
PDMS is a polydimethyl siloxane.
- [32:00]
B
Yep.
- [32:01]
A
That's which is also known as a dimethicone. What, what are we talking about?
- [32:06]
B
It's a material, it's like a gel type of material that you can make microfluidic devices with. So think of it like a, like a. Well, actually I'll show you what it looks like.
- [32:17]
A
So because, and then define microfluidics for me because again I nod my head like I understand, but we are outside of my comfort zone. This is not SaaS economics, so you're going to have to help me along.
- [32:28]
B
Okay, so I'm going to show you this. So this is a microfluidic device. So there are probably several thousand neurons in one well and they are now segmented into these little sections and we actually have channels that we can turn on and turn off. So we can actually control this more like a circuit.
- [32:49]
A
Now what we're looking at here, if you didn't know what it was, you would think these were top down views of water storage tanks with pumps and pipes coming into them. But instead, each one of these little octagons, each one of these octagons has a bunch of neurons in it. And the microfluidics allows you to feed and then also transfer information between them.
- [33:08]
B
Yeah, the microfluidics are these channels and what they do is they constrain the growth of the exon, the long part of the neurons. So they can only communicate with their nearest neighbors, or we can, you know, block them off and say you can only communicate with your east west neighbor or your north, south, mid. So you have a lot more like circuitry, like control over that.
- [33:29]
A
Right. And then you could in theory segregate to have more than one person using the same computer because you can essentially not let the axons get too busy sharing information. Got it?
- [33:38]
B
Correct. So this is kind of like a chiplet model that we've been experimenting with. So we can say, all right, you can have this one here, you'll have 1, 2, 3, 3, 4 electrodes to work from. And if your workload is pretty low, Maybe this is just good enough and you can actually have, if you want, get two, and then you can have them sort of network up like this, and then we can share the instance and stuff like that. So there's a lot of things that we're working on internally to try to figure out, hey, can we increase yield? Can we get more interesting, like learning some. That kind of stuff.
- [34:12]
A
So how much are you raising? What's your target?
- [34:14]
B
Yeah, we're raising. Oh, and actually one more thing. We're raising 30. But I also wanted to show you guys this. So this is the cortical cloud. Ah, yeah.
- [34:24]
A
If you're on the audio version, it looks like a standard developer backend, which is a compliment.
- [34:30]
B
Yes. And then you just hit the instance that you've acquired and there you go. I'm in New York right now, My lab is in Melbourne, and I am live streaming neural activity from a cell culture that my team have assigned me for demo purposes. And what is it doing?
- [34:45]
A
What are these dots that I'm seeing? If you're on the audio version, imagine you're looking out the window of a spaceship and you're seeing stars go by. That's what it looks like in slow mo.
- [34:52]
B
Yep. So the window with stars, this is what we call raster plots. So as we go down, you can see the channels are, like, increasing in number. And if you look on the right side, that's the topological mapping of it. So that's the surface of the chip. And so every time there's a spike or an action potential, or we're marking it on this grid here, and we're also marking it on this. So you can actually see there's like a temporal correlation. You can see bursting activity, you can see synchrony on this view. So, like, you can see this cluster here, there's another cluster there. And, you know, after a while, you can watch this. Oh, look, they're getting really active here. You can just watch this and sort of like figure out what's going on. And then you can also look at the activity views. So here we can actually see the raw waveform. So this is kind of the same thing that a neuralink would be picking up. And if we wanted to give them a bit of a poke, we can just say, all right, I'm going to poke the neurons in this region here, and I can click and I've delivered a stimulus across, and then that wakes them up a bit going on.
- [35:55]
A
It's really good. It's not conscious, otherwise that would be kind of rude. Like I was sleeping.
- [35:58]
B
I know, yeah. How rude. Now is a good dream. There we can see like the spike waveform. So we can actually see the, you know, what, what the shape of them are. So you can verify that they're actually really action potential spikes. And the cool thing here is that if you're a developer, by the way, I'm on Team Human here, so we still need developers to think of really cool applications. LLMs are not going to solve this for us. You write all the application code in Python and they're all in jupyter notebooks. So you specify what programmers actually do with CL1 is think of them as mini architects from mini matrices. Right? So you're building the matrix that the neurons are going to be put in. You specify the parameters of the game, the reward and punishments, the objective functions and so forth, and you just let them get connected to it. And so yeah, this is how you can do it. Documentations on docs.corticallabs.com, there's an SDK you can pip install and off you go.
