23andMe BANKRUPT, Saylor Still Buying Bitcoin, + Founder Interviews | E2102

24:24Right. We're taking a bet. At Etch, our chip is not programmable. It does one thing and one thing only, and that is run transformer inference. But in exchange, it gets much higher throughput and much better time-difference token latency. And for folks out there who don't know what that means, essentially, if you're doing inference, it allows you to have the results come back more quickly, as I understand it. You got it. It means that both you can serve way more users with the same amount of servers, or you can go and give your results back each user way faster. An order of magnitude more than the NVIDIA Blackwell GPUs.

24:59Those Blackwell GPUs have been really not basically in the last quarter or so. So you're saying that what Etch is building could be an order of magnitude better than not old NVIDIA tech, but what's state of the art right now that people are paying, I mean, honestly, hundreds of thousands of dollars per unit for? Oh, yeah. Again, it's a trade-off. NVIDIA is a great company and their chip does so many things better than ours do. We can never run ResNet. We can never run physics simulations, what we are good at is running transformer inference. And there we will outperform by 10x. And this is why I'm calling the company a bet, because if transformers remain the bedrock of most AI models out there, you guys are building a technology that I could see having almost infinite demand.

25:41But if the world of AI moves away from transformers, you guys could be sitting ducks. So I'm curious about the level of confidence you have in transformers retaining their primacy in the AI game today compared to when you started the company because people are working and there's always new things coming up. So I'm just curious how much you guys are like, yeah, Transformers, the correct place to bet the company. Well, when we first started the company, it was a much riskier proposition. We began this company midway through 2022. Back when there were some Transformers like GPT-3, but when there was no big demand yet, GPT-3 hadn't found a use case.

26:17there was no chat GPT. So as a result, you had all the risk, but none of the benefit. So you guys were jumping off, hoping to find a parachute on the way down, and then chat GPT comes out and voila, you guys must have felt like absolute geniuses when a transformer-based consumer AI product went from zero to literally the talk of the entire world overnight. Well, it was certainly good for us, don't get me wrong, and helped us raise a round of seed funding. But there was still a lot of risk. Transformers were one key technology, See, other things ran differently. Stable Diffusion, for example, was a unit.

26:48Maybe that was the future. Stable Diffusion 1, 1.5, and 2 were all units. Dali was a unit, so we weren't clear what would happen. But over the course of 2024, I think we really saw the bet crystallize. Stable Diffusion 1, 1.5, and 2 were all units. But on Stable Diffusion 3, it turned out the Transformers worked better, and they've been Transformers ever since. Or video generation. I'm sure that you saw Sora. Of course. The breakthrough that Sora had was using a transformer for video generation. That was the secret sauce. They talk about it in the Sora blog post. I missed that. Or I'm sure that you saw Oasis transformer as well.

27:23Yes. Okay. So it sounds like you guys made a bet that transformers were going to be the future and we'll build chips for it. The market then drops ChadGPT, accelerates things, but still not entirely clear where things are going to go. But by 2024, last year, you're pretty confident. And I presume that that confidence was shared by investors because you guys raised an enormous$120 million round. I think it was June of last year. It was a good round. I prefer to go talk more about customers. Okay. Can you just break down for people how hard it was to take what I consider to be software and kind of burn that into actual chips themselves?

27:57Because that translation is the part that I know the least about and would love to learn the most about. This was hard. And it's hard for a couple of reasons. The biggest one by far is utilization. On a chip like an H100, there are, you know, about a petaflop of FP16 flops on each one of those dyes. The H100 is the preceding state-of-the-art generation of NVIDIA GPUs. And can you explain to me that second half of your point? Because folks out there are not as deep in this as you are. Ah, thank you for chiming in there. When we run an AI model, it's made of math operations. And one math operation is way more common than all others called the multiply-accumulate.

28:33It's very simple. You have three numbers, A, B, and C, and you do A times B plus C. Okay. And one mechanism you can use to model how good or bad an AI chip is, is to ask, how many of those math operations, A times B or plus C, can we do per second? And there are a lot on modern AI chips. NVIDIA's Hopper class of chips, the H100, can do almost one quadrillion multiply or add operations per second for FP16. One quadrillion, which is a thousand trillion. It is a thousand trillion. Okay. That's the preceding generation. How much faster is Blackwell? Well, the thing about Blackwell is it's actually not much faster.

29:17The highest spec Blackwell chips, the B200s that are water cooled, on each die, they have 25 % more flops. So instead of being 1 ,000 trillion, it is 1.25 thousand trillion. And you guys have written about how we're seeing Moore's law. This is where Moore's law is crunching, that you're not able to fit doubling amounts of transformers on these dies. So the dies are getting larger now for these chips. Right. In fact, MorseLot doesn't even apply for the move from the previous generation of an NVIDIA product to the current generation. Usually, as time goes on, you make your chips on more and more advanced process nodes.

29:52A process node is a sort of assembly line for making semiconductors. You hear about 22 nanometer, 14 nanometer, 7 nanometer, 5 nanometer, and so on. 2 nanometers, the current state of the art as far as I understand it. This is a bit of a misconception that when they went and made the H100 dyes, the previous generation of NVIDIA chips, they built them on the essentially five to four nanometer process node. Okay. But when they built Blackwell for the first time in many years, they did not change the process node. The next gen, three nanometers, was not ready. Oh my. But that does imply, though, that there's a next generation coming that will be, again, programmably quite powerful.

30:33This advertisement is paid by Fidelity Private Shares. All right, founders, we all know cap tables, due diligence, and of course, managing investors is a huge headache, but there's a very simple solution for you. Today, we're talking with Kristen Kraft, an old friend of mine, and she works at Fidelity Private Shares, a new group over at Fidelity. You've heard of Fidelity before, and they have a mission to help startups simplify equity management. They're going to save you money. They're going to give you better service. Welcome to the program. Kristen. Thank you so much, Jason. It's great to see you again.

31:05Yeah, great to see you as well. Maybe just from a product perspective, what are you trying to accomplish with the product? So Jason, we are super excited about our CapTable Management and Data Room platform. We want to make it super simple for founders and startup operators to manage all sort of ownership and equity in the company and essentially prepare to raise. We want to make sure that everybody goes into these fundraising conversations well-prepared, they're ready to share their cap table, and that they're ready to go through due diligence as they're trying to close their round. So from a product perspective, that is where we're laser focused.

31:42And that product is really well built, really strong attention to detail in the way that Fidelity is known and beloved for. So if you want an all-in-one equity management platform, Fidelity Private Shares, they've got you covered. Visit fidelityprivateshares.com. That's one word, no spaces, no dashes. Fidelityprivateshares.com. And hey, mention This Week in Startups. They'll give you 20 % off your first year subscription. Once again, fidelityprivateshares.com and tell them that you heard about it here on This Week in Startups. Don't get me wrong. NVIDIA is a great company and they will keep innovating, but the pace of improvements on the semiconductor side is slowing down.

32:20No longer can you keep getting the performance better for free every two years. It will keep getting better, but at a slower rate. So essentially, if NVIDIA is not able to squeeze more of this math per chip as time goes along and demand for LLMs, which are transformer based, remains high and growing, that does perhaps open up a spot in the market for a transformer specific ASIC. You got it. Okay. So essentially, my big contention that this company is a bet or a wager that transformers will remain, you know, important is almost been answered by the market with a yes. And with the information from NVIDIA, there is space in the market for it.

