Pod Insights Equity Podcast Found: Taking on Nvidia and betting on transformers with Gavin Uberti from Etched About this episode We hope those of you in the U. S. had a Happy Thanksgiving! And as a holiday treat this week, Equity is bringing you an episode of our sister show, Found.
AI startups are everywhere, but there can't be any innovation without proper computing power. Found hosts Becca and Dom sat down with Gavin Uberti, co-founder and CEO of Etched, an AI chip startup focusing on developing specialized chips. Gavin shares the bold bet his startup is making on transformer models for AI chips, aiming to take on industry giant Nvidia. They discuss how Etched is developing specialized chips that they claim will be an order of magnitude faster than competitors, and Uberti shares his insights on the future of AI hardware as models continue to rapidly scale up in size and capability.
Found posts every Tuesday. Subscribe on Apple, Spotify or wherever you listen to podcasts to be alerted when new episodes drop. Check out the other TechCrunch podcast: Equity . Subscribe to Found to hear more stories from founders each week.
Credits: Equity is produced by Theresa Loconsolo with editing by Kell. Bryce Durbin is our Illustrator. We'd also like to thank the audience development team and Henry Pickavet, who manages TechCrunch audio products.
Learn more about your ad choices. Visit megaphone.fm/adchoices
Transcript 0:01 Speaker 00 AI startups are everywhere, but there can't be any innovation without proper computing power. Today's guest is working on bringing a specific AI chip to market. He and his company are making a bet that transformer models remain an important architecture for AI models for the years to come.
0:17 Speaker 01 You're listening to Found, TechCrunch's podcast that brings you the stories behind the startups. And today, we're talking to Gavin Uberti, the co-founder and CEO of Etched, a startup that is taking on Nvidia with its specialized AI chips. I'm Becca Scutac, and here to talk about the future of AI with me is my fabulous co-host.
0:33 Speaker 00 Dominic Midori Davis. And of course, before we get into our conversation with Gavin, we have our two truths and a lie. And at the end of the episode, we'll tell you which one of these statements isn't true. So listen carefully. Is the lie that Etched found out the hard way how early companies need to provide benefits? That Etched believes they're already ahead of Nvidia when it comes to specialized chips? Or that the plan was always for Gavin to be the CEO?
0:54 Speaker 01 Ooh, sounds like there's a lot of fun stuff in today's episode. So you'll have to keep listening to figure out which one of those statements is true and which one is not. But before we chat with Gavin, a quick reminder to rate and review the show. You can do so wherever you are already listening to this episode. So it will be super easy on your end and super helpful on ours. But without further ado, here's our conversation with Gavin.
1:21 Speaker 01 Hey Gavin, how's it going?
1:22 Speaker 02 Going well. How about you, Becca?
1:24 Speaker 01 I am doing well. Can't complain. Happy to have you on the show.
1:27 Speaker 02 It's a pleasure to be on here.
1:28 Speaker 01 Some of our listeners are probably familiar with you guys because you have been grabbing some headlines recently, especially related to funding news. But for those who aren't entirely sure, why don't you start by telling us a little bit about Etched?
1:38 Speaker 02 Well, Etched is a company building very specialized chips. There are many different kinds of AI models, like the filters that run Snapchat, like the recommendation algorithms that power Instagram or Facebook, or like the models that power ChatGPT, or the ones that do image and video generation. In the past, many different kinds of algorithms, they have names like CNNs, RNNs, LSTMs. How are these different kinds of models? One type of model, one application. But that's changing. Now, one kind of model called the transformer powers many, many different kinds of AIs, including GPT-4, including Claude, including the image and video generators we use today. These transformers power most generative AI.
2:20 Speaker 02 And at Etched, we are building very specialized chips. Unlike GPUs, TPUs and other AI chips, ours can't really be programmed. They can only run these transformer models. But in exchange, they're an order of magnitude faster than competing products. And they can do things like read in half a million tokens a second for Llama-70 and FP8. So very specialized, but very high performance.
2:44 Speaker 01 Yeah, and I'm definitely curious because what you mentioned with the focus on transformers, why did you guys decide to double down there and build this company up around that one area?
2:52 Speaker 02 Well, I'm a believer that you have to go take a bet. Nvidia is very good at what they do. And a lot of folks have tried to build flexible AI chips to beat Nvidia, GraphCore, Sambanova, Cerebras and so on. But it's just so hard to be more cost effective than they are that these companies have had some trouble, although they've had some success in the past couple of years. If you want to be an order of magnitude faster, though, you have to go do a different approach. You have to go throughout the specialized applications and bet hard on one single domain. And when we started this company in 2022, transformers were a very risky bet. There was no killer use case yet. There was no chat GPT. The market was very small. And even if you were right, that models would stop changing for two years, then you have to go and assume there'd be a market as well.
3:37 Speaker 02 But luckily, we got lucky. Transformers ended up being very popular. They have not changed materially in the past two years, and the applications for them have only grown over time.
3:46 Speaker 01 That's so interesting to think about the timing here because we've had a couple other AI companies on the show recently, and pretty much all of them have started before the release of ChatGPT, varying amounts of time prior. But it's interesting how much that impacted the whole space. Talk to me about that a little bit more about what it felt like when you had made this big bet to see something like ChatGPT come out and kind of just validate what you guys have been working on and kind of validate the thesis you guys came up with.
