AUTOMATING HEALTH CARE, THERMODYNAMIC COMPUTING, AND PRODUCT VELOCITY | E2155

22:57They're getting this incredible experience. They're going to show up to the doctor at much, much greater clips. Trey, do you eat pizza? I do. Yeah. Yeah. Do you ever seen like Domino's pizza tracker? That's like the 95 % of the inspiration. Just give people a Domino's pizza tracker. We don't need to reinvent the wheel here. You know, the, the, the problem, I guess, for us was really just, um, to, to create the Domino's pizza tracker experience. Everybody it's like, you know, Ronnie Coleman, everybody wants to be a bodybuilder. Everybody wants to live anybody heavy ass weight. So I like everybody wants a pizza tractor, everybody wants a pizza tracker, but nobody wants to build, you know, the integration into the a hundred plus effects systems, you know, the 50 different email inboxes that are going to come in.

23:34And that was really like why it was easy for us, relatively speaking, um, to get to a point that we could offer visibility to both sides. Yeah, no, you're, I do not want your job because it's a lot of work. Everyone that I've talked to who deals with anything in the insurance space deals with a huge plague of complexity as far as I can tell. So I'm glad you're doing it. So what's next? What can Tanner do next after you fix referrals? What's the next thing you can go do with a similar mindset and bent, if you will, to make this system work a little bit better and suck a little bit less? Yeah, yeah.

24:10I just, you know, I'm so allergic to kind of like the oversell or the overhyping or anything like this. But if you look at like where very medically specialized models are going and where this stuff is going to get super, super interesting, is when a doctor refers you for one service, you often actually have like a ton of sort of other issues potentially going on. Now, what prevents you from actually going and getting treated for those services is that it's a pain. You don't know how much it's going to cost. And what's cool about this patient experience is in the very sort of alpha stages, we're seeing that like if I'm getting sent over for something like a CGM, there are on average two or three other services that I am as a patient am eligible for and actually can go and be receiving.

24:52The insurance company is happy to pay for it because they know it's a good return on investment. And yet I'm not taking advantage of as a patient. So the big thing for us is actually creating like net new experiences for patients. Well, it sounded sort of Airbnb-esque, but not quite as sexy. More like net new service is for patients that are getting sent over for thing A. They are qualified and eligible for thing B and C and making it dead simple for them to go and get referred for thing B and C when relevant. I just really dispute your framing of constant glucose monitors as not sexy because if you need one, I bet you there's nothing more beautiful in the world than that exact piece of modern technology, which we didn't have not that long ago.

25:33So I like all that and I appreciate you not overselling it, but I do think that if you've done all the work to learn this element of the internals of how healthcare actually works, you've already done so much of the intellectual lift to figure it out. I hope in time, no pressure, that it can expand. All right, before I let you go, I'm curious about building in your market. One thing that I've heard from founders in venture capitalists over the years is don't sell into government. Don't sell into highly regulated industries because sales cycles are longer. There's just more stuff to take care of.

26:04It's tough, in other words. Your company, you guys had tripled between Series B and Series C, which is incredibly quick, and I'm really happy to hear that. So do you think more founders should dive into your side of the market, or do you have words of warning? You know, when it was like 2022 and nothing worked and everything was completely hellish, we had a bunch of VCs trying to convince us to go into crypto. And it's nothing against crypto, but basically at that phrase, I don't know where I learned it, but Buffett talked about how like, you know, in the 1960s, everybody was talking about the efficient market theory.

26:36And so all of his competition thought it was worthless to try investing because the markets were already efficient. And I guess the learning for myself there was, if everybody is saying that you shouldn't be doing something, it's like, okay, like that actually gets me quite excited. And conversely, if everybody's saying you should do something, you know, that's kind of lame. And maybe that's the reason, but we never said AI as a company. We don't say agents, even though like VCs will say to our face that we are the only like agentic AI thing that's actually working. And we just don't do that because it's like, I would just, I'd rather just talk about the problem, rather focus on the problem.

27:11Did we solve the problem better today or tomorrow? And actually like, you know, if there was some crazy technology that came out tomorrow that helped us solve the problem better, we would, we would do that to solve the, we would use it to solve the problem in two seconds. Well, what I like about your company is that when people think about AI in the healthcare space, They mostly think about, you know, medical imaging and finding more tumors and that sort of thing, as opposed to, you know, automating document flow. But if that's the real issue, that's broken. I mean, I'm glad we're using everything we can to smooth things out.

27:39You give me a modicum of hope, Trey, because if you guys can make this better, there must be other things that we can improve and probably quickly. You know, I mean, you guys are moving fast. The company was only founded four years ago and you're already as big as you are now. I think Fortune said you're in the eight figures of revenue. So clearly you can build a relatively sizable company and attack a big problem here quickly. So I hope other founders listen to this and ignore the part where you said hellish and only talked about the good bits and are encouraged to follow up. Well, yeah, the only, I guess the last soapbox, you know, being on the soapbox a little bit, the last soapbox for me is like, the cool thing is if the scorecard is a problem that you're trying to solve and you can solve it for one after customer after another, you almost get to a point where you talk to a customer, like, just let me solve the damn problem for you.

28:21Because I know you have it. I see that you have it, and I'm not trying to sell you software. I don't care about selling you software. I want to solve the problem for you. If you don't have the problem, I'll run away faster from you than you can chase me. That makes life easy because you're just not selling anything you wouldn't buy, and you either get really good at solving a problem or you don't exist. And that's made my life a lot less stressful because I don't have to think about everything I could possibly sell somebody. I'm really excited about it. I want you to come back on the show in a couple of months when you've had a little bit more time to spend the money to let me know how fast growth is going.

28:55And I want to learn more about what you guys are going to do in terms of how to use the money. But you just raised it. So I left you alone. Trey, where can people find the company on the Internet? And also, what is a job that you're having a hard time hiring for? Oh, great question. So, first of all, that'd be awesome. Hopefully I have something to say. God, I can't believe LinkedIn is probably the best place. You know, listen, we are always one of the reasons we're in New York is, you know, we think there's a there's a great more number of great software talent that wants to be out in New York that doesn't have great opportunities to be out in New York.