- [37:02]
A
You know, it's one of the best parts about talking to founders that I get to do on is just straight up getting to see the future. Usually it feels like I'm looking about 12ft out. This is one of those chats which I feel like I'm looking 12 years out. Like there's, there's enough optimization, expansion, programming languages, how to teach these things, how to. Like there's so much here that isn't done yet. Oh no. I feel like you have like. Well, you have your life's work ahead of you. This is going to be, this is going to be a lot. Okay. I gotta let you come back on in six months and tell us how the cloud's going and how Singapore is going because apparently we're just going to have you on all the time. But for folks who want to learn more, what's the URL and is there a job you want to shout out to the audience in case the right person is tuned in?
- [37:44]
B
Yes. So check us out@cortical labs.com the cloud is at cloud.corticallabs.com documentation docs.cortical labs.com I'm on Twitter as well, Dr.1337 and also we are, we're looking for talented developers. Stay tuned to our Twitter feed. There may be a hackathon coming out at MIT soon.
- [38:07]
A
Are you going to go to that?
- [38:09]
B
Yeah, probably. We'll have to go for that one.
- [38:11]
A
So I only Live like half an hour away from that area. I wonder if I could maybe I can speak Wag. That'd be fun. I'd love to see it.
- [38:18]
B
Absolutely, yeah, for sure. Yeah. All right, so stay tuned. And yeah, more developers channeling the energy bomber. Developers, Developers, Developers, Developers.
- [38:27]
A
At least you're not wearing a sweaty collared shirt. Han, thank you so much for coming back to the show. We'll see you again soon.
- [38:33]
B
Thank you.
- [38:34]
A
We're talking about drones. No, not drones that sit on the water or go beneath the water or just fly up in the sky carrying a single hand grenade. No, today we're going to talk about very large drones, mostly outside of the battlefield context. So I rung up my dear friend Michael Nortra from Pica to come on and tell us about why drones are good for agriculture, we're why they're good for near term cargo transport, and why they may soon be dropping couches from the sky onto your head as long as you're within a radius of 50 meters. Please welcome to show, it's Michael from Pika. How you doing?
- [39:03]
B
Awesome.
- [39:03]
D
Thank you for having me.
- [39:04]
A
Doing well. So I'm, I'm really excited about your company because I thought of you guys as the company that makes large drones that are powered by batteries that sprays crops. But as it turns out, you guys have really broadened out in the last couple of years and do quite a lot more. But Michael, to get to that point, I feel like we should go back to the future. So take me back to the start of the company, the initial vision, and how you guys literally quite got off the ground.
- [39:26]
D
Yeah, great question. So honestly, the start of the story begins when I was like 4 years old. I've been just incredibly passionate about aviation my entire life. So I started, you know, with paper airplanes when I was about 4. Rubber band powered airplanes, little electric planes, pretty much obsessively built aircraft throughout my childhood. Ended up studying physics and then I got to work on a number of different evtol projects in the Bay area. Three or four of them actually. So mostly passenger carrying, vertical takeoff and landing air taxi concepts. Super fun, like, you know, really hard technical challenges. But it was clear this was like almost 10 years ago. There was just no way we were going to certify those aircraft for commercial use for another like, realistically two decades. Yeah, that, that took the wind down my sails. So decided to leave to start Pika. And yeah, that was, that was now almost 10 years ago. So it's long journey.
- [40:19]
A
When was the first flight and how far have you guys gone in terms of like commercial penetration of the agricultural sector with your drone that can handle crop springs? I'm curious if you've managed to gain material market share in that industry.
- [40:33]
D
Yeah, yeah. So first flight was actually like a week before demo day. We went through Y Combinator. So we, it was, it was a wild time. We spent one month designing a sort of 600 pound, what would be about a two passenger autonomous aircraft. We built it in two months, actually a little under two months, and flew it on like week 11 of the whole program a week before demo day. So that was awesome. It actually flew itself. It took off and landed running our own software. You know, great testament to the realities of hardware. Like we had a flying prototype, autonomous big UAV 11 weeks in. That was I think literally like 1% of the work that we've done now. So the other 99% has been going from prototype to now a widely deployed aircraft that is operating with customers literally seven hours, sorry, seven days per week. Some customers are doing 12 to 13 hour shifts. So yeah, where we are today, believe it or not, we're actually the only company on earth that I'm aware of that have deployed a Group 4 UAS to commercial customers at scale. And interestingly, this is happening in the like rural jungles of Brazil. That is the only place on earth that big drones like this are being fly, flown commercially.