4:11 Speaker 02 Yeah, I wouldn't call it validated quite yet.
4:13 Speaker 01 Okay.
4:14 Speaker 02 But yeah, it was really exciting. When we started out, GPT three was still very unique, a very differentiated kind of product. And I could not fathom why folks weren't talking about it. But ChatGPT kind of changed that. But a great user interface, put a great kind of application onto these language models and showed folks that, yes, there is demand for half a million tokens a second for one of these big models. So I think that showed that, yeah, people want what we're buying. And from there, we were able to go raise our seed round. We're able to hire our first employees. We were able to get this thing started.
4:47 Speaker 01 Oh, so that's interesting with the timeline that you guys had the idea, but like really hadn't gotten some of that basis stuff, that first round of funding until after that. How much do you think that's impacted you guys since then? I know you guys raised a huge round of funding. I think I actually put it in my roundup of AI funding rounds that I did a couple of weeks ago. How do you think the release of ChatGPT has helped you guys progress to where you are now?
5:08 Speaker 02 I mean, I'd be foolish not to say that we are standing on the shoulders of giants. There are a lot of other really cool products that use transformers that are what we hope to sell to and not just chatbots like ChatGPT. One of the other really big transitions just this year was the fact that image and video models move to be transformers too. Early on, models like Stable Diffusion and Stable Diffusion two were powered by UNets, which is a kind of convolutional model. We can't run those on etched hardware because we're specialized for just transformers. But earlier this year, it was discovered that transformers actually work better for doing image and video generation than UNets do. So Stable Diffusion three is a transformer. Or Sora, OpenAI's video model, is a transformer. Maybe attention really is all you need.
5:53 Speaker 00 Yeah, I was actually going to ask you about what makes Etch different from a huge player like Nvidia in terms of the transformers that you're building.
6:02 Speaker 02 I mean, the fact is, Etch is making the biggest bet in AI. We are betting that these transformer models are going to stick around and be the default for the next couple of years. Every other AI chip company out there is flexible and can run transformers and CNNs and LSTMs and RNNs and so on and so forth, that many of them think we're a little bit of crazy. But if we're right, then we'll be the best performing chip on the market for these by an order of magnitude very soon. The fact is, I'm not interested in specialized chips for transformers. I'm interested in the trend more broadly as models get bigger and bigger and bigger. The way that we've gone from GPT-3 to GPT-4 and will soon go to GPT-5 is not by inventing new algorithms, it is by scaling up models.
6:44 Speaker 02 We have gone from $10 million training runs like for GPT-4, sorry, GPT-3, to 100 million for GPT-4. And GPT-5 and the other models currently training are going to cost more than a billion dollars. And the key point they're crossing is one where, starting with GPT-5, it'll be more expensive training a whole foundation model than it will be to build a single specialized chip just for that foundation model. So if you spend 100 million bucks of that billion-dollar training run on a single chip, you may be able to get an order of magnitude. And as models get bigger and bigger and bigger, specialized chips won't be an option, they'll be the only option.
7:21 Speaker 00 Yeah, and you brought up the cost of it. So it seems like the manufacturing of chips has become political in some sense, or at least there's been a lot of talk about the sustainability behind it, especially since it's using up a lot of natural resources. As someone who's in this space building, what are your thoughts in terms of how much natural resources the building of these products take and kind of the environmental impacts this innovation could have?
7:46 Speaker 02 These things do require some environmental impacts. Takes a lot of water to manufacture semiconductors in particular, but I think it's very important that we get good at making these in the US. And it's hard. Chip making is not an easy process. Chips on the front end and beyond like ours are made using these big expensive extreme EV lithography machines. And these machines cost hundreds of millions of dollars and break pretty often. And to make them make money, you have to run these things 24/7. And that means that if the machine breaks at 2 AM, you have to have some PhD worker come in and fix it at 2 AM, not 9 AM, not the following day, right away. And in the US, they have a lot of job opportunities for PhDs. That if you're one of the world experts in lithography, you don't have to wake up at
8:30 Speaker 02 2 AM and fix lithography machines. And for that reason, it's hard to go get, I think, great talent to want to go build these chips in the US. But I think that's going to be essential for keeping America ahead.
8:40 Speaker 01 I think one follow up with Dom's piece, looking a little bit further at the competition piece and some of like, why wouldn't Nvidia make something like this? Why has no one focused on this area before you guys? Why do you think that is?
8:53 Speaker 02 Well, the fact is, it didn't make any sense when we started, that it was super risky. And there was no payoff. In 2022, if you had a transformer ASIC, be able to run $20 million of GPT workloads. And that was kind of it. And that is what has changed. Now, today, it makes a lot more sense. And today, I think it's a matter of time before Nvidia and Google hyperscalers begin to compete. But the fact is, it's too late. I think we have more than an 18-month head start on them, and we're going to get to market way earlier. I think the right analogy here is Bitcoin mining ASICs. Whether or not you're a fan of cryptocurrency, the Bitcoin mining ASIC companies have been able to do quite well for themselves. They make the chips that mine Bitcoin, although they don't run them, they sell them to other parties. And back in the day, these weren't a thing yet.