29:27That is always universally in a world where the sort of the efficient market theory of today is everybody thinking you don't need engineers. I hope desperately for any competition to believe that you don't need engineers because engineers are still, you know, that's our stable of of incredible talent here. and we keep obsessing over it. So it'll always be hard. So if you're an engineer and you want to work in the Chelsea neighborhood of Manhattan, Trey Holtman, look him up, go talk to him. In the meantime, Trey, we appreciate it. And we'll talk to you in a couple months. Talk to you guys soon.

29:56Thank you so much for having me. We're all familiar with AWS, Amazon Web Services. That's the cloud platform that powers so many of your favorite brands. But do you know about AWS Activate? That's their program for startups where they provide up to$100 ,000 in AWS credits for all startups. Whether you're backed by an investor or you're bootstrapping, money is time to keep innovating, time to delight customers, and time for you to keep gaining traction. You need runway, and AWS's Activate is going to help you with that runway. We hear this story from so many of our founding university companies.

30:33They're finding product market fit, the words getting out about their product, just a little boost, finding some savings here or there to get a little extra time can be the difference between getting traction and bringing in revenue or, hey, let's call it what it is, running out of money, okay, and shutting down. AWS knows this. That's why they've created the ultimate toolkit for early stage startups looking to boost growth. With AWS Activate, you're going to get up to$100 ,000 in AWS credits, hands-on support and training, plus exclusive discounts with some of our favorite companies and tools.

31:04So start getting the support you need at every stage of your startup journey. To learn more, visit aws.amazon.com slash startups slash credits. That's right. aws.amazon.com slash startups slash credits. The most valuable company in the world today is not a software company. Instead, NVIDIA's$4 trillion market cap puts a hardware company at the absolute top of the corporate world. Why? Well, because building and selling chips to power compute is simply an enormous business. For perspective, NVIDIA generated revenue of about$44 billion in its most recent quarter, and that was growth of 69 % on a year-over-year basis.

31:45Naturally, startups are circling. Cerebus Systems and Grok are building chips to handle inference at scale, while companies like Rebellions are reducing power demand for AI compute. The list goes on. But there's a startup called Extropic that is not simply trying to build a faster or more efficient GPU using known principles. Instead, the startup has an entirely new approach to crunching AI numbers that could consume far less power and potentially topple the most viable corporation in the world. That's why we're excited about Xtropic and put it onto our Twist 500 list. The startup is gunning for the richest vein in the modern economy with hundreds of billions of dollars worth of potential yearly revenue up for grabs.

32:22So please join me in welcoming Gil Verdon, the founder and CEO of Xtropic. Gil, how are you doing, man? I'm doing good, doing good. Thanks for having me. Happy to be here. My absolute pleasure. And if folks don't know, you're also based Beth Jesus over on Twitter. People might know your Twitter account and not have made that connection. Great to have you on the show. I want to start with what the problem is. A lot of folks out there have seen AI models improve. We keep hearing stories about bigger and bigger data centers. It seems that to some degree, the current model of GPUs and LLMs works. So Gil, what are we running up against and why does Xtropic want to fix that?

32:58Yeah, I mean, you know, it's been it's been a highly successful sort of scaling of AI in the past years. But I feel like we I think every hyperscaler is feeling the pain of trying to keep up with the demand. I think the thesis is that, you know, as you have more compute, you have more intelligence. And if you have more intelligence, it has more value in the market, you get more revenue. And so the pull of the market is nearly infinite for more intelligence and that sort of transduces to the pull for more compute and and we've clearly seen that uh and a few years ago we would warn you know i used to warn that uh intelligence will scale and we're going to run out of power and everybody would laugh uh three years ago uh why are you working on a paradigm of computing that's far more energy efficient and now no one's laughing because they have to build five uh they have orders to build uh several data centers with five nuclear reactors pretty to center and they're pulling their hair out.

33:52And so essentially, you know, what we've seen is that, you know, over the past few years, there was scaling the training worked really well. You got to GPT, one, two, three, you know, one, two, three, four, 4.5, and then they went back to 4.1. That's kind of the top end of that sort of scaling. And now what we're seeing is test time compute, they call it, or reasoning, right? The models reason for a very long time and you use the same model instead of a bigger model with more parameters, you use the same model kind of recurrently. You make it think for a while and then you optimize that sort of thinking trace with reinforcement learning often on, for example, at the moment, it's mostly on math problems because we can check whether it was right or wrong.

34:39And really that's kind of the future scaling law is at least our thesis. and I think the thesis of OpenAI and of Google is that scaling reasoning, scaling test time compute more generally is the way to go to have smarter models without having to keep scaling the training compute by 10x to get sort of diminishing returns in terms of the performance. Then again, the test time compute scaling law tells you that the more compute, both at test time and when you're training this reasoning system, you get more performance. I think it's about like quadratically more performance. It's like, I think if you put in 10x more compute, you get about at reasoning, you get about 100x, the equivalent of 100x more compute at training and the sort of old school method.

35:34If inputs go quadratically, when we compare training versus test time compute and just doing more inference work, What you're saying is your method will be more efficient inherently because you're focusing on the more efficient area of AI improvement today. Yeah. So what we're seeing is that if you make your AI algorithm more probabilistic, so instead of having just a single forward pass that's deterministic and it starts looking like a Monte Carlo algorithm, you're kind of probabilistically thinking, probabilistically searching over responses, then you get more performance. And that was the thesis of the company from the get-go, is that the algorithms will become more probabilistic.

36:11But then, and, you know, we're catering to that because we are building probabilistic computers. But also, if you look on the hardware side from the bottom up, as you try to make your transistor smaller, you try to use less current, less power, the transistors become probabilistic as well. So then we saw an opportunity. Well, hey, most workloads are going to be probabilistic. The hardware naturally wants to be probabilistic. Why don't we just cut the Gordian knot here and run probabilistic algorithms on probabilistic computers, right? And that sounds like heresy. You know, we've been doing things a certain way for many decades.

36:44You know, a lot of people rooting us on. A lot of people are skeptical. Totally natural. It's very disruptive. It's a very fundamental change in how we do computing from the ground up. But what we've demonstrated really is that we can build these new probabilistic primitives in silicon using mass scale manufacturable techniques. So using one of the large fabs. And that these are very real. They're here. They're in the lab. We submitted a bunch of papers to peer review. And now we're going to put this chip and this product. This is our test board with our first extropic silicon chip. We're going to put it in customers' hands, similarly customers, for people to try out what we call thermodynamic computing.