32:58So I guess actually maybe this is the right time to say, okay, your system is called Sohu and you burn transistors into it. So please get back to the translation from software to hardware part of this. Well, on a NVIDIA H100 or B200i, you have some headline number of the amount of math operations per second it can do. Say, a thousand trillion multiplies or adds per second. So you might expect you could actually get that, but this is not the case. Usually, the number of operations you'll actually get is only about a third of what they claim in the headline number. For example, when Meta trained Llama 3, they got between 38 % and 42 % of the headline number of flops.

33:37And they hailed that. It was one of the key points in their paper as a really great accomplishment. They worked hard and they got less than half usage of the flops that are claimed on the side. If 38 % is good, Gavin, what's an industry normal bad result? You can easily see 20%. Whoa, that's one fifth. It's a massive, massive thing. Even with this, NVIDIA is able to get that 20%, that 20, 30, 38 % because they spend a massive amount of software resources making their stack quite good. For other competitors, Kuda. Kuda and PyTorch and the whole ecosystem around it. But for other products, you'll see even less.

34:14But I presume that we're working our way towards that Etched has a much higher yield on flops. Well, that's the key thing. With an NVIDIA product, they have to build it to run any kind of program. And as a result, it's much harder to model, hey, how do you pack data in for a transformer? How do you get that really high utilization number? But on an ASIC, for example, on a Bitcoin mining ASIC, you will get 99, 100 % utilization of the Bitcoin mining hashes per second that they claim on the side because that's all that it does. We should define ASIC here. I'm so sorry, Gavin. But it's application-specific integrated circuit.

34:48It's essentially a chip designed to do one thing. Okay, back to you. Yes. On a Bitcoin mining chip, what you see is what you get. It only does one thing. And on Sohu, it'll be the same way. of the number of headline flops that we claim, you will find you will get more than 90 % utilization of those flops because the whole thing is built around getting to that really high utilization number. Why does that matter? Because if I can get 38 % and NVIDIA is making bigger, more powerful chips, do I really care about the yield? Is it an economic point? Is it simply a space and storage point? Why does that number matter?

35:22It's a massively economic point because to go run a model, to go get a token out of an LLM, you need a fixed number of math operations. You can actually calculate it very easily to a first approximation. If you just take the number of parameters and multiply it by two, that's the number of math operations you need. So to get one token as input or output of, say, Lama 70b, a 70 billion parameter transformer from Facebook, you need 70 billion times two or 140 billion math operations. So if you're only able to get a third of the theoretical maximum or the amount of math your chip can do, you can only serve a third of that number of customers.

36:01But if you get 90%, you can serve three times more customers. Now, this comparison works well if a etched Sohu ASIC, specifically built for transformers, has a similar power draw to a competing chip because we're talking economics and it's three times as math efficient, but 800 times less electricity efficient, you're losing ground a little bit. So what can you tell me about the power and cooling requirements of the eventual Sohu chip, which I don't think is in the market yet. We have some very exciting stuff coming out later this year. Okay. But how is it looking from an efficiency perspective?

36:35We're going to be 10x better in terms of tokens per watt. Now, watts don't matter for the reason I think most people expect. And I'll show you some back of the envelope math as to why. Okay. For a NVIDIA 8x H100 box, you can expect to pay around$300 ,000 for that system. Which is eight of those chips inside of one cage. Inside of one server, you got it. That will draw around 10 kilowatts. And the rule of thumb is that a watt costs around a dollar a year. Okay, so that would be$10 ,000 per year to run that 8H100 cluster. You got it. What this means is that if, say, the lifetime is five years, you're spending$300 ,000 in CapEx, and you're spending$10 ,000 times five or$50 ,000 in power costs, that the economics of big AI chips are just dominated by the cost of how expensive the chips are.

37:26Ah, it's so interesting. Okay, so if you're talking about a 10x production, though, in compute per watt, a equivalent system from Edge would cost 5k over that same time frame versus 50. Well, the key thing is, in terms of power, we draw the same amount as the new Blackwell chips. But in terms of throughput, you're going to get more than 10x more. So as a result, the watts per token goes down by an order of magnitude. So you guys actually had a riff on your site that just says, what would happen if AI models get 20x faster and cheaper overnight? It's sounding more and more like that's not actually polite height.

37:59It sounds like you guys have a pretty good shot at executing that. So how close are you guys today to getting this Sohu chip out? Is it going to launch this year? Or are you going to show off more of like a demo run of manufacturing later this year? How close are you? And we've got getting a ton of customer interest from a huge variety of customers. We have publicly traded AI companies, clouds to labs to enterprise to high frequency trading firms. For those folks, time to market is everything. The fact is, demand is just not the bottleneck for us to get the product to market. It's all manufacturing.

38:34Have you guys discussed who your manufacturing partner is? Oh, yes, we have picked this out. It's not just one person. that sure you need to have one partner who does the assembly of these servers, but there are so many other parts of that supply chain who's going to go ahead and make the actual dies. How much allocation do you have? Very early on, we were able to work directly with TSMC's Emerging Businesses Group and 4P, which allowed us to get pretty close access to their fab team. How much capacity can you get through that TSMC program? Because I know that everyone right now is trying to get higher yields on their chips and to get more manufacturing capacity.

39:07So from my end, it would seem like NVIDIA is going to buy up all the available manufacturing capacity and leave you guys out in the dust. But it sounds like you do have a good amount of available manufacturing capacity. We're in a very good place there. We have our HBM locked down. We're in a good spot with TSMC as well. HBM is high bandwidth memory, which is an incredibly important part of the AI data center infra stack, if you will. If you have a lot of GPUs, but not a lot of HBM, it's still slow. You need both. And also you need to have a quick fiber and blah, blah, blah. Basically, it's very hard to do this because everyone wants to do it to the nth degree.

39:38But the good news for you is that that means market demand is going to be bonkers. I think it's such an exciting time right now in the AI inference space, especially given some of these new reasoning models and some of the advances in video generation technology too. So this brings me to something that I wanted to dig into because when I was going back to the etched stuff and we were joking before we hit record that you guys have been pretty quiet. So most of the material is a little bit dated, but you were talking about how, quote, scale is all you need for super intelligence. And also you guys were talking about trillion parameter models, which is the pre-training side of scaling.

40:14And then we have post-training scaling. And then we also have test time computer reasoning, which is what you're referring to now. Does the Sohu chip from Etched work in each of those three areas of AI scaling? Absolutely. And I am extremely excited about a reasoning time scaling. Tell me why. I'll give you an analogy. Back in 2006, if you were a telecom company, you were limited by the size of your market. There were only so many Americans. And if you're billing per minute, there are only so many minutes that Americans will hold talk on the phone. And what changed this model? Cell data. Yeah.

40:48The moment the iPhone came out, they were able to go sell way more bits in the same amount of time and grow revenue pretty considerably. And I think we're seeing a very similar moment right now for LLMs. In the past, the limit was the human eyeball. that if every token that comes out of the model has to be read by a person, the bottleneck's probably the person's time. There's only one token in a token. But now that is changing thanks to reasoning models. If every token the user sees, the model has one, 10, 100 thinking tokens, and you get to bill for those, then suddenly the size of the market has gone up by 100x, 1 ,000x overnight, which is very, very exciting for us.