- [41:55]
A
That's weird. Talk to me about group four. I don't think that's a particularly well known metric outside of the UAS space.
- [42:00]
D
Yeah, so it's a more like military classification of drones. It's basically drones that are bigger than 1320 pounds. So you know, things like the MQ9 reaper, that's the most common group for UAS.
- [42:15]
A
All right, now you said you'd only done 1% of the work when you'd gone from no plane or no big drone to having a big drone that can fly. A lot of people out there that watch this or listen to this are software founders and they're probably familiar with. If you code up a feature that does not mean product market fit, it doesn't mean the sales cycle is over. There's still a lot of work to be done once you have some code. But for folks outside the hardware space, tell them why it was only 1% of the work. Because from the outside, going from no plane to plane sounds like you made a lot of progress.
- [42:43]
D
Yeah, totally. So I mean the thing with our products are a blend of hardware and software and so really like the last nine years have been this just steady march of maturing both the hardware and software in parallel. The hardware I think is the particularly difficult one to mature quickly. So that first aircraft couldn't actually complete an interesting customer mission. It didn't have the features it needed to spray a crop. It couldn't really move cargo. Even so there was a lot of work there. But I think the most interesting thing really is actually what's been happening in the last three or four years for us with commercial deployments, with customers. So a prototype can be operated by n number of engineers. It can be operated by your entire company if need be. You know, your uptime requirements are non existent. If you fly for one hour in one week, you'll be like, hell yeah, we got an awesome video. That's all we need.
- [43:43]
A
Put it on YouTube immediately. We're raising the next round.
- [43:46]
D
Exactly. With customers it's completely different. You know, our customers are very, very upset if the aircraft is down for more than 24 hours. And these customers exist in like extremely remote regions of the world. Like they are literally in some cases a six hour drive from the nearest city. And so getting to that sort of level of uptime takes an extraordinary amount of time and effort. So it's, you know, it's deploying multiple aircraft with customers, flying them, learning, retrofitting, fixing, re engineering over and over and over again until you achieve, you know, it's like the equivalent of product market fit is a tool that a customer can actually rely on.
- [44:30]
A
Yeah, yeah, product rely, no market. Product reliability fit maybe we could call it.
- [44:36]
D
Yeah, exactly.
- [44:37]
A
Why are we not using more of your ag drones here in the States? You talked about, you know, Latin America, remote regions. I know that agriculture is different across different areas based on what we're growing and seasons and so forth. But you know, I, I, I've laid drip tape on a farm, I've driven large pieces of equipment and to me like going above the crops here would make tons of sense. And your system seems to be very robust and it's got a good spray width and why can't I use this in Nebraska? Or why aren't you?
- [45:06]
D
Yeah, yeah. So it's coming, it's coming fast. So the, the things that drew us to Latin America were twofold. One is regulatory. So Brazil actually deregulated agricultural drones about two years ago. Just very convenient for us. We do have commercial approval to operate our aircraft here in the United States. We actually have the largest drone approved for commercial use by the faa. But it has some cumbersome limitations. We can only Fly the drone about 4km away from where it took off and landed from, which is not commercially viable today.
- [45:36]
A
No, that's ridiculous.
- [45:39]
D
It has to do with like line of sight to the vehicle.
- [45:42]
A
And oh, Zipline told me about this, that if you want to fly outside of line of sight, there's another entire set of regulations on top of that. But it works in Brazil, right?
- [45:51]
D
It does, yes.
- [45:52]
A
Is the air different in Brazil compared to the United States? Does something change to navigation and computing when you cross the border?
- [45:58]
D
Yeah, actually, it's a totally. No, it's not. Yeah. So in Brazil, we're operating typically anywhere from like 5 to 15 kilometers from where we took off and landed from. So, I mean, the good news is we have data about this. No one has ever done an operation like this before. And so the FAA understandably is sort of apprehensive about what that looks like. We've been working with them over the last year to actually remove that limitation and expand it to more like what we're doing in Brazil. Making really good progress on that. So like I said, we're, we're coming to the US we do have one customer already here who's, who has an aircraft, but we're going to have far more in the coming year.