37:24I don't think everyone is familiar with deterministic versus probabilistic computing. So if you could just break that down for the layman, I think that'll help people understand the rest of our conversation. Absolutely. Absolutely. So computers, they're in a deterministic state. You have a bunch of transistors. There's zeros and ones. And they're definitely not in a state that you don't know because if you don't know the state of your computer, it crashes in current architectures, essentially. You get a blue screen, it's over. And then I came from a different world originally from quantum computing where there we're using quantum physics, which is like superpositions and you have any noise there, it crashes.

38:01Same thing, right? So it seemed to me like, okay, well, to scale quantum or scale deterministic computers, the bane of our existence is noise. So randomness, just from the jitter of matter, of electrons, of radiation, whatever it is, sources of noise come in, seep into your system, and add entropy. They add uncertainty as to the state of the computer. And that's usually a problem. But if you use the uncertainty and you shape it, that's actually the algorithm we're trying to run in machine learning. You're just trying to shape uncertainty into a blob that is the same shape as the data set blob.

38:35And that is machine learning. And so that's what we've been doing using essentially the natural jitter of electrons that occurs at very low current. Because if you have very few electrons, then whether a handful of them are oscillating does matter to the population average. Right. So you could think of it as a mosh bit of electrons and we're kind of steering it. Right. And that's kind of the art. That's the very real challenge there. And, you know, we're creating software and compilers to compile well-known textbook algorithms onto the physics of this device, onto the physics of this mosh pit of electrons.

39:12Okay. So determinants compute, everyone knows is what we use today. Quantum computing, everyone knows that we're working on it. All the majors are, you worked on this at Alphabet. I know Microsoft, Amazon, everyone's working on quantum, super low temperatures and an entirely different set of problems. But in between those is where you guys seem to exist. And so I think it might be time to bring up energy-based models and how those tie to the underlying structure of neural nets and other AI techniques that matter. So talk to us about that. And then I want to get from there into production. Yeah, for sure.

39:43So energy-based models are both very old and cutting edge at the same time. There's kind of some founding fathers of deep learning and AI. They're an active working group and working on energy-based models at Meta. But originally their, you know, back propagation algorithm we use to train neural networks came from Jeffrey Hinton, who was working on how to train networks of these energy-based models. And if you look at averages of energy-based models, averages of their inputs and outputs, because naturally they're more of a distribution, their probability distribution over outputs. But if you just look at the average, then that's how we actually got neural networks.

40:22And then if we looked at how to train, you know, deep networks of these energy-based models using the averages, that's how backpropagation came about. So in a way, we're going back to the ancestors of deep neural networks and running them natively in hardware, right? And that's pretty awesome from a theory standpoint. But at the same time, you know, the energy-based models have been known to be far more parameter efficient, again because you kind of have to use more test time compute to use them. Traditionally, sometimes people didn't want to throw more compute at test time to actually use them.

40:58But if you do, then you theoretically have the most parameter efficient model and hence data efficient. Because the more parameters that you have, the more you got to chisel away using gradient descent. So you need more hits of the chisel, which is data points every time, Right. So unlocking the ability to run energy based models scalably and tractably is what we're working on. And that's going to unlock these models that are more data efficient, more parameter efficient. They use far more compute. But if compute is 10 ,000 X cheaper for us in terms of. Then why do we care? Why not just use 100 ,000 X?

41:34Exactly. Exactly. Right. And so so we've been not only working on the hardware, but we've also been working on on the algorithms in order to to, you know, get people to use more energy-based models, right? So how to modify diffusion models to use more energy-based models is some research we've been doing, for example. Yeah. So I want to get to software in a second, but I want to go back to the underlying work that you're doing here to make chips more efficient because you guys talk a little bit about brains and how they are incredibly efficient, but also deal with noise in an efficient way. So does the brain provide a bit of a map for the company or is it more of an analogy for how you guys are approaching dealing with noise inside a signal inside a compute?

42:19I mean, these energy based models were kind of a physicist model of how the brain works. That's why they were then called neural nets. Right. But again, it's like a loose inspiration. But, you know, we're not obsessed with sort of biomimicry. This is not a normorphic processor. but for us you know unlike quantum computing there is a proof of existence of a large-scale thermodynamic computer out there that is very performant for ai and runs on 20 watts and we're using it to converse right now it's our brains exactly right so so that's very encouraging and again the brain to us instead of how you know we we build it with electrons sloshing around in a mosh pit the brain is really a mosh bit of of neurotransmitters right which are bigger chemicals They're heavier and slower.

43:01So to some extent, you could imagine someday we might exceed the brain in terms of energy efficiency because we're using kind of particles that are lighter and they're less energy. They're less costly energetically to slosh around. But, you know, I don't think we're at brain energy efficiency yet or in the near future. The manufacturing process is going to have to improve, but we want to end up in the 10x, 10x, you know, away from the brain. So if you were to scale this chip, this tiny, tiny chip right here to a whole wafer, it would run on 20 watts and have about 1.5 billion pbits and 17 or 20 billion parameters per layer.

43:48And then you could do multilayered programs. So you get a hundred billion parameter model. And because these models are 10 to 100 X more parameter efficient, that's effectively should be the intelligence of a trillion parameter model. And that's what we're getting for. Right. And so if you have a trillion parameter model for 20 watts, that seems pretty valuable. That's insane. You can run that on the edge. I mean, you could run that on on a smartwatch. So obviously we have to scale our system. Right. We're aiming to scale 1000 X year over year, which is, you know, really crazy. We started with three.