9:38 Speaker 02 And Bitcoin used to be mined on GPUs. But the moment that the first ASICs came out, they were better than GPUs by an order of magnitude. And in fact, they were so much better and so much more efficient power-wise, that it became cheaper to leave your GPU turned off. Then to use it to mine Bitcoin because you'd lose money on the electricity cost.
9:56 Speaker 01 Now, that's a super interesting analysis. And with that being said, because as you mentioned on the Bitcoin side, like this has made that sort of transfer over to that different type of newer technology. Where are you guys in the process of building these chips? When do you think you'll be able to hit the market? What is your roadmap there?
10:13 Speaker 02 Until we ship our chip, from my perspective, our valuation is zero. We have to ship. And we're gonna have some exciting announcements coming out in the next couple of months. So stay tuned.
10:21 Speaker 01 So it sounds like things are closer than further away.
10:24 Speaker 02 But I think you're right on the power front as well. I don't think people talk enough about the electricity requirements that are going to be needed for running some of these huge data centers. Not just electricity, but heat. When you have a hundred megawatt data center, the hundred megawatts doesn't just go away. That is all turned into heat that has to be vented into the atmosphere. A very literal warming of the surrounding area. And AI chips are getting hungrier and hungrier for power. People are talking about bigger and bigger data centers. There are going to be one gigawatt installations built out in the next couple of years, and people aren't going to stop there. And one of the things that gets me most excited about our chips is that they are way more energy efficient on a per-token basis. And I'll tell you how we do it. Whenever you run one of these big language models, there are a bunch of weights.
11:08 Speaker 02 There are 70 billion for a model like Llama 70. And each one of these weights, for any given user answer, is used exactly once per word that comes out of the model. You have to load it from memory once and do one multiply add operation using that weight. And then you discard it. And the problem is, these 70 billion parameters of weights are too big to keep on the chip. So you have to go reload them every time for every word that comes out of the model. You can only generate one word at a time because that's what these models do. One word, iterate, put it back, run the model again. So one memory load, one multiply add operation, and you discard the memory load. But it turns out, the math operation there is way, way cheaper than the load from memory, like 100x.
11:51 Speaker 02 So what you can do is group together many different words, many different tokens, either from the same user or from different users and do a huge, huge batch. So load the weight in once and then reuse it across 10, 100, 2,048 different people. And thus amortize the one very expensive memory load from a power perspective across a huge number of very cheap math operations. And this trick allows us to get way higher throughput, both for ingestion, it allows us to go get very low latency for doing token by token serving, and it allows us to be way more power efficient. I want to be clear that GPUs do this trick too. This is not a novel idea, but we're able to take it much farther. GPU typical batch sizes are 64 to 96, but we can go into the thousands.
12:35 Speaker 01 And I'm curious, obviously you don't have the chips out in the market yet, so you can't measure exact customer demand. But what has demand been like? What feedback have you heard from potential customers as you've been working to build this up?
12:47 Speaker 02 There has been a ton of demand from potential customers. We've got a number of strategic investors for this reason, because we're only able to build so much capacity of these chips on that first production run. We have tens of millions of contracts for these products. And anyone who was not able to get a contract, some of them invested.
13:03 Speaker 01 Oh, interesting. So people who couldn't get one of the first round just became backers of the companies to share in everyone else's contract?
13:10 Speaker 02 For other reasons too. Again, I love our strategics. I think it's really a superpower to have them on our cap table. I'm very grateful for their support.
13:17 Speaker 00 Yeah, and that kind of leads us to what we were going to ask about next, which was what fundraising for this was like. I imagine, or maybe I'm just assuming because investors love AI right now, that maybe it wasn't a problem. But what was the fundraising journey like? Were investors all over you guys?
13:34 Speaker 02 They say that the number one rule of fundraising is raise when you don't need money. Raise a little bit early. For us, it was kind of a spontaneous happening. Where two of our existing investors, Brian and Mary, in a positive sum. Saw there was demand from customers to go do a strategic round. So they went and led one, added in a few great new investors. I want to give shout outs to Hummingbird, Fandomo and Oceans in particular, as well as to Sigma, a few of our strategics like Skybox. I'm very excited to have them on the cap table and have them along for the journey. But it was done pretty quickly. The whole thing happened in about a month.
14:05 Speaker 01 Oh, and thinking about where you guys were in the journey. I know you mentioned a little bit earlier, you consider your valuation to be zero just because of course you don't have the product yet out in the market. And knowing that you and your founders are pretty young to be building in this space. Did any of these things become factors when you guys were looking to raise? Or did it seem like people weren't as concerned about things like that? Or did you get any feedback on those things?
14:25 Speaker 02 Well, let's be clear that until you have the product, the valuation is zero. And investors are well aware of this. That's the risk that they're taking. They're expecting and hoping, as am I, that once these products come back, the valuation of the company will be much greater. As for age, it's largely been a non-issue. I am much younger and especially some of the other focus semiconductors. There's a great story about it. The way I met some of our TSC contacts was at a dinner McKinsey event. Where I was the only person there under thirty, the only speaker there under forty. I think that Silicon Valley investors in particular care about the idea, care about the quality of the team. And if you're smart, they take you seriously. I never really had any issues either on the management side or on the fundraising side because I'm young.