41:26What is this Sohu chip going to cost me? I can't say too much on pricing here. we price 10x cheaper than nvidia in terms of dollars per token so it sounds like i won't be able to see these out in the market for a couple of quarters what's the next milestone from etch that i can have my eyes and ears out for as you guys work towards commercial viability and availability and so forth i can't say too much now but i am a big believer in showing and not in telling and if you like seeing much of their like showing i guess let's have to look forward to in the next two quarters. Okay, so before the end of 25.

41:59Okay, good. Okay, that makes me very excited. If you guys fundraise, let us know. Although given how much money you raised last year, I didn't think you'd be doing that for some time. But I'm blisteringly excited because I love everything I'm using in AI. I have this fancy open AI account. I'm paying Google. I'm in a complete slap for this stuff. So I can't wait to see the future when it's cheaper, faster, and better. So Gavin, thank you so much. And for folks who want to find you online, what's the URL and what is a role you're hiring for? Yes, our URL is etched.com. You're hiring for many roles.

42:30If you are a fantastic firmware, fantastic kernel engineer, we'd love to go bring you on staff. Alternatively, if you're good at manufacturing, we'd love to go talk as well. So if you're doing real world stuff or digital stuff, call Gavin because he is hiring. Please, please do. This has been yet another fun Twist 500 interview and I'm excited about Etched. Can't wait to see how they do. It's great to be on the show. Take care, Alex. I had so much fun with that one. I love talking to people who are so deep on something that I don't, you know, I don't talk a lot about ASICs in my day to day life because my spouse doesn't work in technology.

43:03So it was so much fun to nerd out with them. My big takeaway from this was how much you need to know, how deep your knowledge has to go about what your company is going to do and what your competitors are doing and like where, because he does have like a market fit. Like here's where I have an opportunity. Here's what NVIDIA is not focused on that I can focus on. And like, here's this narrow segment of the market that's open that I can sort of exploit. But if you didn't have this kind of a nuanced understanding of AI architecture and chips and how, not just how the technology works, but what NVIDIA is doing and how you can sort of counter and complement their strategy, you couldn't do this.

43:47Like you could only build this company with this level of institutional knowledge my favorite kind of founder takeaway from this is don't be afraid of hardware i mean they're working with you know tsmc and all this but i mean just if this chip hits the market the way they think it's going to this company is going to print money and i think that we talk a lot about software which is great but there's just fewer companies out there trying their hardware game on and i think we're going to see some real generational companies built in the next couple of years and i'm i mean if i had five if I was an investor, I put five bucks on etched, you know, versus a lot of other companies.

44:21Well, I think with, with software and, and, and these are, you know, our prejudices are our preconceived notions about it, but it feels like, oh, there's so much wiggle room. Like if they're doing this, you could build this other thing. That's not exactly the same. And with hardware, it's just, and, and, you know, like, I don't think we're right, but there's this sense of like, well, if they could build it any better, why wouldn't they already be building it better? You know, like there's a best way to engineer something. And once you've done it, that's it. And, you know, like you guys were even talking about in the interview, like Moore's law.

44:51And like, we have all these like hard, like numerical, like this is how much better chips can get over this much time. And, you know, because it's physical. It's like how, how, how small can you fit something, you know? And so I think, yeah, that's, it's an interesting approach to like, well, there aren't these hard limits in hardware that we necessarily think are like built into the physical world. Yeah. Imagine how cool it would be if overnight the cost of a inference dropped by 10 X. I mean, that would be explosive. it would be it would change everything and i mean that's what we keep we keep seeing this sort of promise like even the the deep seek sort of news and it's like is somebody going to figure out a way to like do the things we want to do but with dramatically fewer resources and power and and and everything else for a lot less money a lot faster and it does feel like we're we're tantalizingly close to that level of revolution we are we're very close but let's drop the hardware game let's pivot over to software.

45:48Our next interview is with a company called Skyflow or Skyflow, if you want. The CEO is Anshu Sharma. I've known Anshu for a long time. Lon, Skyflow is a company that I've been tracking for a while. We'll talk about what it does in a second, but it's a company that started kind of cybersecurity and ended up in the AI game. So for founders out there who think that they cannot expand their vision with time, well, here's an example of a company that did that and is doing quite well. Hey everybody, welcome back to Twist. So today I'm going to bring up a founder that I I've known for a very long time.

46:16I met him back when he was a venture capitalist. Previously, he worked at Salesforce and Oracle, and he started a company called Skyflow. When it came out, it was a company that was built to help companies keep PII or personally identifying information private and safe. But since the company was born, AI has seen a massive boom, and the company has also begun to help other companies keep their data secure in an AI context. I put them on the Twist 500 because of their fundraising history, their growth history, and because I just happen to know the founder and think that he's quite smart. So please help me welcome to the show Anshu Sharma from Skyflow.

46:49Anshu, hey, how is that for a polite introduction? Well, that's very nice of you, Alex. And it's always a pleasure talking to you and arguing with you about the future of everything. Yes. Happily, though, we are not doing this on the phone this time. We're doing this live so everyone can hang out with us. But listen, Anshu, I've known you for a while. I've known the company for a while. Not everyone who's with us today does. So can you go back in time to when Skyflow was incorporated and put together and you guys launched your first product and just explain to me the market you were going after at that time?

47:21I like to talk about it from the perspective of an end user. So this morning, you probably went and bought coffee either at Starbucks or Dunkin', since you're on the East Coast. You know, coffee companies used to sell coffee, which meant that you walk in, you walk out they know nothing about you i'm 99 sure you ordered ahead which means you have a mobile phone with an app that app means that they have your phone number they also now know that you visited this location in rhode island so they know where you live approximately in fact if you think about your coffee company knows more than maybe your wife does because you know who you're having coffee with i can coordinate that and that's just your coffee company now think about the streaming company you interact with, the airline company.

48:06When you used to buy flights 15, 20 years ago, you bought a flight, you know the end of it. Now everybody has a frequent flyer number and that's not enough. They also have a credit card number. Well, why do they have credit card numbers for the frequent flyer numbers? Because they want to collect data. So that's the world I was living in when the big breaches at companies like Capital One, Facebook kept happening. And I got sick and tired of it. And I was like, what's happening behind the scenes inside of these companies? And the answer is they are drowning in data and the data is about you. And the sad truth is we want them to have our data.

48:44You really do want to order your coffee ahead. You really do want that 13 % off coupon from company that sells online dog food. At the same time, we don't want our social security number, credit card number, date of birth, email address, in plain text, in all of these places. And your coffee company was not built like a bank, even though they have more data about you than the bank. And so the question I asked, what does bank-grade security infrastructure for data look like? And the answer was, what do banks have? They have a vault. So we decided we're going to build a vault for your personal information that we sell to enterprises.

49:23And that was the founding story of the company. Fairly simple. What if there was a company that could securely manage this information for the pharma companies that you buy drugs from, the coffee company, down to the last small bit of information exchange you're going to have? So we live in a world where we want to have personalized service. We want agents. Now we want AI. And none of that can happen unless we trade personal information. The coffee analogy works for me because the problem that I've had with cybersecurity and its progress is that it seems that the tools have gotten better over time to keep things more secure.