44:18Now we have, you know, nearly nearly a thousand degrees of freedom in the in the devices. We're going to hand customers. And then and then next year it's going to be on the order of a million. Right. And then we keep scaling. So people are definitely familiar with qubits in the quantum case. Explain for people what a P bit is and how it might relate intellectually to a qubit. Yeah. I mean, you have a classical bit. It's it's definitely zero or definitely one. And then you have a quantum bit. It's actually people used to say zero and one zero or one. it's not it's not and or or it's a it's a complex valued superposition right so they have a phase they have a complex number and and the value of that phase matters because they you know uh zero plus one and zero minus one are completely different states a p bit for example you can have it be 20 of the time sitting in the state one and it's dancing between zero and one and 80 percent uh time in the other state and then you could flip that you could go from 20 zero 80 percent one to 80 percent zero 20 percent one and that's sort of a fractional bit flip so we can do fractional operations and that's how we save power right because there's fundamental limits for any operation you do of in terms of your edit distance how much are you editing how much compute you're actually doing but because we maintain the computer in this uncertain state we can technically do fractional bit operations so this is really useful for for low precision neural networks right and and sort of any sort of probabilistic algorithms, it's not necessarily the best for, you know, super high precision scientific computing.

45:49No, no, God, no. But it's great for AI, which is inherently probabilistic versus deterministic. Exactly, and we're seeing that, you know, low bit quantization works pretty well, right, in practice. And that's what the labs do to scale their product, right? Otherwise they couldn't afford the inference. And so, yeah, that's what we're targeting, you know, from the ground up. So it's a Herculean effort to, I mean, reimagine, understand the physics of the transistor and of silicon, reimagine the whole architecture from the ground up for AI, which is needed because we're investing. I mean, we're at the point of investing trillions of dollars into building out nuclear reactors and GPU-based data centers.

46:28I think we can afford to have some startups with tens of millions of dollars taking a different bet in order to save the planet, right? I mean, yeah. So that's actually a great segue into what I want to get to next, which is just growth of the company as a commercial entity. So you guys put out this slide here, which shows essentially your first superconducting prototype back in 24. This year, the test chip that you've been holding up and showing us. Thank you for bringing that. And then next year, first production chip. So much more capacity on that thing. As far as I can tell from public information, you guys have only raised 15, 16 million dollars.

47:03This all looks to me incredibly expensive. I'm used to, as you said, you know, GPU spends being in the budgets of billions, tens of billions and so forth. So how much more capital will Extropic need to get out into the world at commercial scale? And is that capital hard to access? VCs have been known historically to be a little bit skeptical of hardware companies. Yeah, yeah, no, I totally understand that. I think VCs have been burned by, you know, hardware plays because if you're not ambitious enough, you're in the NVIDIA kill zone, right? If you're trying to make a better GPU and competing with NVIDIA, you're in for a world of pain.

47:37But if you create something that's entirely alien to them, they're going to dismiss it at first and you're going to keep scaling it. And then and then when you're at scale, it's going to be it's going to be too late. So you have far less market risk there. Yeah. In our case, I mean, you know, we've raised actually thirty eight million dollars so far. We haven't we haven't burned it all. Obviously, we're in a very good position. We've burned only about 20, 20 million so far. But we have a very, very, very high talent density team, a lot of physicists, a lot of applied mathematicians, electrical engineers.

48:10And we've been building the whole stack from from the algorithms, compilers to to to the hardware itself. And we've done several tape outs. We have done all these experiments and superconductors. We took our learnings, brought that to semiconductors. And now we're we're taping out our much bigger chip and it's we're aiming to make it available next year. But this chip turning it into a product, a dev kit, is what we're working on right now. And in the coming weeks, we're going to be able to ship it to the very first enterprise customers. So we're pretty stoked about that. Yeah, summer 2025 is the time this stuff actually reaches the customer.

48:44And that's something I'm very curious about because you can take NVIDIA GPUs and put them into many different data centers because they kind of have some compatibility. I have no idea what bringing Xtropix chips into an existing data center might look like. So maybe the right question goes, can you do that? Or will people have to build? Oh, okay. So talk to me about that. Yeah, yeah, yeah. So people saw our first experiment. So the thing with thermodynamic computers is as you make them smaller, they can run hotter. And we started with, you know, the chip was small, but the P-bits were macroscopic.

49:16You could almost see them with the naked eye. They were pretty large. But to have that operate in a thermodynamic regime, we had to super cool it. So it was in a big fridge. We put out this movie. And everybody was like, ah, that's never going to scale. It's like, we're just learning here. This is our experimental platform. And this is the 2024 superconducting prototype. You were working at quantum temperatures. Quantum and then a bit above. We're going hotter, right? So essentially we cracked how to do it at room temperature. So originally we thought we were going to have to cryo cool it with some, you know, liquid cooling or some crazy like pro gamer overclocking setup.

49:51I see you've seen my gaming rig. All right. Yeah, but we figured out how to do it room temperature in silicon, right? Using the right process. And so now, I mean, it's just like this. You know, you could have a card, you could slap it in into your server, and that's what we're going to sell in the early days. But over time, you could imagine you want to have it as integrated as possible with regular forms of computing. And so I think people, you know, what people should plan is, you know, for their build outs that there will be heterogeneous compute. You're going to have some GPUs or you're going to have some Grox, some Cerebra, some Etched or whatever.

50:27And you're going to integrate some thermodynamic computers. And then they're going to work together, right? So just like a diffusion model is like a big neural net and then it proposes a jump, a probabilistic jump for how to denoise the image. You're going to collaborate between a thermodynamic computer for that proposal and a big neural net running on a traditional processor. But that sort of collaborative algorithm, you get about 100x less compute on the classical side needed. And of course, that's most of the energy consumption. The thermodynamic computer consumes very, very little. But if you were to figure out how to run a whole algorithm on the thermodynamic computer, again, it's going to look like a different architecture.

51:03So far, what we've been able to do CIFAR, you know, kind of benchmarks that are kind of 2012. Again, we're going back to the drawing board in terms of the architectures. So we have to redo, you know, 10 years of deep learning from scratch. Hopefully the community helps us with that as we release it. But, you know, with CIFAR, we were able to get 10 to the 8 times greater energy efficiency, 10 to the 8, that's pretty significant, that's 10 million times greater energy efficiency than diffusion models running on GPUs, right? And so there's something between 100X and 10 millionX, right? Where some architectures that run really well, either mostly on a thermodynamic computer and partially on a classical one, or fully on a thermodynamic computer, sit somewhere in the middle.

51:46And anywhere in that middle is awesome. Because that means we have 100 to 1 ,000 to 10 ,000X less nuclear reactors to build to get the same intelligence and performance. and to us frankly I think I don't think robotics and embodied AI will be tractable unless you have a paradigm like this that is at scale right because right now you have a whole cloud being the brains of these robots and even with the the biggest clusters you know running these video models these video world models they're extremely expensive they're not even close to being at the edge and ideally we want them at the edge and so so you need that many orders of magnitude for us to unlock robotics and, you know, American...