- [46:34]
A
All right, so I think this is a good time to talk about fuel because it's going to come up when we go over to dropship in a minute. You currently make mostly electric drones, and I know that when you sell the Ag drone, you send with multiple sets of batteries so you can kind of hot swap them and move them in and out. But if I'm six hours from a town, what I probably don't have is as good of access to the grid. So to me, it's lovely that I don't have to cart fuel in a truck up some mountain roads. But I'm curious about just recharging these damn things out in the woods or mountains.
- [47:04]
D
Yeah, yeah. So, so farms have electricity. Like, you know, some farms are pivot irrigated. For example, they have these giant like circular pivots that go around big water pumps, blah, bl blah, blah. So the farmer will always have electricity. The question is whether they want to run the electricity from the sort of center point to the actual Runway. And I would say half of the customers that we have today do that. The other half use a diesel generator. But what's really remarkable so the, the kind of competition for our aircraft is a vehicle called the air tractor. This, it's a big, you know, roughly 8,000 pound vehicle, it burns roughly 55 gallons of jet fuel per hour, which
- [47:46]
A
is not small at today's prices.
- [47:49]
D
No, yeah, that's a lot. The Pelican in the worst case, if you're charging it off of a diesel generator is about 2 gallons per hour.
- [47:59]
A
That's a lot better. Even if you're doing dirty charging, as you might call it, it's still much more efficient. Okay, now I know that you guys say on the site that the current Ag drone starts at, if memory serves, 550,000. As far as farm equipment goes, that's not super crazy, but it's also not that cheap. So I'm curious about the, the ROI here and kind of like time to repay the purchase because the way that I think about it, more efficient flight probably need fewer human pilots because it's autonomous and you're using electricity and blah, blah. So I presume there's some savings baked into this. What is the time to recovery for people that are changing over to this type of ag work?
- [48:37]
D
Yeah, that's a really good question. So it's, you know, depends highly on how, how heavily you utilize the vehicle. So that's sort of one other benefit of Brazil actually is they're just crazy about the way they do agriculture. They have multiple seasons back to back and they just have an intense culture of like work. So you know, we have customers who, one of our newest customer is planning on running three shifts with the Pelican. And so the thing at most could operate literally 24 hours in one day.
- [49:10]
A
How comfortable would that make you as the guy who has to take the phone call if it goes down while it's working 24 hours a day? Because that sounds like a maintenance nightmare, but maybe I'm overestimating.
- [49:19]
D
It's intense. But I mean we're already, we have customers doing 12 hours a day right now. So the difference between 12 and 24 is it's not that big.
- [49:28]
A
In time are we going to see like smaller farmers be able to like do ride share equivalent of rental for these things to, to handle their crops? Because I don't think everyone's going to need their own personal.
- [49:40]
D
Yeah, correct. Yeah. So in the US for example, you would need to have a farm that is almost 20,000 acres in order to utilize, to fully utilize the aircraft. And that's huge. That is a, by US Standards, very, very big. By Brazil standards, that's actually like mid size. So yes, in the US it'll be much more common to have spray as a service where a contractor owns the vehicle. And then services a number of customers in a surrounding radius.
- [50:09]
A
We can make an acronym out of that. We could call it SaaS. No one's ever used that before.
- [50:12]
D
Yeah, totally. Yeah. Maybe our valuation would go up. Anyways, so you asked about economics and, and payback period. So yeah, it depends on the utilization. Two to three years, roughly. But the thing that people don't understand is pelican. Yes, it is a lower cost solution. Yes, there is a different type of labor. You use a much easier to access labor pool. You can train someone to operate a pelican in two to four weeks versus a aerial application pilot is 18 months. But really the reason people love the pelican is because of its ability to spray really, really well. And so, you know, the simplest way to understand this is the cost of the chemical that is being sprayed is typically about four times the cost of the application. So it's this very precious resource. Oh yeah. And you're trying to deposit this fine mist of chemical very evenly over an entire crop, including around the boundaries of the crop, which is very difficult. So the pelican has like a bunch of things about it that make it just a like massively superior solution for actually applying the chemical.
- [51:24]
A
And right now, I mean, as you and I record this, the straight of hormones is still totally blocked off. And I presume that behind this comes out, it'll still be blocked, which is killing fertilizer prices around the world. So people are probably extra price conscious right now about these chemicals they're spraying over the crops.
- [51:38]
D
Yeah, totally. So input costs are, you know, like chemicals are a very, very significant portion of growing food.