50:01But the rate at which more companies have taken on more of our personal data has grown faster. And so even though I think the state of cybersecurity has improved, I felt less secure. And that's why originally when you were telling me about polymorphic encryption and privacy vaults and so forth, I was into it. But that's all back in, I think it's the company was founded in 2019. Yeah, that is right before COVID. That is heading into the 2021 boom. So talk to me about the time from founding up until ChatGPT launches and how growth went at Skyflow for the cybersecurity business. Yeah, any company starting out, the first question is, will anybody ever trust you with information and security?

50:41Because you're not follow-up to networks and you're not IBM. Although people do get fired for buying IBM now, sometimes. So I think for us, what we discovered very fast was, if we went and talked to companies that were building fast growing companies, they were the fastest to take us up because they intuitively knew that security was important. So think about companies like Nomi Health, which is a health insurance company, companies like Equal Identity, which is building the clear of India. All of these startups, they were like, hey, this makes sense. I don't have the same 40 engineers that Amazon's pharmacy team has or Netflix team has building their equivalent version of API.

51:25So let me just buy it and I can just turn it on. And then over the next few years, we climbed up from these startups to bigger and bigger logos. So we got bigger customers all the way up to GoodRx, which has publicly known information, over 20 million subscribers, 100 million people's information they interact with. If you want to buy insulin or you want to buy a GLP-1 drug, you can get it for 80 % off if you use our customers' coupons at GoodRx. But that was the journey before Sam Altman changed the world. Yes. But before we get into the launch of ChatGPT and everyone wanting to use their enterprise data inside of AI safely, talk to me about how much your ACV went up.

52:09A lot of founders watch the show, aren't you? And I bet they're curious, as you went from startups up to even now public companies as customers, how quickly did your average kind of contract value change? And looking back, would you change anything about the choices you made along that journey? Yeah. So I always tell founders, you have a choice. You can build a company in a known market where everybody knows they need storage, they need computing, they need CRM. And if you go in that market, you can call up any customers and all the question is going to be like, why you? As opposed to why should I buy this in the first place?

52:45The challenge is, once you get your first 100 customers, eventually you have to become differentiated and then the long pain journey starts. On the other hand, there are companies like MongoDB. If you started MongoDB in 2006 or whenever they started, for the next five years, nobody knows what a NoSQL database is. Even today, if I ask you what's a document database and what's NoSQL, well, you know it's not SQL, but what is it? And the answer is, when you're creating a category, because it's trying to solve a problem in a new way, you have to educate them, show them the new way, and that's how you get started.

53:23So we got lucky, frankly, because security is such a high priority for these CTOs that we were able to get to our first million year are pretty much within the first nine months of our company starting off. But we didn't then go out and hire 100 salespeople overnight. we knew that we had to prove with these first set of customers that product is reliable, product is secure, and the product just works. So we took a measured approach. We've been trying to double or more every year for the last five years, and we've hit our goals. So we're excited about the rate of growth we are at. But the goal is not growth for the sake of growth.

54:02The goal is to prove to the customers now. today we have, you know, you might have seen press releases with ServiceNow, Visa, and all these logos. Yeah, it did. It takes many years of getting good Rx and Nomi Health to be passionately happy with you, to convince a Visa or ServiceNow or Walmart to engage with you. But just putting this in kind of concrete terms, how big were your earliest ACVs with startups? And then I know you can't tell me exactly how much individual customers are, but how big have your contacts become now? Our early customers traders paid us anywhere between a few thousand dollars a year all the way to a hundred thousand dollars a year okay and now that would be considered a very small deal that maybe one of our inside sales rep handles and the size of deals with some of the largest companies of the world is truly orders of magnitude higher got it i can't share those numbers but think about the largest fintech company in Indonesia, right?

55:02The largest e-commerce company in all of Asia, they have 380 million users data. When they sign up with Skyflow, you can multiply that number with, hey, it's still an extremely large number. Thank you for that. But let's keep on this chronology. So, ChadGPT comes out, the world goes, oh my God, LLMs are the best. Everyone races to begin to experiment with them. Then as time passes, people get a bit more sophisticated. And I think people really got the idea that if you want to apply AI to your business, to your software, you're going to want to use your own data. And there was some news items out there about companies accidentally leaking information into LLMs.

55:42So I presume pretty quickly on shoot that the market was like, okay, if we're going to go this route, we're going to need to have a solution that will keep our information both secure. And I think segregated in such a way that it had a different, I'm going to butcher this permissions structure so that way llms didn't reach it and grab the wrong thing at the wrong time so talk to me about market reaction from your perspective to the post chat gpt world up through today yeah so when chat gpt came out you know i started trying it out and i was like this is fun i have no idea why like people are worried about privacy and security because all i'm asking it is what i call wikipedia questions right they're in english language and so i can say tell me more about why India got independence in 1947 versus 20 years ago.

56:30But the data I'm talking about is just data that's available to everybody. It was my chance conversation with actually the co-founder of Salesforce, Parker Harris, where he started telling me about what they were beginning to build for their customers and how they had to worry about security and privacy. And that led me to start thinking about, wow, when you take this technology, apply it to customer data, you now have a whole set of challenges. There's challenges for software vendors, whether you're starting a company that's automating call centers or legal documents, you now have to earn the trust.

57:10Or the enterprise that's going to consume that technology, how are they going to hook up their most valuable data to these startups or even large enterprises and still trust them. And so this trust problem became very apparent. And it actually took me a few months of these conversations repeatedly with very, very large companies to realize, oh, polymorphic engine that we built for securing data in your databases, in your CRM, in your snowflake and data bricks can perhaps be purposed for unstructured data. And the light bulb moment for us was a healthcare company that was building AI models. They just want to be able to build models with real data because you can't buy fake synthetic data.

57:56If you bought synthetic data, you would find out that 20-year-olds have the COVID at the same rate as 80-year-olds, right? Which is just not true. So synthetic data, in one sentence, is amazing to do what I call scenario testing. You know, what if the car started driving on the road in the opposite direction? It's hard to create real data for that on a one-on-one. Synthetic data is easier. So when you're testing by creating outlandish scenarios that you may not have thought about, or just testing the basic functionality of your software, should I punch a bazillion dollars into this field and see if it implodes?

58:33Synthetic data is great for that. Okay, that makes sense. But the key word in AI is machine learning. You don't want the machine to be learning off of bad, fake data. You want the machine to be learning off of real world data, but without knowing my date of birth and social media number. Right. So we started entering the market there. Can you de-identify and certify the data that's going into models? Can you do it not just for structured data, but also for PDFs? Actually, one of our customers passes all the voice files through us. So how can you build models off of real data? And then you can test it with synthetic data, by the way, later on without risking any of the sensitive information.

59:17Sensitive information is not just personal information. If you have a rare disease, that's sensitive information. If you are a drug manufacturer and running a clinical trial on some new drug molecule, name of the molecule is sensitive information. So it turns out all of those pieces of information need to be protected in the same way that we will protect yours and my social security number. So we took that technology, applied it to all of the sensitive information, and started shipping our model training and drag infrastructure product line. Which was the Skyflow GPT privacy vault that you guys put out in 2023, I believe.

59:54Exactly, right. Six months after we started playing around with it. And honestly, at that time, it was more of an experiment, right? You know, I learned both at Oracle and Salesforce, ship early versions and see the test the market. And the response was interesting. And startups were immediately on top of it because what was happening to them was they were building these amazing demos, getting those ARR that's not really repeatable because innovation team has greenlighted the product. Right. And then getting blocked by the CISO, getting blocked by the chief privacy officer. Because who wants to save$3 in a call if the personal information about a customer's life insurance policy is going to get leaked?