15:08 Speaker 02 And I also think I am in fact standing on the shoulders of giants. There are a few other notable Harvard dropouts and people do make some connections. And I hope that one day I am as successful and able to live up to the expectations of a Harvard dropout, just like they have. But there is a long way to go.
15:23 Speaker 00 I actually wanted to ask you about the dropping at a Harvard. Cause like I always hear the stories about founders dropping out. And I always wanted to ask when like, is it a hard decision? When you, when you say like, oh, I'm going to drop out to start my company? Or was it like a no brainer? Like, yeah, of course I'm going to drop out to start my company.
15:38 Speaker 02 Well, let's not say it's a secret. I did not drop out to start my company. I dropped out originally to go and travel the world, working for another company called OctoML or I did a micro kernel development. Other AI models like CNNs and RNNs and LSTMs rely on convolutions and major multiplications. And to go get these running really fast, you have to go assembly code. So code in the machine's language, not in a programming language, to get them to run really fast. So I did this work for a processor called Arm Cortex-MOS, while traveling the world a little bit in South America, Southeast Asia. It was very unconventional. I did not have a company at the time. I think my parents were a little bit skeptical of this. But while doing this work is when I became convinced that transformers were going to be the future, that attention really was all
16:20 Speaker 02 they needed. And also where I realized just how inefficient most flexible chips work. Most people don't know that only 3% of an Nvidia H100's transistors are used for the fuse multiply and math blocks that do math models. 3%. But if you talk to a kernel programmer, then this is all they'll talk about. Doing the math is easy when you're writing these microkernels, but getting the data to the right place at the right time, that is way more challenging. So when you're fighting against this day after day after day, it just makes sense. So I was doing this work for about six months, then got started September, October 2022 alongside Chris, an old friend from college. We began working on Edge seriously, and then we raised. It was great.
17:03 Speaker 01 More from this conversation right after a quick break.
17:09 Speaker 01 And I think sort of taking it even back a step further from that, obviously what you were working on before you started Edge was similar in some ways, obviously not 100% the same. But how did you end up getting interested in the space to begin with?
17:20 Speaker 02 I think it really was GPT-3. That was a magical moment for me. To see a machine kind of talking like a human could without any kind of crazy rules and just statistical learning. Wow, that is wild. And I just couldn't understand people weren't talking about it. That the promise of AI was so, so large at being able to do what a human does. That was just a huge amount of math. And the idea that with GPT-3, we got there by scaling, was also really interesting. That there was no trick. There was no magic secret. All it was was just making the model bigger. And that worked well. And I didn't know whether or not we'd have better algorithms. I didn't know whether or not attention really was perfect. But what I did figure was that if GPT three got so good by just scaling up,
18:05 Speaker 02 and there was a lot more than ten million dollars people were going to be willing to spend. So I figured it would keep scaling, and so here we are.
18:11 Speaker 00 Oh, wow. Yeah, chat GPT three. So now I'm really interested, like, growing up, did you always see yourself working in this space? Or was there something else that you had in mind?
18:22 Speaker 02 I'm young enough that GPT came out in twenty twenty.
18:25 Speaker 00 Yeah.
18:26 Speaker 02 Right when I was graduating high school. I'd worked for a company before then, too. I said folks had a X store. They made a on-device AI for a couple of different applications before being sold to Apple. Very interesting company. One of my hot takes is they should not have sold so early. Could have been much bigger, but.
18:41 Speaker 01 We love a hot take.
18:42 Speaker 02 It was a great team, and I think that working with them, they'd be even more convinced that this AI thing was going to be the future. People were going to figure out how to make it work. And then GPT three coming out then made me even more bullish.
18:53 Speaker 01 I'm so curious, too, because I know, obviously, people working in this space, and you weren't like the only person using chat GPT three. But I feel like so many of us, even those of us like Dom and I who cover tech, really weren't super knowledgeable about it, super in the weeds. And of course, we're not AI reporters, so of course, our AI team would probably respond differently about chat GPT really until chat GPT four came out. And that swell that followed after came with it. And it seemed like everyone became interested all at the same time. But what was it like getting in really early? And seeing the space change and knowing you wanted to work in it, seeing it change so much. But like, how do you feel about the fact that you did kind of time it just right?
19:32 Speaker 02 I don't want to give myself too much credit. I think that other people ended up seeing what we did. And a lot more people in this space. I think that this space is going to need way, way more smarts and talent than it has right now. There are a lot of really hard problems in AI today that people are working to solve, but we need more people. Think, for example, our training clusters. I talk a big game about, hey, models are going to get huge. Ten billion dollar, ten billion other training runs, hundred billion dollar clusters. But the fact is, the chips aren't reliable enough for that either. Meta published a great paper, their Lama three paper. That said that their cluster for training their Lama three crashed on average every three hours. That's really bad. And to restart the training run from the save file every three hours. If you make that cluster ten X bigger, it's going to crash ten X more often.