52:27Are you saying that eventually robotics will not be essentially tethered to remote data centers to handle compute, but it'll be literally on device thanks to lower power demand AI chips? I just want to make sure that I understand. I mean, that's the goal, right? Like to have a fully autonomous robot that's not necessarily tied to the fleet. But I would imagine the manufacturers were always going to have some connections for all sorts of other reasons. But, you know, in practice, you could have a brain that runs on the on the robot's battery and doesn't take all the power and is very highly performant and can help do motor control and reasoning in a 3D environment.

53:03And that's the goal. It's also very important for AR and VR. Right. And AR and VR, they're kind of, you know, Apple Vision. I have an Apple Vision. It's pretty heavy. Right. Like Apple have their own silicon. They have some of the best team in the world, but it's too heavy. So it needs to miniaturize. and needs to have way more compute with less power, right? And so that's the sort of stuff we can unlock with much denser intelligence in thermodynamic silicon. So really briefly before we go, is your target customer the hyperscalers? Is it the foundation AI model companies? Is it companies that just want to run their own inference in-house?

53:38I'm just trying to figure out who you're going to sell to to start. Yeah, I mean, you could look at who buys GPUs, right? You have companies that put them in devices at the edge. You have companies that put them in their clouds. You have developers that slap them in their computer at home. And I mean, that's what we're going for, right? We're going to sell cards and dev kits. So the dev kits you're going to be able to have on your desk and you can have one or two chips. And then the cards are going to have more chips than that. And then you could build servers. And so it's a bit of everyone. But frankly, everyone needs to migrate because there's a sort of ecosystem effect when everyone's using the same software tools and the same researchers can go from a robotics lab to to an ML lab in the cloud, you know, for, for, for, for open AI or whatever.

54:23And the same researchers, their skills carry over because they are using the same chips and the same tools. So ideally we kind of spread to all these verticals, but, you know. Are you referring to kind of like a NVIDIA CUDA system, but for X-Tropic processors? Yeah, yeah, absolutely. So, so we have our own compiler and everything like that. And for some background, you know, co-founder and I, Trevor, CTO, we built, originally we built TensorFlow for quantum computers, right? How to integrate quantum computers into deep learning compute graphs and how to compile to them. And so that's what we worked on at Google as basically kids like eight years ago.

55:00And so we have experience basically creating a compiler from scratch for a new type of weird alien computer that we don't know how to program, right? That was kind of our, you know, we've done that. And now we're doing it for an entirely new form of computing that, you know, there was basically no community that was preexistent. You know, I think really there was some DARPA research on like, hey, you know, neuromorphic is kind of dead. We should think more like thermodynamic. Yeah. One of our advisors, Todd Hilton, worked 20 years at DARPA on neuromorphics. And he's kind of the godfather who coined the term thermodynamic computing.

55:34So we met up and I was like, hey, I'm leaving Google X. I'm going to do this. I'm going to build the paradigm. And he was just over the moon. And we've done it now. And we have the components. Of course, as we scale it, it's going to have more impact in the world. But right now, it's going to be the early adopters, the early believers, and also kind of the companies that can't afford to be disrupted by this next year. You could think of a hedge fund, a defense company, or even this big AI lab. Now they're in such a tight race. It's almost like hedge fund. They're looking for alpha. They're trading researchers for$100 million a pop.

56:08It's getting crazy out there. So any sort of edge they can get, they'll be interested, right? So that's where we can end. I think if your guys' technology risk goes all the way to zero, your market cap's going to scale to essentially infinity. And that's why I was really excited to have you on, because I really do think that we're going to have to really go in a different direction to get to what we need to go. Gil, an absolute pleasure. I could keep you here for two hours. You have to let it go here. When you do get the dev kits out, let us know. And oh, what is this gonna cost? Oh, that's a good question.

56:36It's gonna be a case by case basis. There's not that many. So for me,$5, that's what I heard. Something like that. Well, we'll see. All right, well, thank you so much for having me on. I really appreciate it. Cheers. Okay, everybody, your pitch is the face of your company, right? What does everybody say when you ask them for money or they're going to join your company, they say, send me your deck. And some founders, no offense, you're walking around with a little schmutz. You got the schmutz on your face, right? It's messy. And people judge a book by its cover. So you need a clean, easy to follow, beautiful pitch deck.

57:10And not just to let investors know about your company, it's also to show them you're capable of making a beautiful deck, right? It's just kind of table stakes. Let's call it what it is. That's why we've joined forces with our friends at Gamma, G-A-M-M-A, for a twist pitch deck competition. Why am I doing this? I want to see better decks. So go to gamma.app, G-A-M-M-A.app. These decks come out so beautiful because you use their AI power platform to create stunning, stunning two to three minute pitch presentations. And you can import your existing deck from whatever incompetent, ugly, disgusting platform you built your deck on and you can get started from scratch with just simple text prompts, right?

57:54AI first. It makes sense. And Gamma is working with 20 of the top AI models. That means you're going to be able to use one of their hundred pre-made themes and templates to help you elevate your ideas and your script and your designs. So here's what I want you to do. I want you to submit your finished pitch deck at gamatwist.com. Gamatwist.com. Go to gamatwist.com. We set up that special domain, and you submit it there. 10 finalists will pitch me right here on This Week in Startups, and I'm going to give the winning company$25 ,000 in cash, or I'll even make it an investment if you want my name on your cap table.

58:35You choose. If you take the$25 ,000 as a prize, it may have a certain tax treatment. If you take it as an investment, it may have a different one. So talk to your accountants. Make your deck today at Gamma.app, and then go to GammaTwist.com to submit it. And hey, get the schmutz off your face, all right? Clean it up a little bit. All right. So last week on Twist, we dug into Substack's massive$100 million Series C. It was a big event for the newsletter and online creator space. The investment pushed the value of Substack up north of a billion dollars and was, I believe, the largest round in its space since Beehive raised$33 million last year.