1:00:36Yeah. So as startups running into that wall and bigger companies started running into this wall, we started getting interest from them. And then it took another six to nine months before enterprises started realizing that because it went from being a vendor problem to being an enterprise problem. So when we spoke back in, I think it was kind of early 2024, you had said that, you know, you were now seeing about 30 % of your revenue. I think net new revenue was coming at that point from large language model related usage. How much of your business is now predicated on AI versus kind of the cybersecurity picture?

1:01:11And how is the revenue mix changing as 2025 gets underway? An amazing thing happened. We took our data security product, applied it to AI security and privacy. Meanwhile, Ali Goetze and the new CEO, Snowflake, either had this vision that actually AI and data are the same platform, right? Things you're going to build on your data is the AI applications. And so weirdly enough, our product is no longer two distinct, complete solutions because our customers have their data in BigQuery, in Snowflake, in Databricks. They're trying to build models, they're built in chatbots. So they actually now view this as the modern AI data stack.

1:01:57The way that I've been thinking about things, as in you guys protect PII and also help companies use their data inside AI models safely, really, those are not distinct use cases anymore. It's all kind of the same product and sales cycle. Because what are you building AI on top of? It's your data. Yeah. Okay. Where is the data? Exactly where you are. But it implies that companies that were probably even prior Skyflow customers are working their way towards building their own AI services, solutions, models, whatever. Is my read correct that what you're saying is that every company wants to use their data inside of AI and this is not a subset of your prior customer base?

1:02:34It's pretty much all of them? Yeah, so all of our conversations now start merging with each other. It still happens that some companies will say, hey, I want to build an internal IT help desk AI, need to protect the data going in. Let's start there. And then we say, hey, so where is this data coming in from and where is this data going? Oh, you know, my application is ServiceNow and my data lake is Databricks. Okay, now you have to protect the data going in and out of these places. So a perfect example is, you know, we just announced a partnership with Visa and ServiceNow. Now, ServiceNow has this amazing new product called credit card dispute management, right?

1:03:13You charge back, you dispute a charge. Well, in the old days, that would be a manual process. And ServiceNow, with their AI platform and their data platform, they want to make it such that a bank or a merchant doesn't have to spend weeks figuring out what's going on. But to process that, what information are you processing? What is that data warehouse app or model looking at? Well, it's the details of your transaction. It's your credit card number. It's the thing that you bought, maybe a medication that you bought, but you didn't really buy because someone else bought it on your behalf. Well, all of the data has to be carefully protected before you will let that information go between multiple systems.

1:03:56And that's what we are doing with ServiceNow and Visa. So very, very simple use cases, but deep, vertical use cases. suddenly you realize, oh, AI is all about data and workflows, and therefore it has to be protected in that context. I mean, that's bullish for future Skyflow growth. I'm curious though, you mentioned earlier that you guys have been, I think you said something along the lines of doubling every year or shooting for that. One thing we have seen from a number of what I might call AI first companies is historically quick revenue growth. And I know you and I both share a little skepticism about how durable some of that ARR is, innovation budgets, people testing, etc.

1:04:35But no matter how you slice it, growth for the leading startups today has never been, I think, faster. So have you felt any pressure as a founder and CEO to try to grow SkyFlow even faster than old-fashioned startup growth norms? I think I learned this from looking at the growth charts of Salesforce and HubSpot and every other company that's durable and large. The key is growing 2x or more in the early years and then being able to sustain that growth of 50 % plus for another 5-10 years and then 30 % for another 5 years. What we are selling is so core to the infrastructure of these startups, these enterprises, these AI solutions people are building that you can't just go in and say, hey, why don't you just try it out?

1:05:26If it doesn't work, come back to me later. I would rather have 10 customers like ServiceNow, Visa, and others who are referenceable and happy than have 18 customers who are trying it out. So our revenue tends to be much stickier. It's kind of like, you know, once you build it in, you leave it on. Nobody turns off data security. Unless they want to have a breach, I guess. Exactly. You're saying that your gross retention at Skyflow is very close to 100%. Yeah. So companies that die are the only companies that leave Skyflow. There's only one way out. It's in a casket. A little dark, but no, it's good news for you.

1:06:05I mean, last year, a lot of fintechs did die. And now they're back in growth mode again. So it does happen. So we're not immune to the market forces. Sure. But customers that rely on data security, data residency, building the AI models, tend to stick with it for a very long period of time. And I think you can always grow faster by hiring more salespeople and there are times and phases for that. And we are planning to grow faster, but I think it's been exciting. So it seems that in and around the Skyflow area of the market, there is elevated compared to the last couple of years, M &A interest.

1:06:41And so I'm curious, Anshu, are big companies sniffing around now? Are they trying to kick the tires on Skyflow and see if it's something that they want to pick up for themselves? You know, I've worked at most of these companies that are most of these CEOs. We are always talking to them, working with them, but in the context of partnering with them, right? So we have partnerships with Snowflake, we have partnerships with Amazon, we have partnerships with ServiceNow. You know, my goal is to build a generational company and the best way to build a generational company is to have these companies as partners.

1:07:14People have been talking a lot about where value will accrue in the AI stack, if it's going to be at the data center layer, what Jensen's now calling AI factories, the model layer, the application layer, or kind of where I put Skyflows, if you will, which is like enabling the app layer on top of models, if that makes sense. I'm just curious from your perspective and the other founders and CEOs you talk to, where do you think most of the value will kind of pool? Because I'm sure that other founders want to know where they should be building? A very simple way to think about it is where is actually value getting created?

1:07:49For certain types of businesses, the value is in legal know-how. So how we can capture a lot more value than whatever else they're using under the hook. So verticals aside, for most of the other things, it comes down to actually how replaceable you are. In a world where there is only one leading cloud provider, all the revenue growth was going to one large cloud provider. Today, we have three or four cloud providers. So the price is competition. We only have one GPU player that was meaningful until a year ago. And that's mostly still true. And if that's still true, you have pricing power. So pricing power goes between layers, not based on how important you are or how big of a keynote you do.

1:08:36Pricing power shifts on how many other companies similar to you can deliver almost the same functionality. I call this, are you one of N or N of one? So one of N is there is N companies that can do this, right? Compute, virtualization, storage. And the moment N companies can do it, you're basically doing feature function price analysis. N of one is there is only one NVIDIA three years ago that's building a GPU on which you can tell you more. You can complain as much as you want about Blackwell being late by three and a half months. What are you going to do? Unless you're Google, you can't build your own TPUs, right?

1:09:18So I think that's the framework where we go. As a result, something has fundamentally changed. 18 months ago, I assumed there would be a few companies just like GPU was the only layer. There was only one model company that was winning the whole thing. There's only really one or two platform companies. But today, Oracle's in the game. So is multiple model companies, DeepSeek, Meta. So as it is built, I believe the more other layers get commoditized, the more unique companies and unique assets, not just us. You just saw Weights and Biases get acquired for like a billion dollars, basically. Why? Because there's not any players that can do what Weights and Biases does.

1:09:59If there was five companies that were very similar in success as Weights and Biases, they'd be acquired for$50 million. As a founder, my job is to create value for customers, which is revenue, ARR. But my job is also to create the multiple. And the multiple gets created when you are truly building something differentiated and you are end of one or end of two. All right. Well, Antje, thank you so much. We'll have you back on in a couple of quarters. In the meantime, thank you, everybody. And don't forget twist500.com, the coolest companies in the world. Takeaways, Lon, from the Skyflow chat, what you got?