20:17 Speaker 02 So we need great folks to go and fix that, or power. One of the tricks that's going to be played here is very low voltages. We'll need new kinds of DC regulators to do power switching at that low voltage. Machines don't like AC at this point either. Maybe that has to change too. Or cabling. Optical cables are very promising technology, but are still very prone to breaking. It's still very, very expensive. Chips like ours or Nvidia products don't use optical cables for level one interconnects just because copper works way more reliably and is way cheaper. So when these models get way bigger, we don't have the tech yet to go and expand. So to go take two steps back, I'm very excited at the publicity space is now getting. Because that's going to drive the world's smartest people to keep working on this and solving these really hard problems.
21:00 Speaker 01 Yeah, I was gonna ask about building the actual company as opposed to the tech and knowing that once you guys like ChatGPT four came out, you guys were able to raise that seed round and kind of start hitting the ground running. What has it been like hiring for the company and sort of bringing in talent? Like you mentioned, a lot of people are flooding into the space, but are you finding you're hiring more people who are already working in AI or like? How do you hire for something that is so specific in such a space that is so hot right now?
21:25 Speaker 02 My most important job is hiring is getting the world's best talent here at this company. And I'm very privileged to be working alongside folks like Mark Ross, who was the CTO of Cypress, sold for nine billion, or folks like Ajat, who is our VP of Hardware Now and used to be a stainless engineer at Broadcom and a one of Intel's VPs who led their custom silicon group, or folks like Sapadip, who was one of the co-founders already at a company worth more than half a billion dollars today as well. Those people are the folks who make etched the company that it is. And I need to hire way more of them. We are not just building a chip. We are building a system. And that means that we need a platform. That means that we need really high quality software. That means that we need to actually sell these things and have a customer success organization.
22:08 Speaker 02 And we need similarly talented folks running all of those orgs too. So I spend most of my time, and I love spending most of my time, hiring, recruiting, getting these folks excited. It's not just the leadership, but the ICs we have at it too, I think are world class. And they do. You need to bring them along, explain the vision, explain where this company is going. Taking folks out to dinner is a great use of my time. So I cannot overemphasize. Hiring is so critical to what a company does and a founder has to spend a ton of their time getting it right.
22:38 Speaker 00 And something I love asking companies with multiple co-founders is how did you decide who was going to be CEO? Like, was there drama or did you always know like, hey, I'm going to be the CEO of the company? Like, what went down when that decision was made?
22:51 Speaker 02 All the true answer here is we started the company in September and we weren't sure yet. We never really talked about it. And then to go talk to some of the alternative chip tool providers, we need to have a legal entity. They set one up on a clerk. And I asked him, hey, I'm doing the docs. Can I be the CEO? And he was like, sure. And that was the whole, the entire conversation. And Chris has since taken on a leading role in our software team. I'm very lucky to have such a talented person running our host software division because of how critical that is and how often I think it has screwed up other chip companies. The way you win on software is not by supporting the most models, especially if you build basics. It is by supporting them really well, supporting use cases like tree search and beam search and chain of thought decoding. So I can think of no one better.
23:34 Speaker 01 So that's so interesting, thinking just based on like needing to have something and then realizing later that you guys probably did end up in the right places at least. I'm definitely curious how you think about your role as a leader and being the CEO. Obviously, you've already mentioned that hiring is a huge piece of it, but how do you think about it otherwise? You're obviously working on this sort of cutting edge tech in this hot field. How do you build culture at the company? How do you foster a place people want to come to work? Like, how do you describe how you guys all work together?
24:02 Speaker 02 I'm old fashioned. I believe in fully in-person company, people who are really committed, really excited about working late, people who want to go build the future technology. And that's not to say that this is some sweatshop. We pay people very well. We have a much larger section of the company set aside for an option pool than a typical company at our stage. Because I want to recruit the very best and the very brightest and they deserve to be compensated fairly. But still, we work hard. And I think that as a leader, you see some CEOs talking about their seven, eight jobs. I disagree with that methodology. I think I have three jobs. Hire the people. Set the long-term vision and don't run out of money. That's the end of the list. And for other things, like for, say, designing the chip. I did a lot of that in the early days, but I need to hire somebody better than me at that. And I'm lucky to have Sopedia.
24:46 Speaker 02 For things like running the tech team on a day-to-day basis, I need to hire folks better than me for that. Like Mark and like a shot. And of course, for the individual work. Well, I'm not terrible at Verilog. I am not good at all, so I need to hire folks better than me at that. And that is my job as a leader. Get the world's top talent excited about working for Edge.
25:04 Speaker 01 And it's so interesting to hear you say that because last week we had Parker Conrad on, which obviously Rippling is a very different type of a business than what you guys are running. And he had just this really interesting thing to say about leadership where he was like, I think leaders and founders and people like himself, they're like, I find the things that I'm not good at, and I force myself to spend all of my time there, which is not something we generally hear when we have founders and CEOs on the show. So I'm curious what you think about that and sort of balancing both hiring people who are better at areas that are not your strengths, but also like locking in on those areas at the same time and sort of like being able to have eyes and expertise in like every part of the business you're building.