59:11Our question after chewing on Substack's latest fundraise was pretty simple. What's the 10x case for both Substack and Beehive over the next 10 years? Venture investors backing the two companies really do expect material upside, but can that be achieved when media economics are in tatters? Now, we've had the founders of both companies on the show before, but with new capital flowing in and new revenue milestones being announced, talk about that in just a second, it's time to look to the future. So please welcome back to the show. it's Beehive's co-founder and CEO, Tyler Dank. Tyler, how you doing, man?

59:43It's great to be back, Alex. Thanks for having me. It's been almost three whole months, so it's about time we got back together. And since we last spoke, you dropped some news that in early July, I believe, Beehive scaled to$20 million in ARR, but that's not the full revenue story. Tell me about where Beehive is today. Yeah, so$20 million ARR as of a few weeks ago. And then we also have an ad network and a tertiary monetary revenue stream called boost, which is like a co-registration network. And the combination of the ad network and boost accounts for another 10 million in revenue. So altogether about 30 million revenue run rate as of today.

1:00:20But just to be clear, this gorgeous chart that you shared, I believe it was in your newsletter, big desk energy. This chart is not inclusive of boosts and the ad network. This is just the software revenue here. Yeah. If you, if you multiply that by 12, that would get you the 20 million ARR. And yes, that's just the SaaS revenue. I wanted to put this chart up because it's very, very, yeah, I was going to work of art, beautiful, impressive. It's pretty much rock solid. And I want to talk a little bit about how this chart came to be, because one thing I have seen is that Beehive has incredibly quick product velocity.

1:00:53You guys are always launching new stuff. So can you just, for folks out there who are building, try to connect for me product velocity and shipping things and MRR growth or ARR growth? Yeah, so the product velocity, I think is kind of like, it sounds cheesy to say built into our DNA, but we entered a very competitive industry back in 2021 where there was MailChimp who just been acquired for 12 billion. There's tons of email 1.0 platforms. Substack has been around for four or five years. There's other platforms in the space. And we entered with an MVP prototype type product where we knew that it was very rudimentary relative to the competition.

1:01:27And so the only way we can possibly catch up or the philosophy we took is let's just ship as many valuable features as quickly as possible. And what that does is one catches up to the market that we entered very late, relatively speaking. But it also creates this narrative that we can ship quality product quickly that addresses our users needs. And so that gives users or potential users the perception that, yes, they don't have what I want today. But because they ship so quickly, maybe they will solve my use case next month or in a few months. And that buys you a lot of patience with these early adopters of the platform.

1:02:00And we really just haven't taken our foot off the gas since. So I was thinking about Beehive and Substack and people like to talk about Uber and Lyft and kind of other traditional well-known technology rivalries. But I think the best analogy or perhaps analog for what Beehive is building is Ramp. Because when Ramp came out, they were targeting a similar-ish market to what Brex was doing. They had a particular angle on it. They were focused on saving you money versus kind of corporate cards for startups. and since then they've managed to grow from what was immediately or initially dismissed as an also ran into being one of the most valuable companies in in silicon valley so i guess the beehive story saga if you will is that you can start with something basic and then iterate incredibly quickly and catch up to any incumbent or just one that was only a couple years in the market like sub stack i mean sub stack is a formidable competitor for sure i think the market's massive right and like we said like i just said mailchimp got acquired for 12 billion there's dozens and dozens of other email platforms.

1:02:56As of two weeks ago, we just launched a website builder as well. So you could argue that we are going deeper into the creator or just builder space, competing with a WordPress web flow framer as well. And so I guess we're going to make the pitch of where we can 10x from here. I think there's a lot of opportunity, both in email, but also in the website space, which is growing very competitive. But I think the wedge of being vertically integrated with a website builder that has a full CMS and newsletter and the growth and monetization of your audience is a really compelling offer that other website builders don't offer.

1:03:30Also, I'd say your analogy of Brex versus Ramp is interesting. I think the way that I typically look at it is I think us versus Substack is pretty analogous to Amazon versus Shopify, where Amazon has millions of third-party sellers, but to access their products, you have to go through the Amazon website or the Amazon app. the product shows up in an Amazon box and Amazon owns the customer relationship. Very similar to how Substack owns the distribution and reader relationship with its users. Whereas Shopify is more similar to Beehive in the sense of it's the infrastructure and tools that kind of sit behind the scenes to empower the success of our users.

1:04:09And we don't take a cut of revenue, unlike Substack. So I think the Shopify and Amazon analogy is pretty interesting. So just pulling up the numbers here, Shopify currently valued at about$161 billion, Amazon at$2.41 trillion. Stretching out the time horizons here, do you think that in time Shopify's model will generate a larger company than Amazon just based on market value? I'm trying to understand why you're taking this side of the bet when Amazon has done so well with its centralized model and frankly, owning the customer relationship, as you said. Yeah, but you could also probably make the, and granted, I'm not a financial analyst viewing both of these companies.

1:04:43You're not? That's why I had you on, Tyler. But you could also argue that Amazon benefits from AWS and they own Twitch. Not that that's a massive revenue stream, but they have a diversified product suite outside of the pure e-commerce play where Shopify has really owned, like if you want to build your own brand and your own product, they are the storefront. They're the infrastructure. They will facilitate payments. So I think that is kind of what we are trying to do. We don't want to be, we don't say sign up for my beehive, right? Doesn't sound great. And like, that's not the purpose of what we're building.

1:05:13We want you to create your blog, your website, your newsletter, your storefront, whatever that looks like. We want to sit in the background and empower our users. I think Substack has very intentionally created the alternative or taking the alternative approach of sign up for my Substack, download the Substack app. If you want content created by our users on our platform, you get that by Substack.com and find different content that suits you. So I think it's like a philosophical difference in how we're approaching the business. So as I was thinking about the future of Beehive, actually, one of the questions I wrote down was central application, question mark.

1:05:47And it sounds like you're saying that that's not something that Beehive would want to build because you're not trying to aggregate demand. But notably, in the case of Shopify, they have put out an app, just sticking to your analogy there, that does centralize shopping demand for sellers on the Shopify network. So does the analogy to Shopify fall apart when it comes to centralized mobile applications in the case of Beehive or am I just taking that analogy too far? I mean, it's a great point, right? I think what we've seen time and time again is when platforms get in between their users and their audiences, it usually doesn't end up so great for the users.