1:10:32I think it was one interesting thing for this chat was, you know, you normally think of sales as being this, you know, sort of aggressive, like you got to close the deal or whatever. And he was really talking about it in terms of trust that when you work in security and cybersecurity and protecting people's data, it's less about giving people the hard sell and more about establishing like, we know how to do this. We're going to actually protect your data. And then once you sort of do establish that and people agree to sign with you, they don't leave. You don't then lose those customers because it's about building up that level of trust.

1:11:06So it feels different than a lot of other kinds of companies or B2B sort of businesses where, you know, the idea would be you shop around. There's always somebody who's going to maybe make a better deal. And this is really like, no, it's something more. It's a different kind of market. There's a luxury, though. And so I don't think what Anshu is saying applies to every single situation. As we said, their gross retention is near 100%. And so if that's the case, then you're not going to be dealing with kind of standard SaaS churn. And so probably longer term relationships are even more important, ergo, trust, et cetera.

1:11:41So this is pretty far away from like PLG, AI, consumer SaaS, right? Like it's the other end of the fence. But it is interesting to see how a company that does sell to the largest companies approaches that. So if you are thinking about going up market, I hope Onshoe's comments were useful to you. I think Skyflow is a cool company. I think they actually did a small down round to their last funding round because after their 2021 valuation. But I suspect they're going to be okay and keep on trucking. So that's why I put them on the Twist 500 lawn. And we have one more company. Shall we hear from them?

1:12:13Please. Can't wait. All right. So we're going a little bit of field this time. You know, we started talking about chips and AI. Then we talked about cybersecurity and encryption. Now we're going out to outpatient clinics. We're going to talk to a couple of guys from MedServe. MedServe is a startup that is building essentially secure cabinets for medications. Now, why do we care about that lawn? Well, it turns out there's a lot of rules about that. And sometimes people might want to take a medicine that's not for them at that time. So hospitals have large secure cabinets with relatively intense security in them.

1:12:48But those are big and costly and they don't fit in an outpatient thing like an urgent care clinic or a partial care program, for example. So MedServe is tackling that particular niche and they want to build a similar level of security, but for smaller clinics out there in the market, replacing not pen and paper this time, but instead small filing cabinets that are not sufficiently secure. So let's hear from the MedServe guys to take us home. Hey, everybody. Welcome back to Twist. We have been banging on about AI for quarters and quarters and years now. But sometimes lost in the startup game are the companies out there that are really impacting the real world slightly out of sight.

1:13:26And in the United States, healthcare is one of the biggest industries, period. And so when we look into the startups that are working in that space, we often find interesting things that are just outside of our vision. One of those companies is Medserve. They are building a very interesting cabinet. We'll get to that. It's going to help with both drug dispensing and preventing drug diversion. I'm really excited about this company, so please help me in welcoming. We have Alex and Matt, the CEO and president of MedServe. Hey, guys. How's it going? Hey, Alex. Nice thing. Hey, great to be with you.

1:13:55So I'm just going to start at the top here. Teal is the company color by the looks of it. That is. Yeah, we're matching today. Yes. I love when people show up in coordinated colors. It reminds me of like old YC demo days when teams would all wear like bright purple T-shirts so that we could find them in a little group. Ah, old days. That's just how me and Matt are all on the regular. Well, I appreciate that. Now, another thing that I like about Medserve is that you guys are actually a Midwest-based company. You guys were founded, I think, in Milwaukee. Is that right, Alex? Yeah, yep. Milwaukee-based.

1:14:26And I got to say, I kind of grew up in the Chicago tech scene back when I was in university. So I do have a big appreciation for the Midwest, but I don't know jack about Milwaukee. So tell me about the startup scene there. Is it active? Is it burgeoning? Should I go visit? Well, actually, let me turn this over to Matt. Matt's got the background in Milwaukee's tech ecosystem as in startup ecosystem as he was one of the builders. Ah, Matt, please. Yeah. So Milwaukee has, I think, a lot of great potential. You know, I think building a company in Milwaukee has allowed us to access talent that we may not have necessarily been able to find elsewhere.

1:15:02And what's really cool is, as you can see behind Alex, we manufacture medication cabinets here in Arcata cabinets. And we do that right down the street. And we're seeing Wisconsin about 20 minutes away. So it's a great place to have a startup that you have to bend metal as well as push lines of code to be successful. You know, I heard that all American manufacturing was gone. So I guess that wasn't true. Is there still a pretty solid industrial base in the Milwaukee Midwest area? Yeah, we've had a great, you know, time in terms of co-developing. Our technology was originally developed by a manufacturing company that didn't necessarily see the, you know, technology application of, you know, cloud-based software and doing the reporting and stuff that we do.

1:15:54So, you know, we were able to take that idea and kind of take it to the next level by acquiring kind of some of the intellectual property around MedServe's cabinets and improving on it. All right, we're going to get into how a cabinet talks to the cloud in a second. But first, I want to talk a little bit about the problem space because my spouse works in healthcare, so I have some visibility into what we're talking about, but a lot of folks don't. And so when we think about drug dispensing and drug diversion, it might help, Alex, if you just kind of run through the terms, what they mean, and then just anything you can tell about the scale of the diversion problem that we have here in the States.

1:16:30So Alex, I think probably to level set, my background is I'm a pharmacist, been in the pharmacy world for nearly 20 years. That's hard to say, but nearly 20 years. I grew up professionally within a large health system, you know, health systems and fancy hospitals. They have all the fancy technology. Outside of the hospital system. However, you're like traveling back in time. So the problem that we're solving, we're providing digital medication security and management at the point of care. So the point of care, what I mean by that, you can think of it as a surgery center that's out in the community, a dentist office, a clinic, an imaging center, a veterinary's office.

1:17:07There's so many different areas where medications are housed and administered to a patient that are outside the hospital that don't have the hospital technology or budget. Okay. So is this a little bit like if I go to rob a bank, like I know there's a vault, there's security, there's cameras, but if I want to go to like an ATM, there's probably a little bit less security. It's a little bit easier to, well, maybe not ATMs, but you know what I mean? Essentially, if you're out in the field, the fences are lower. Yes. Yes. So I guess if you went to a bank that's inside of a grocery store and try to rob that, maybe that's a good comparison.

1:17:37Do they have banks inside of grocery stores in the Midwest? They do in the Midwest. Okay. That's a great spot. In the little community you got everything in one place i think that's because it's so cold there you don't want to go outside and change venue and i recall that from my chicago time yeah okay so we're talking about surgical centers dentist office these are all around the country and because they do need to deal with controlled substances even addictive substances they need to keep them safe so if i did not have access to a med surf cabinet what is the standard legacy solution you guys are going out You're not going to believe me.

1:18:10So these facilities literally use like a kitchen cabinet that has a lock on it or two locks. Somebody has the keys on them. So sometimes one of the stories we hear all the time is people go to lunch with the key. People leave the office with the key and then they can't access narcotics because they left. Some people will use a metal, like little more secure metal box with a key, but the key problem remains the same. Many facilities will also invest in the little like realtor key holding box that's on the wall. It's got the little pin code you can put in, but the pin code shared. When somebody quits, nobody replaces it.