25:41 Speaker 02 Parker Conrad is legendary. I'm a great fan if you're listening, but I think Parker's right about some jobs that the CEO cannot delegate. Some things I can, like chip architecture. While I love that personally, that's a thing that I am not the world expert on. I have Sopedia too, but things like hiring and recruiting, I have to have to become world-class in that domain. And this is not my background. I was a world champion at mathematics in high school, went to Harvard for math, and those folks are not known for being the most sociable, but I like to hope that I've gotten better at recruiting. And I think the smart folks get it. You just explain to them what you're doing, why it makes sense, on why they should be a part of it. They'll often come and join you. Now you'd have to go be respectful.
26:23 Speaker 02 You have to go and acknowledge that they're fantastic, that they could go work anyplace, but you think they should work for you. And you can tell them why. And that I have to be world class at that. That cannot be delegated.
26:35 Speaker 01 No, you definitely make a good point because of course joining a startup, people know that they're taking a bet. A lot of startups fail. Of course, great ones don't. But it's like you do have to kind of sell the company in a way that you don't if say you're hiring at Salesforce or Intel or like a quote unquote steady type job. But I wanted to ask you, did you ever think you would start a company? Like, did you ever imagine you'd be the CEO of a company you co-founded? What is this like for you personally? Like I just mentioned, a very different type of career than, you know, if you had completed Harvard, gone and worked at Meta, gone and worked at an existing large tech company like? How is this been for you personally? Did you ever picture yourself here?
27:12 Speaker 02 Well, the thing I love doing is building and shipping products. Unfortunately, it turns out that building and shipping great products is not a one person problem. As I heard a man much smarter than me once say, building computers is a social problem. So if I wanted, I had to go build unique products like our first chip, Soho, like other specialized chips down the line. It also have to get good at these jobs of a CEO. I have to get really good at recruiting, really good at setting this vision and really good at not running out of money. So I didn't see myself being a leader early on, I think. I saw myself building products, but I think that a key part of building products is recruiting a team and doing a little bit of leadership. And of course, I'm still getting better at this. Continuing to improve is my job. I think I am a good CEO of a series A company.
27:55 Speaker 02 But as this company gets bigger and bigger and bigger, I'm gonna have to get new skills and that'll take a lot of my time. But I'm excited for it. I think we're doing something pretty unique here. It makes me happy coming into work every day.
28:06 Speaker 01 Do you have any like early mishap stories that I guess now you look back on and you laugh?
28:12 Speaker 02 Oh, so many.
28:13 Speaker 01 Yes. Oh, my gosh. I love asking this question. What are some of the early mishap funny stories you have?
28:18 Speaker 02 Well, we're not gonna name the vendor at fault in this story, but we originally raised our seed round and flew out to California. My co-founder Chris had a broken leg. So I think it was April of twenty twenty-three or something. We got here to California. Chris's leg is broken, but we're a real company now. Chris can get it checked out. We have to go get health insurance. But we forget about it for a while until we are hiring our first employee in May or so. And we're like, I'm signed for some health insurance. And I think the date's right now like May fourteenth. We reached out to this provider and they say, great. You want health insurance? Sure. We can get that set for you. It'll take three weeks. And currently May fourteenth or so, so we think, ah, great, June seventh or so. And they say, no, all the insurance has to start on the first of the month. So you're gonna have to go start this on July first.
29:03 Speaker 02 And that's kind of bad because Chris has to go get his leg checked out and x-rayed. So we can make sure that he's gonna be okay. But okay, whatever, he can wait three more weeks. So we wait and we wait and we wait. And, uh, June fourteenth rolls around or two weeks away from the date. And Chris has now been like two months without seeing a doctor. And I go call up Brooklyn and I say, hey, I haven't heard anything. Where are we at here? And they say, oh, person you were talking with, she quit.
29:28 Speaker 01 Oh, my God.
29:28 Speaker 02 Who are you? I'm like, we're etched when you held the insurance. And they're like, oh, great, we can help you. It'll be three weeks. And I'm like, actually, they're like, no, because you gotta start on the first of the month. But don't worry, Gavin, September first, you co-founder Chris, you can finally go to his healthcare. After a great deal of pushing, we finally got August first as our start date. But wow, that was not ideal.
29:48 Speaker 01 Oh, my God, yeah.
29:49 Speaker 00 Is his leg okay?
29:50 Speaker 02 Oh, yes, he was okay in the end. But I felt terrible for how long I had to wait.
29:55 Speaker 00 That's crazy, American healthcare.
29:58 Speaker 01 I know, big healthcare.
29:59 Speaker 02 Oh, like, again, we had funding. We were a real company. We had employees, like we had other people with families who needed healthcare. It was a huge problem.
30:06 Speaker 01 No, that's so interesting because like every time you started a new job, it's like setting that stuff up is such a pain. And it's literally just you setting it up in like a system that already exists and already like knows you're coming and like stuff like that. Like, we haven't had anyone on the show, I don't think, talk about that before. So that is interesting to think about, like when you do start setting those things up, like when you go from that, oh, it's two of us with an idea of working on something to like needing things like health insurance. And sort of those like things that you need at the bottom of a very well running company. That's what I'm trying to think about.