- [51:44]
A
Man, that's, that's. It's worrying to me that the solution to that, which is apparently you guys, is not as applied as the need for fertilizer is because there's a mismatch between solution and crisis in our food supply. Anyways, for everyone listening to this, if you're confused why we're talking about farming here on Twist, it's because I'm working this towards a point which is that Pika has put together a not just drone, but also software and the integration with the hardware to make a cool system that allows the company to really quickly build new devices. And this brings into play dropship, which as you explained to me before the show, is really the second generation of your cargo plane. So how did you get dropship up and into the air so quickly? What have you learned? And then when will it reach the market?
- [52:30]
D
Yeah, great question. So just as general background, dropship is a Dual use product. Within the commercial sector it's used for logistics. Within the defense sector it sort of forks. Yet again it is used for contested logistics and then it's also a very versatile multi mission aircraft good for carrying large sensor payloads, et cetera.
- [52:53]
A
Could it carry bombs in the little container? Sorry to be rude, but I'm curious.
- [52:57]
D
I mean technically it could. That's not the most like interesting use case for it I would say.
- [53:04]
A
Okay. And why, why not? Because when I think about drones today, we think a lot about you know, drone based warfare in Ukraine and in and around the Middle east and having a large drone that can carry more. Boom. Sounds good to me.
- [53:16]
D
Yeah. So our vehicles are designed to be quite reliable and low operating cost. And so like if you want to deliver, if you want to do more kinetic stuff, you don't build a vehicle like we did. Our vehicles last too long basically. They're too nice. If you wanted to make a really low cost bomber, there would be different decisions that you'd make than what we did with dropship.
- [53:41]
A
And probably also like you know, electronic emissions and noise and there's other factors that would have gone into making this a bomber if you wanted it to be one.
- [53:49]
D
Correct. Yeah. And so yeah, yeah.
- [53:52]
A
So you guys had a test flight recently of dropship. Really quick process to get it, you know, from as you said, CAD to the Runway talk.
- [54:00]
D
Yeah, so we went from like initial CAD renderings to first flight in 180 days which is pretty awesome. So that wasn't the fastest. I guess what we called Big Bird, the plane that we built in my parents backyard during Y Combinator technically was faster but dropship is way more complicated and useful. So yeah, it's the second generation cargo plane. First generation cargo plane was all electric cargo aircraft, really cool plane. We built eight of them. We're actually going to probably build quite a few more for commercial customers. That aircraft only has a 200 mile useful range. So that was a sort of key limitation for defense. Anyways, kind of the backstory there is we built eight of those. We sent three of them to the Air Force, got very positive feedback from them, sent one to the army, got really positive feedback as well and invitations to some big demos that are coming up in the next two and a half months. Air Force, we won a follow on contract as well. I think the thing that was so cool about this is we got really, really good insights into what it is that both Air Force SOCOM and Army want from a attritable contested logistics and multi mission uas so there's no like requirements written for this type of vehicle yet. This is too new. And so the main feedback that we got, willing to share it now because dropship is out in the market is basically like Pelican cargo, the electric thing, really awesome. Super easy to use, practical, easy to maintain, needs much longer range. So ideally 1,000 miles of range with 500 pounds of payload. Check. Dropship hits that it needs to fit in a 20 foot shipping container. So all of our vehicles actually fit in 40 foot containers already. We designed them to do that, but they're just like two and a half feet too long to go into 20. So dropship meets that requirement. The tail is removable, which is really cool.
- [55:52]
A
Ah, okay, Got it.
- [55:53]
D
Yep. And then the last really big one was. Or sorry, two more operation off of heavy fuel. This is true for any defense focused UAS. So JPA, JP5 diesel. And then the. The most critical one was the ability to airdrop payloads. So the.
- [56:10]
A
Because the first cargo plane opened up in the back like. Like a C130 versus opening up in the middle like. Sorry for the analogy. The bomb bay of a B17.
- [56:18]
D
Yeah, exactly. So the. The nose opens on both of them, but on the electric one, the whole floor was a battery. So there was no the nose that
- [56:26]
A
opened up, not the back. Oh, I totally misread that image. I'm so sorry for funny.
- [56:30]
D
No, no worries. Yeah, the nose opens on both. But yeah, we had this giant battery in the way, so we couldn't make the electric one airdrop, which is.