1:18:44They should replace it and reset it, but it doesn't happen. So they're opening stuff up to so much risk by doing this really old school process. Now, this is just the physical storage of the medication. Now, there's also a documentation component. The documentation component, I don't know if you've ever, if anybody remembers the prescriptions handwritten by a doctor, imagine those scribbles in a bigger hurry done by an anesthesiologist on the paper log where they're just jotting down how much they administer to who and trying to make sense of that. And people forget how to do math. People make mistakes, forget to enter data.

1:19:17And a poor nurse or director of nursing or a manager at the end of the day has to sit there and try to figure out the numbers and why they don't add up. Matt, so if the numbers don't add up, does that automatically ring an alarm bell? And if so, how shrill is that signal? Well, it's alarming because unfortunately, a drug diversion or drug theft. Yes, Alex, you're welcome there. It really is a growing problem in health care. Unfortunately, health care providers, about 7 % have self-reported that they struggle with substance abuse issues. We unfortunately see horror stories of health care providers diverting or stealing drugs intended for patients, patients going through health care procedures without proper anesthesia.

1:20:04And when this happens, it often, you know, it has to be reported to the DEA. The DEA is the reason why, you know, we exist in a lot of ways as we're helping, you know, with the compliance, automate that workflow, secure narcotics to their standards. So all these facilities, you know, when drugs go missing, it is very alarming. And then just to add a little bit to it, so it's alarming in a number of ways. One is, you're short a controlled substance that you are responsible for. So the center, that is the way the dispensing, holds the bolus of responsibility. Yes. So it's the center, it's the person in charge of the center, so let's say the administrator, and it's the physician whose name is on the license, the DA license, who's really got the most to bear.

1:20:51So if there are issues, let's say this is negligence, there's not great policies in place, procedures aren't being enforced, the facility can be shut down, they can have their license suspended, so they can basically be put out of business. And that's without having horrible headlines like Matt was mentioning. And, you know, we can dive into that a little bit later if you'd like. So you have that as an alarming issue. A controlled substance that you're legally responsible for is going missing. Second is, well, where did it go? Is one of your employees under influence taking care of patients? Or is like Matt pointed out, is one of your patients going through a procedure paralyzed, can't move, can't yell, but doesn't have pain medicine?

1:21:30So the liability here is multifaceted. The problems are diverse. And it sounds like that you're really protecting the drugs from the providers. I was curious if this was keeping drugs out of the hands of patients, but it sounds much more like the concerns on the provider side. Yeah, it's on anybody who works in the facility and has access to the medication. So it's nurses and providers. Does this also apply to like nursing homes and urgent care clinics and that sort of thing too? Yes. Anywhere and anything and everywhere where there's medications, there's a risk for things to be diverted. So whether it's a controlled substance that people are using to get high, whether it's a medication that's just being stolen because it's expensive or because it's just not tracked.

1:22:11I, you know, I'm hungover today. I took a Zofran, which is an anti-nausea medication to make me feel better. It's only a dollar or whatever it may cost. I'm stealing from the organization. I'm just thinking about what I've heard about compensation for staffers who work at like nursing home like facilities and it being incredibly low. That plus the cabinet with a little lock just sounds like an absolute catastrophe, not just waiting to happen, but must be happening all the freaking time. How have why did it take so long for you guys to build a better solution? Like, why did why didn't the market solve this like 15 years ago?

1:22:44And I say all this as a recovering addict who went to rehab. So like I am sympathetic here to the issues at play. Yeah. So Alex, I think this problem has been being solved for the last 20, 30 years. It started with hospitals. So technology for this didn't exist in hospitals 30 years ago, but it started. It started slowly with certain more progressive hospitals. Now every hospital has an automated dispensing cabinet that is on every wing of the hospital floor. centrally located for medications for the patients. But it hasn't spread outside of there because economics really weren't there. These smaller facilities can't afford the premium price tag for a product that's built for a hospital.

1:23:27Okay. Talk to me about dollar six. I know that in the healthcare world, prices get a little bit inflated. So if I was a hospital, let's say a 500 bed facility, and I was going to put the legacy enterprise solution, if you will, onto each floor, how much would that cost me? Millions. Millions? It's a filing cabinet with a cool lock. The systems in the hospitals are very complex and very comprehensive for good reason. They follow different workflows. They house a ton of different medications. There's a lot of intricacies and complexities. So each machine on every floor may have two or three machines.

1:24:07Each machine is probably$150 ,000. Wow. Okay. Times however many floors, plus however many other areas around the hospitals that may have those machines. Okay. So just to summarize, we have a ton of healthcare facilities in the nation. We have a ton of controlled substances. We have a theft and abuse problem. And we have existing technology that only applies essentially to the biggest players in the market. You guys want to bring it at a lower price point to the facilities to prevent theft. So then the question becomes, how much cheaper is the MedServe cabinet that I can see behind you compared to those$150K examples you just mentioned.

1:24:40Sure. So the cabinet that you see behind me is$8 ,000. Oh, well, that's like 6 % as much. Yes, it's very, very affordable. So they're priced right for the market. The beauty of our system is our system is scalable and it's modular. So a facility can go in with a smaller cabinet or scale up their locker to be much bigger based on their needs and thus their budget. The bigger the facility, the larger the budget, of course, that they will have. and same thing. Most of the products in this industry are hardware and software products. Same with ours. So the hardware is$8 ,000 and the software can range between$150 to maybe$600 a month.

1:25:16And this is thousands if you talk about the other systems that we talked about earlier in hospitals. All right, so what I want to know, Alex, is how this thing actually works. If I was out in the field putting this to the test, walk me through a standard use case, if you would. All right, perfect. So let's say I'm a nurse and I'm at a facility and I need to take out some medication for patients. So we'll sign in. So I'll scan my badge to sign in. And I'm going to go and take out some fentanyl. As probably many listeners and viewers know, fentanyl is a highly abused controlled substance. So it's behind a digital double lock.

1:25:49It's a two-factor authentication. I have to put in my pin code as well as scan my badge. So I'll put in my pin code here. It helps if you put in the right pin code. Well, it just shows that the technology works. So then we'll go through and now I can take, I can return medication, but we'll take some for a patient. So we will take, let's say, one vial of fentanyl for a patient and select our patient. This comes through from our integration with the EHR or electronic health record. So we'll select the patient here from the list and go ahead and it unlocks the compartment for us. And now this is a safety step here where we're asking the nurse to count how many vials are in the cabinet prior to him or her taking them out.

1:26:27So this is where if somebody takes an extra cabinet or extra vial or extra pill, because as you can see here, you have access to the entire content of the cabinet, you're double checking the person before you. And if the count is incorrect, so let's put a 11 or 10, and hopefully it's going to be wrong. It gives us a warning that the count isn't correct, and I can try again or create a discrepancy. When I create a discrepancy as a nurse, I'm able to continue and take care of my patient. However, my administrator or my manager just got an instant notification letting them know that there is a miscount on medication so they can come and address it right away and not have to wait until the end of the day for them to figure this out.

1:27:03And there's no siren that goes off, but essentially there's a digital bat signal sent to the right person to alert them right away that there's a potential discrepancy or loss of a drug. Correct. And then also there is a little icon that now appears on the screen in the corner here to let us know that there's an alert and that something needs to be addressed. All right. And then if you close it up, let's see what happens. So now it's just closed. Now I can go in and access a different compartment. So here I have four medications. Then I can go through and take out another medication for another patient or the same patient.