30:34 Speaker 02 Yeah, and like, you know, as the company grows, you need a lot more stuff like this, too. Now we're better equipped to go handle it. And now I'm getting to be dealing with frustrating vendors. Yeah, stuff you don't think about. Startup idea for somebody.
30:45 Speaker 01 To catch people right when they spell the company and make sure they get health insurance. But yeah, I think we're pretty close to coming up on time. So I definitely wanted to ask you how you guys are thinking about the future, both for you guys personally, as well as the space that you're playing in the AI space. What are you thinking about for the next eighteen months, three years? What's next? What's the future?
31:06 Speaker 02 I think the biggest question here is how good are models gonna get? A lot of people look out at the space right now, seeing like GPT four level models and say, oh, there's a bubble. And while there might be an Nvidia bubble, there's not an AI bubble, because these models are gonna keep getting better. That this trick of scaling the damn model up has worked for ten orders of magnitude, very reliably according to these scaling laws, and seems like it will keep working. So there doesn't need to be new technology to make a GPT five level model. There just has to be more money thrown at the problem. So I think we're going to see a number of GPT five level models in the next months to years. And that is gonna be pretty transformative for the space, seeing models that can be smarter, can behave like H's, that can behave somewhat autonomously. That is going to, I think, really change many folks' minds to, yeah,
31:51 Speaker 02 this is the biggest revolution of the past hundred years. And I think as well, for us, moving on from just transformer ASICs to even more specialized chips, not for transformers broadly, but for one singular individual model, that's how we're going to go make a name for this company. So much of the stack's gonna have to be rebuilt to build this ten gigawatt data center. You need new cooling. You need new power. You need new connectors. You need new cabling. There's a lot of work to be done.
32:15 Speaker 01 Well, cool. Well, it sounds like that means we have much to look forward to. But we're right about a time. So thank you so much for coming on the show, Gavin. This has been fun.
32:22 Speaker 02 Pleasure being here.
32:28 Speaker 01 And that was our conversation with Gavin. Dom, what was the lie?
32:31 Speaker 00 The lie was that there was always a plan for Gavin to be the CEO. Apparently, it was a little bit more spontaneous than that.
32:38 Speaker 01 I'm always interested in that story because I know we've had people on before. Like, I remember we had that healthcare AI company on last summer where there was like four co-founders. And I remember the guy we talked to was like, oh, we did kind of have to go back and forth on it for kind of a while because everyone wanted to be CEO. And so it's interesting thinking about it here. But on the other side, I'm like, I guess when companies are so young, if like, you're small enough that everyone's kind of working on everything, like, it probably doesn't matter as much.
33:02 Speaker 00 Yeah. Every time I ask that question, I'm always expecting like some succession level drama of like, who's gonna lead the company? And so I'm like, waiting for a company to come and say like, actually, we had the duel for the CEO role and I won out or something like that.
33:15 Speaker 01 Right.
33:16 Speaker 00 But yeah, the company is so young. They're so young. He's so young.
33:19 Speaker 01 I know. And I think that's what's one of the funny things about the story he talked about with the health insurance and that they didn't think to get health insurance. It just like hadn't occurred to them. And I'm like, well, one, I'm like, maybe hadn't occurred to him, he'll probably be on his parents' health insurance still. But two-
33:34 Speaker 00 I was gonna say, he's probably, thanks Obama.
33:37 Speaker 01 Which I'm like, fair enough. I just thought that story was so interesting because like, one, it talks about like how fast some of these companies really are moving currently in AI. But two, I thought we've got a lot of startups that have stories like that, that we just have not heard.
33:52 Speaker 00 Yeah.
33:52 Speaker 01 Well, I guess it depends where you set up because some states, like I know in New York, if you hit a certain threshold of employees, it's New York or Massachusetts, you hit a certain threshold of employees, you like automatically legally have to have it. But I'm sure that's not the case everywhere. So yeah, definitely not the kind of thing you want to realize you should have implemented earlier because something has happened. But I mean, hey, that's probably not what they tell you about when you're starting a company is to like, think about benefits and things like that.
34:16 Speaker 00 No, I mean, remember last summer, I used to ask founders about like, how do you take care of your workers? Were the employee benefits that you give them? And for the most part, everyone, it was not like a main thought. It was like an afterthought. And so I thought about that when he was like, because I think he was also saying like, yeah, you know, we have employees and everything and they need health insurance. And I was like, so these people were just working for you and they have no health insurance? I was like, that's-
34:36 Speaker 01 Yeah.
34:37 Speaker 00 Is that just like extreme startup life? Or is that common?
34:41 Speaker 01 I know. Now I'm curious. Based on like the rest of the conversation and stuff, I'd have a hard time believing like Gavin and this company were like, the only people who like overlooked this area or like, didn't know how important it was to start this early. So I'm like, this must be a more common-
34:55 Speaker 00 No, wait, we need to talk to some of the employees because I bet there was a group chat and they were like, and we don't have health insurance or something like that. The employees were probably like, I have to go to the doctor, but I can't, there's probably like a lot more. But then again, that's probably why I wanted a corporate job. I was like, I need benefits immediately, actually.