- [56:38]
A
I feel like we're beating around the bush. The new Dropship is hybrid, so it has both a diesel engine, as far as I understand it, for getting somewhere. And then it has batteries for very quiet looping in and around the area once it arrives.
- [56:50]
D
Yeah, yeah, exactly. So it's a really cool architecture. I don't think there's any airplane out there that's a hybrid turbo diesel architecture. So it's a parallel hybrid like you said. So the diesel engine actually has a propeller attached to it. It can also charge the batteries in flight and then the electric propulsion system. So the diesel engines behind the fuselage is a pusher propeller, kind of normal UAV configuration in that sense. And then up in front of the wings are two very high power electric motors. And so those are used during takeoff. That's how we get this really ballistic takeoff performance. Like takeoff and landing.
- [57:25]
A
And under 600ft, all three at the same time.
- [57:28]
D
Oh, yeah, absolutely.
- [57:29]
A
Oh, so it's just supposed to be going Mad like a little beep.
- [57:31]
D
Yeah, it sounds really cool. It's this like combination of turbo diesel and then the electric. So yeah, the diesel engine is about, it's like just over 30kW peak. And then the electric motors in the front are each about 25 kilowatts. So when the electric propulsion system is running, you know, our peak power is essentially three times our cruise power. Yeah, so yeah, so we use that for this like ballistic takeoff and landing performance. Use it to climb to altitude and then once that altitude actually shut down the entire electric propulsion system, we have these neat passive folding propellers and then the airplane freezes just on the diesel.
- [58:14]
A
You know, I was really excited about dropship, not because of its dual use capabilities, but more because I was thinking about domestic commercial applications, like getting stuff, places like can I send my mom a couch? You know. But it sounds like our prior, our prior point that we talked about with your agricultural drones is that regulations may not be ready yet to turn into a domestic FedEx, if you will. Are, are we going to shoot our own foot here as a nation and not end up with a lot more autonomous flight? Because it seems much more economically viable, environmentally friendly, convenient, like. It just seems better to me.
- [58:49]
D
Yeah, 100%. So I mean the, the reason that we kind of pivoted away from mass manufacturing Pelican Cargo, the electric one, wasn't because there was a lack of interest from customers. It was really exactly the reason you mentioned we could not get regulatory approval to fly scaled beyond visual line of sight operation in a timeline that was relevant to us. So that was the biggest factor. Are we shooting ourselves in the foot? Yes, absolutely. I think there's this very. Just getting back to the original point, like these hardware, software, blended products are so important to our society. They're also going to create so much value, but their value is 100% determined by their exposure to the real world. SpaceX is worth how much it is because they blew up rockets, figured out how to stop doing that, and now the idea of a self launching, recovering rocket is just taken for granted. There are very few companies who have figured out a way to actually collect that data. So Pico, we've done it. We had to go to Brazil, Zipline, they've done it. They had to go to Rwanda. Yep. And then there's a bunch of other companies that have gone to Ukraine. So Ukraine is the other place to learn essentially. But outside of that there are very, very few options, which is troubling.
- [60:12]
A
I want to make sure we get to supply Chains, components, and, and kind of sovereignty. Now if you were just making agricultural drones for Brazil, I wouldn't really care. But we are talking about the military, we are talking about duties a little bit. So how, how well are you able to source materials, components or any sort of input without reaching into Chinese supply chains? And how much better can you do in a couple of years?
- [60:36]
D
Yeah, good question. So I think kind of because of when we started the company, there wasn't like a bunch of stuff we could just buy off the shelf to build these drones out of. And so we actually vertically integrated essentially every critical component on our product. The motors, batteries, motor controllers are our own design. All the avionics is our own design, airframes, et cetera, et cetera, et cetera. So we're already able to create NDA compliant versions of our product. There is some difference in sourcing. You know, for example, the battery management system we will manufacture here in the United States for an NDA product for commercial products will manufacture the battery management system in China.
- [61:15]
A
What's the cost differential there?
- [61:17]
D
It's about 2x.
- [61:19]
A
That's. Is that a high price component compared to the overall cost of the device?
- [61:24]
D
No, not. I mean, so for dropship, no, there's two BMSS in the entire vehicle. For Pelican, it's more. There's 15 batteries that ship with each airplane.
- [61:33]
A
Yeah, yeah. So more complicated. Okay, that makes sense to me. I was just thinking that 2x didn't sound as bad as I was expecting and I wasn't quite sure why.