1:27:36All right. A certain step here require two people. So for example, if I wanted to restock the cabinet, we require a second person so that you always know accurately how much stock was actually put into the system. Oh, so to prevent people from not putting in as much. I see. Yeah. And actually any of these steps here that you see, any of these can require two people. So different facilities sometimes will have a lot of staff turnover and they always want to have extra security. So they will require a second person. And how tall is the overall cabinet that we're zoomed in on here on the video? So this one's 36 inches tall, but they come in different sizes.

1:28:12So there's a smaller and a larger cabinet. And just because I'm now curious, how tall is your tallest cabinet? This is actually the tallest cabinet. Oh, I'm sorry. But they are stackable. And so you can have multiple cabinets together. So typically for average person's height, this is the great height from top to bottom where they can reach it. And then they would just have multiple next to each other. Think of it like a locker wall. Two more questions. And they're both about security. Now, normally I talk to folks who are just building software. So we only talk about cybersecurity, but you guys actually have a thing.

1:28:42If I was to bring in the sledgehammer or a cutting torch, how robust is it versus me, the unintelligent man with a hammer? so uh this is this is a good point uh so big distinction between a safe and what we do we're not a safe we are we are a cabinet that keeps things separated and tracks all of the activity and yes we keep it more secure than a regular cabinet but we are not a safe so have you if you bring in a jack a sledgehammer or whatever it is that you bring in yes you can break into it now you have to rely on the facilities security systems because they have a certain level of security that protects their fence and all that's sitting in the kitchen cabinet door door locks employee badges uh cameras yeah exactly some facilities even go as far as put um shocks uh motion sensors or shock sensors on the cabinets if somebody shakes it i it sets off the alarms similar to like your window uh shatter sensor you mentioned scales earlier those are kind of like the las vegas mini bar effect right that if you take uh yes so this is that that's uh something that we have on our roadmap is it has to be at the right price point.

1:29:47Oh, no, no, no. I'm not saying you guys are lacking that. I'm just trying to understand the various levels of security here. Okay, so that helps a lot. Now, in a cybersecurity concept, it has to be much more durable than a cabinet. So I know there's a lot of regulation about medical data, but how hard was it to get your cybersecurity set up to the point to which you guys were confident to take on that responsibility for your customers? Well, it's been a journey and we're continuing. you know it's it's cybersecurity is something that obviously continues to evolve and whatnot we take it extremely seriously we've built out a security team with you know various vendors and internal resources um here to industry standards and best practices around data privacy and security and um that's allowed us to gain the trust of some of the largest health systems in the country who are excited to count as customer so and uh no major breaches problems so far all right well i i don't know yeah i haven't seen your logs um but i'm really glad to hear it um and then finally just one last question about the the long-term future here um are you guys gunning to eventually go public are you hoping to sell to a major provider in the space what's the uh what's the end game you're working towards i definitely see a strategic and end game here um a lot of providers in this space and definitely are looking to build something that they have not necessarily figured out how to do, which is go to market these tens of thousands of endpoints in healthcare.

1:31:21Alex, do you agree you're not going to take this public? You're going to sell out before you reach Wall Street? I'm kidding. Yes, I think the market is just prime with a number of strategic buyers for whom would be a great add-on, either providing them a market they don't have or a solution they don't have. Well, I mean, if there's a plethora of possible providers who might buy you, make sure to, you know, make them pay through the nose when it's time. Actually, sorry, one more question. When will you guys feel like you've like gone far enough alone as a company that you would be willing to entertain an acquisition offer?

1:31:57Because it's kind of a matter of taste, I think. yeah and i think uh so today when we talk about it we talk about it being a three to five year timeline from where we are today okay um but i think things evolve so they've seen what the opportunity is ahead of us uh and uh we it may make sense to do this for longer it may make sense to uh exit sooner if the right opportunity comes up well i will be pulling for an ipo but in the meantime i'll be keeping my eyes peeled for m &a but i'm really happy with what you guys are doing I love the idea of keeping drugs out of the wrong hands at the wrong times.

1:32:32It's an issue that's very important to me personally. So thank you very much. And if folks want to find you online, where should they go? And what is a role you are hiring for that you are having a hard time landing a candidate for? Well, I'll start with how to find us. So we are MedServe and MedServeRx.com is our URL. We're on LinkedIn and quite visible there. and I think we're always hiring for people that are going to help med serve to the next level expand and grow so always looking for account executives that have experience and healthcare technology sales or capital equipment medical equipment sales all right well there it is ladies and gentlemen med serve it's not med server x.com it's med serve rx.com that threw me off and I was happy and then prepping for the show guys thank you so much and come back when you reach the next revenue milestone.

1:33:24All right. Thanks, Alex. Thanks, Alex. I think all founders should be forced to wear color-coordinated tops when they come on the show because that was too much fun. The most intriguing part was when you asked, is this like a safe? Like you're imagining like the Tom Cruise dropping down from the ceiling. No, because that was the question that I had. Like, I think it's really interesting. If you thought about this as a consumer-facing product, that's what it would need to be. It's like, this is the next generation safe because your apartment's not that safe. It's like, if you're going to keep your valuables in here, we're going to protect it from a guy with a crowbar.

1:33:59And what these guys are saying, they had that level of knowledge of being insightful about working at a doctor's office or a pharmacy, or they've already got protocols. Like you're not in the back at the pharmacy anyway. They're only looking out for the people who are already licensed to be there. Yeah. Not, not the, uh, impossible mission force breaking in through the ceiling. Yeah. It also just goes to show how specialized things are because I would have just presumed it was a lock with, you know, the spinny little, you know, like a safe style thing. The three metal bars that come through to make sure that Gerard Butler and O'Shea Jackson Jr.

1:34:35can't just crack in, you know. You see, and that's why I love having Lon on the show because he actually knows the names of people who appear in films, whereas I just go, it's that guy. I've seen him before. They're in the den of thieves, Alex. They're in the den of thieves. Well, I just want to say that I did, frankly, kind of come of age in the Chicago technology scene. So I will always have a deep and abiding affection for people that are building in the middle of America, even though I have now lived on both coasts, which is kind of where I culturally fit. It's great to see people out there building great companies everywhere.

1:35:06Proof, founders that are watching, that you don't have to live in the same eight blocks of San Francisco as everyone that you see over on X. Lon, I think we can wrap here. We're back on Wednesday. We're back on Friday live shows. Jason will be back with us. Tons of news coming. I'm already kind of pulling in stuff for the Wednesday docket. But Lon, where can people find you on the great wide internet? Go on Twitter at L-O-N-S. That's the easy or not. It's X now. We should be saying Twitter long dead. Go on the place that used to be Twitter. I'm at L-O-N-S. That's the easiest place to find. And I just want to say that because I knew Lon's Twitter handle before he was on the team, I keep trying to at lawns in Slack and it keeps absolutely not working.

1:35:48So can you, can you fix that for me? It's not my real name. It's not actually my name. Well, it's how I think of you. Anyways, I'm Alex, uh, Alex over on X. Of course, this week in startups comes out three times a week. We're on every podcast platform. We streamed live on YouTube and LinkedIn and X and Instagram. And I think also TikTok now, and we're trying to go everywhere. So wherever you get your podcasts, we are going to be there. Back soon. See you then. Thank you.