35:13 Speaker 01 Yeah, for real.
35:14 Speaker 00 Imagine having a broken leg and then like waiting that long.
35:16 Speaker 01 No, I know. That's such an aspect of these small companies moving fast. And I definitely feel like when people talk to you about hiring the best talent, which obviously also came up in this conversation, stuff like that never comes up, which is funny because it's always about like pay and equity stakes, which of course matter, but like for someone like me, and obviously like you just mentioned it too, health insurance kind of matters. Cause like if I had to leave Yahoo today and like join my boyfriend's health insurance, I thought that we would probably even be able to do that because we're not married. But like his health insurance is awful and very expensive. It's like, that would be a huge change. Like that would matter.
35:52 Speaker 00 Yeah, families, I couldn't imagine. You know something, last night I was at this founder event for like, it was for black founders, black women founders. And they were talking about how hard it is to get like any money, like to just get anyone to give them anything. I started thinking about our conversation because he raised so much money so fast, like.
36:15 Speaker 01 I know.
36:16 Speaker 00 So fast.
36:17 Speaker 01 And it was interesting to hear about like the timeline, like he was talking about how like they had the idea. And then once ChatGPT four came out, investors were like, oh, hello. And they were like, okay, great. Like let's hire, let's raise money. Like let's actually start doing this. Like we've had founders on before who have been like, oh yeah, fundraising has been easier now that we've been able to kind of ride the AI boom, whatever. And Gavin was literally like, yep, that came out. And we were like, cool. We can actually like formally start the company now, which is so interesting to hear of just like how gun dry it was. Like they probably would not have been able to raise it if not raise money or even get started with that for the ChatGPT.
36:55 Speaker 00 I know. Like, look at Sam Altman's influence. I guess we're starting to see it. I didn't even think about it because I think you were saying how when ChatGPT came out, I also thought it was like a fad or whatever. I was like, okay, chat box lol.
37:09 Speaker 01 Yeah.
37:10 Speaker 00 But seeing all of these things come from it now, now I'm thinking like, maybe there is, I mean, like, everyone's saying it's too early to tell if there's a bubble. I was like, I don't know. Like, imagine if ChatGPT never came out, would we have etched?
37:22 Speaker 01 No, I don't know. And it's like, but would it have use cases? Because I mean, I like the way he talked about the company's like valuation. Like he was like, yeah, we've raised money, but our valuation is zero. If we don't actually produce something people want to buy and use, which is definitely a refreshing take on it because lord knows how many companies we know who have raised a lot of money and have a quite high valuation and really don't have much
37:42 Speaker 00 Yeah.
37:42 Speaker 01 going on.
37:43 Speaker 00 Yep.
37:44 Speaker 01 They're targeting something so specific in a market that not a lot of people were even talking about when they were like first coming up with this idea. So it's like saying AI is a niche market even then probably isn't fully accurate because of course companies were using it, they just weren't talking about it. But like being able to like build something like very specific within a specific area of tech. I don't know what would have happened if not for that.
38:06 Speaker 00 Yeah, and then to take on Nvidia, obviously it's possible. But I don't know, to take on Nvidia.
38:13 Speaker 01 I know.
38:13 Speaker 00 That's something.
38:14 Speaker 01 And there's definitely things where it's like you can disrupt a big company by doing something so much better than them. But I was just writing about this with software, based on the conversation we had with Parker Conrad, with like one feature, SaaS startups, etc. There's very few areas you can actually do that and be successful. Like I've talked to people about it multiple times and they're like, okay, if you want to be a one feature startup, you pretty much have to build in cybersecurity. Or no one else will like pay for you just to have this one feature. They're like, we're looking to do this like one chip, we're looking to do it well. And like, we're going to beat them because we're going to do this super well. We don't have case studies in AI to say if that's going to work or not. And he at least acknowledges how big of a bet it is because it really is. But definitely be interesting to see how it plays out.
38:57 Speaker 00 Yeah, I mean, I guess founders have to take big bets. It reminds me, there's a hip hop lyric that reminds me of this where it's like, I'm coming for you if you're coming for me, so you better be fast on your feet. That's, I mean, hey, that's, yep.
39:09 Speaker 01 Yeah. And one other thing I'll add is just I thought it was surprising when he talked about how they're like so, so, so focused on this area. And then at the end, he kind of like threw and he's like, well, we'll probably build other types of specialized chips down the line. And it's like, oh, well, are you guys bullish on transformers being the future or not?
39:23 Speaker 00 Okay, see follow up episode.
39:27 Speaker 01 Found is hosted by myself, TechCrunch senior reporter Becca Scutac, alongside senior reporter Dominic Midori Davis. Found is produced by Maggie Stamets with editing by Kel. Our illustrator is Bryce Durbin. Found's audience development and social media is managed by Morgan Little, Alyssa Stringer, and Natalie Kreisman. TechCrunch's audio products are managed by Henry Pickavet. Thanks for listening, and we'll be back next week.
Previous Aerospace Corp's CEO talks literal moonshots and Space Agenda 2025 Next 'If you build it, they won't come': Bison Ventures founders on scaling frontier tech Page updated October 12, 2026 · transcript and summary are machine-generated