1:06:23And so in 2018, when Facebook revamped their overhauled their newsfeed, much to the detriment of publishers and brands on the platform in 2022, Twitter deprioritized external links and the newsfeed, which hurt traffic to publishers. And then just recently, as of a few weeks ago, you've seen publishers claiming 40 to 50 % decrease in traffic due to Google's AI summaries. And so I think that there's something around aggregators and trusting a large platform that has historically always come to the detriment of users and publishers. And I think one of the winning narratives for Substack is distribution and algorithms that can boost quality content.

1:07:02and just know when that distribution works in your favor at times, it also can work against you. Just like Facebook used to give tons of free traffic as they Googled all of these publishers and users until it didn't. And so I think actually it's like a cheap dopamine hit to get that type of traction early. But I think the more sustainable businesses have their own foundation and can stand alone on their own. And that's what we're trying to empower our users to do. Yeah, I mean, in one sense, I think that the dopamine hit, the short-term sugar rush is the right analogy because it never feels bad back in the days of Google being a great traffic source to have Google send you 5 million people.

1:07:40But if it's going to send you zero the next day, it's just simply not sustainable as a model. So that brings us all the way back to kind of the core Beehive product, which is email newsletters. You and I both write one on a regular basis. So we're in the trenches here. I also write the Twist 500 newsletter, which is on Beehive. And I'm curious about how much that market can grow because I think you and Substack have done a good job of getting all the folks like myself and yourself who want to write for a regular audience on our own terms-ish onto these platforms. But what I'm not clear about, Tyler, is just how many more folks out there in the market who might want to sign up for Beehive, launch a newsletter, either use the ad network or sell subscriptions, and just kind of join the party, if you will.

1:08:23Are we halfway there? Are we a third of the way through the market? How much is left? Yeah, it's interesting because when I raised our seed around three and a half years ago, that was the number one point of feedback that investors would give me. They're like, how big is the market? Because everyone kind of looks at, oh, newsletters are at the time, Morning Brew, The Hustle, Axios, The Skim. And they're like, how many more, how can you build a business? Like how many more of those exist? And I think the narrative you just gave is like, yes, there are writers, there are people who want to build their, who write content and build their audience.

1:08:50But how many of those like journalists and like written word first creators exist? And I think the one, the answer is a lot more than people give it credit, which I think we've proven over the past three and a half years. Two, I think you're seeing that content creators that are YouTube native, Instagram, TikTok, or just businesses are seeing the value in communicating with their audience directly via email, whether that is a customer comms email that you send weekly, whether that is an investor update. I think we've all come around to like email is not going anywhere. It's actually, it has a ton of staying power.

1:09:22And so I think that there's more opportunities to bring these types of net new creators on board. But I also think there's a lot of markets that are not email first yet, but can diversify and leverage email as an additional channel to engage with their audience. And I still think we're pretty early on that. Email is the form, but the content can be variable. You and I write a lot. You also do a lot of graphics and illustrations and your newsletter is in the, what is like Windows 3.1 historical style or something? Yeah. Windows 98. Oh, that's, well, I just dated myself there a little bit. But one thing I'm curious about is emails and newsletters as essentially a way to send stuff that isn't words, videos or podcast transcripts or audio or whatever.

1:10:05So how does Beehive think about expanding support for other types of content inside of email to hopefully get those TikTok first creators possibly onto your kind of economic engine, as you call it? Yeah. And so what you're pushing up against is like the limitations of email where like native video is just not supported, which is one of the downsides of email. right? It's antiquated and it has staying power, but it isn't as dynamic as a webpage. What I think we are hinting at is where I see a lot of opportunity in the business. And so, because I know that the preface for the conversation is like, how does Beehive 10, 100x from here?

1:10:36And so I'll paint the picture of what I think what we're building is so interesting and like the three pillars that overlapped like a multi-billion dollar opportunity. One is email, two is website, and three is monetization. And I don't think there's a platform that exists that does all three of those as well as what we're trying to build in a way that's a truly one-of-one opportunity. So focusing on email, we own the newsletter space now. So newsletters are bread and butter. I would argue that email and content-based email is actually the smallest segment of emails, e-commerce and marketing emails that you get for 20 % off Black Friday.

1:11:11Like there's a reason why MailChimp sold for 12 billion and why Klaviyo is a public company. So I think not saying we are definitely going into marketing emails, but that is an expansion opportunity. If we are great at email, what would be stopping us from expanding there? And another type of emails like transactional email. So whether it's SendGrid, Mailgun, basically, if you are creating a mobile app or a website that needs to fire a forgot password or log in, like transactional emails is a whole different segment of emails that we could potentially explore. Not tipping our hat that we're going into either one of those, but I think there is a wide range of email type companies beyond just content, which we are suited to be able to expand as opportunities present themselves.

1:11:51So let me, let me explain the two other pillars. So the website builder, we acquired a website builder company a year ago called Type Dream. We've been building in beta for the past six months. We launched it two weeks ago. The website as a key differentiator that's vertically integrated with the email is like a great selling point, right? So you can have a website on WordPress, but if you have an email newsletter, you're copying and pasting into MailChimp or to Beehive, two different systems, two different logins, two different pieces of infrastructure that you have to manage. We've just gone vertical with the website builder and where website leads to is kind of like the classic creator play of link in bio, community courses, digital products.

1:12:29And so I think there's a whole roadmap that can be built and will be built on the website side that allows us to onboard coaches who want to sell their time and they have a storefront of booking time with them directly. And rather than paying a 10 % fee on a lot of these other platforms that do booking, you can have all of your booking is done on our website product that is fully integrated with the email product so you can communicate with your users. So I think website as something that we made a big bet on is bearing fruit already. And I think there's a huge opportunity and roadmap to be had on the website front.

1:13:01And then the last pillar is monetization, right? And so we offer paid subscriptions just like Substack, Patreon and other platforms. We're the only platform in the space that doesn't take a cut of revenue. So you can charge your readers$10,$20,$30 a month, and you keep the 10 minus the Stripe fees, you keep what you charge. We have an ad network, which has brands like Nike, Netflix, HubSpot, AG1. And so as a publisher and focusing on the ICP of publishers, you receive ad opportunities every week. You see that AG1 wants to sponsor your newsletter for$5 CPM. You don't have to run it. If you want to, it's a high quality ad.