- [61:39]
D
Well, so if you own the design, it's not that bad. I mean, so you can, for example, if we don't source from China, we could source from another low cost region. The 2x is actually doing it in the United States though.
- [61:51]
A
Oh, no way. Oh, yeah. Huh. All right. Well, that's better than I thought. Okay.
- [61:56]
D
Yeah. I mean, I think if you own the design and manufacturing in the US versus China is order of magnitude 2x difference. If you don't own the design and you're buying from an OEM and it's like OEM in the US versus OEM in China, then it's probably going to be like 4x difference.
- [62:13]
A
Yeah, yeah, quite a lot. Okay, that makes good sense to me. But you mentioned hug. When you started the company, things weren't available on the shelf, so you managed to do it all yourself. Did that slow you down materially? And I ask that because it doesn't feel like the company is moving slowly, but you've taken on a Lot more what we might call technical risk by building solutions in house that I'm kind of shocked that you've gotten this far given that you really blank sheet of papered this entire thing.
- [62:38]
D
Yeah, yeah, for sure. So, I mean, I think we have an extremely strong technical team. It has slowed us down. There are some components that have matured really nicely in this sort of OEM space that if we knew now how far they would have come, we maybe would have paused on developing our own. So the motor controller, for example, was something we did in house. They've gotten bigger and bigger and better and now there's off the shelf options that are vaguely similar.
- [63:05]
A
Got it.
- [63:06]
D
But you know, on the flip side, it's really hard to say because it's kind of the grass is greener scenario. Like we've run into issues with our own designs and our own products. You know, lots of issues. We've resolved them all or we haven't resolved all of them.
- [63:20]
A
We're working on the last couple of resolutions. Everybody give us two, three weeks. Yeah, yeah.
- [63:26]
D
Whereas we've had issues with our off the shelf products and getting to a robust resolution with those often is far more time consuming. So yeah, it's tough to say. I think the biggest thing for us though is if you think of the most successful hardware companies out there are a blend of hardware, software, Apple, Apple, great example. Yeah. Tesla dji, perfect example.
- [63:53]
A
Sure.
- [63:53]
D
And I think like what those companies are able to do is make something incredibly complicated seem just really, really simple to the end user. You know, when you take a picture with your iPhone, it's phenomenally good. It's like better than my dad's super fancy DSLR camera now. And I don't know how or why, but like there's so much going on. There's so much.
- [64:16]
A
What are all these, what are all these little like camera things back there? I don't know what a single one of those does not. Once I looked it up. And you know what, my pictures of my kids look fantastic. So thank you.
- [64:24]
D
Totally. Yeah. And so like the way you do that is if you, you have to own everything, you have to own the hardware and the software. In my opinion, like, that's how you make that really magical experience. If you try and integrate a whole bunch of different systems, you end up with this like customer experience where you can tell. And so, so that's, I think that's like the biggest benefit. But it's.
- [64:46]
A
Are we just talking around Boeing's decision to start making things in house and just supply everything externally and integrate all the systems. I feel like we're slowly circling the Boeing story here. Michael, I. I have to let you go, but one last question before I do, which is really simple. There is a lot of money flowing around the venture capital world today. However, I talked to a lot of people that are now building enterprise agent orchestration NCP servers, and they tell me that, oh, my God, people are not interested in us. So does the company have enough access to capital today to continue growing if you still need more money to bring this vision to reality? Because I can see a really cool future in which we have tons of quiet, safe, friendly drones in our skies doing quite a lot of work for us, and I want to get there quickly.
- [65:29]
D
Yeah, yeah, Good question. I would say yes and no. I mean, we. We've been able to raise the money that we need. If we had more money, we would move faster.
- [65:39]
A
So if you're a VC listening to this. Hello. Come on. Stop backing SaaS. Companies that are slapping on an AI wrapper back something cool. We're building drones, y'. All. Michael, a treat. Working people find the company. And is there a job you are hiring for? You want to shout out into the void?
- [65:54]
D
Yeah. Great. So you can find us on our website, fly pika.com p k a.
- [66:00]
A
Not P I K A. Yeah, Correct.
- [66:02]
D
P Y K. Very active on LinkedIn. Actually, very active on Instagram. Most of it's in Portuguese, I'll warn you, but really, really cool shots from Brazil of our actual customer operations there.
- [66:13]
A
Fantastic. All right, thank you very much. We'll see you soon.
- [66:16]
D
Great. Thank you.