1:13:35And you've never spoken to anyone at AG1, yet they're paying you$150,$250, up to a few thousand dollars for running the ad in your newsletter. And I think we're very early on the ad network front. And in a way you could argue we could maintain our user base we have now, scale the ad network, and that is zero marginal cost and additional revenue as long as we're filling ad rates in these different newsletters. So I guess going back to your initial question, like why not jump into marketing yet? I think we have owned this ICP of like content first email newsletters. And if we can help them grow faster, build a website, build products on top of their website and monetize.

1:14:11I think that's a really compelling offer. And I think there's a lot of room to run with that. So, you know, we used to think about companies get into 100 million in revenue and then going public. I know that's a 1997 perspective on things, but I still think about 100 million in revenue is like here is a company that will last for a very long time. It has reached essentially unkillability in some way. It sounds like, Tyler, that with the addition of websites, which I believe will unlock incremental software or ARR for you, software revenues, and growth on the advertising side, you have enough components just there to grow the business from a$30 million run rate today to$100 million in some amount of time.

1:14:50That answers the question of how do you 10x? Because if you can get to that scale, your last valuation is going to appear tiny in comparison to what you're doing now. So maybe the right question then is to ask what could go wrong? What could disrupt you? Because I've seen the revenue charts. I use the product. I've seen the releases. Things seem to be going pretty well. What's the potential turbulence on your horizons? Yeah, I mean, it's worth calling out for all the revenue projections, too. We just launched three and a half years ago. So I think we are just getting warmed up and really just beginning to hit stride.

1:15:20Up until two months ago, we had a team that was manually provisioning ads in the ad network to our publishers. Now we have a machine learning model that's using data and optimizing both campaigns for advertisers and for publishers. So it's very nascent. It's successful, but very nascent. and the website builder has been live for two weeks. So I'd say like, these are big bets that we've been making and investing in for years, but we're three and a half years old. And I think like just starting to see a lot of these things go live. What could go wrong is I just pitched you eight different business models and eight different ICPs of we could do transactional, we could do marketing, we could diversify here.

1:15:55And I think it is very hard what I've learned to build a single startup that does one single thing very well that customers will come back to and pay. We are trying to build almost like three companies in one with the website builder, with the email newsletter and core email platform, and then an ad network that is globally scaled with thousands of advertisers across all of these different publishers. And each one of those is almost like a startup in itself and very difficult. So can we thread the needle and do all three equally well and have them compound where one plus one plus one equals 10 is like the bet that we're making?

1:16:31but each of those has tons of complications and i could see ourselves and if there was an area of concern it's that they all need to work well together in lockstep and it's very hard to do all three at once but it does seem like you have a pretty clear conception of your icp or ideal customer profile if people don't know we're not talking about the insane clown posse and if you stick to knowing what your icp is it seems like you're going to have the right direction to follow to make sure that you do thread the three parts together into a kind of one cohesive whole maybe I'm underselling how hard this is, Tyler, but to me, it sounds like what you're describing is doable with your team.

1:17:05Just given what I've seen you put together before, you don't need to be modest. We're on a podcast. You can, you know, you can beat your chest a little bit. Okay. Yeah. I mean, again, it's because we've been here for three and a half years. I remember like yesterday, the days that we had$10 ,000 of revenue and everything feels so fragile, right? So we're in a hyper competitive space. A competitor just raised a hundred million dollars. We liked the valuation though. That's good. Um, and, and it's great. It's great for you guys to have that comp and they're jumping into advertising which is something that you guys really kind of showed a new way forward for for folks who don't know you usually don't get a five dollar cpm from an advertiser you're often selling banner ad space at a very low effective cost per thousand or cpm rate um which is why i started off this little chat with the media economies and tatters um because the old models of advertising don't work and yours seems to work at least for newsletters So I don't know, I'm pretty bullish, Tyler.

1:17:57But just to kind of close things up for today, Tyler, why is Beehive and Substack doing so well when traditional media is in such an absolute dire situation? Because to me, the split screen between Beehive's growing, newsletters are doing so well, and also, you know, my old home of TechCrunch has cut a huge chunk of its staff. It's now down to the absolute bones. So what's driving that? And is that a long-term opportunity for Beehive? Yeah, I think the underlying thing there is the owning your audience and distribution. So grouping media as a whole as struggling, I think, is partially incorrect.

1:18:31Morning Brew had an amazing exit and post exit continued to crush it. Axios had a great exit. Politico had a great exit. The thing all three of those have in common is they were newsletter first and they own their audience distribution, where if you flip it to the other side of the equation, you see a lot of publishers who outsource traffic to Facebook, Twitter, Google. they monetize with a few pennies on a page view on banner ads and google adsense and when twitter facebook stops prioritizing external links and google now is leaning into ai summaries your traffic dries up as does your revenue so the underlying business model i think struggled you're seeing new business models like puck and other publishers kind of lean into subscription first and they're having a lot of success as well so i think it was again going back to like that easy dopamine high that you can get from like network effects and distribution.

1:19:21I think it feels good, but those who can own their own distribution, build their own moat, have their own brand and product and have distribution directly with their audience. I think those are like the tried and true media companies and individual content creators that stand to benefit from this. All right. So the 10X wager from Beehive is email's not going away. Building websites and vertical integration inside your service is going to have a big effect on helping people do more. The advertising system can scale well. You're going to have lots of software revenues. And as media falls apart, well, the people are going to be building and wanting to own their own audience.

1:19:57I can see that being a 10x wager. So Tyler, when you do raise more money, you call me and I have five, five dollars ready to go into the next beehive round. So watch out world. We'd love to have you. I'm going to tell you, thank you so much as always for your time. I love hearing you explain this stuff. And when you do hit the next ARR milestone, call me. Yeah, or subscribe to my newsletter. It's the first place where you'll find any of the milestones that we had. That's actually true. I had to go back through your newsletters to pull stats for today's chat. For folks who want to find Big Desk Energy, where can they?

1:20:25Just mail.bigdeskenergy.com. It's not a podcast appearance if I don't promote the newsletter. I worked it in the first two minutes of our chat. I know. I respect the game. I appreciate it, of course. All right, Tyler, thanks a lot, man. We'll talk to you soon. Cool, thank you.