Episode Transcript
[00:00:03] Speaker A: Welcome to INA Insights, where prominent leaders and influencers shaping the industrial and industrial technology sector discuss topics that are critical for executives, boards and investors. INA Insights is brought to you by Ina AI, a firm focused on working with industrial companies to make them unrivaled. Segment of one leaders to learn more about Ina AI, please visit our website at www. Dot ina dot AI.
[00:00:40] Speaker B: Good morning everyone. Welcome to another episode of our INA Insights podcast. We will continue our focus on AI early stage companies today. Our guest is Mister Jo, a company that specializes in AI powered process control for semiconductor manufacturing and merges physics insights with advanced AI and machine learning for unparalleled automation and control. 7 million in funding prior to this, John was vice president and CTO of VMware's cloud management business unit, which generated $1.2 billion per year for the company. He also founded several other startup companies, PI Corporation in 2006. John is a former tenured professor of computer science at Oregon State University and his highly cited academic research work was awarded the award for contributions to the field of recommendation systems. John, welcome to our podcast. We are excited to have you here today and we look forward to talking about Tegnas and your journey as a founder and CEO.
[00:01:41] Speaker C: Thanks for having me. Perfect.
[00:01:44] Speaker B: So let's get started. Tell us a little bit more about Tegnas. John, why and how did you start the company?
[00:01:50] Speaker C: Yeah, so we started, let's see, almost seven, eight years ago now that we started Tegnus, when I was CTO of the cloud management group at Bull company, we had customers everywhere and had a chance to kind of see where the market was going. And we were, I was really addicted to starting startup companies. That was kind of an anomaly that I ended up as an executive at a large company and it was sort of through an acquisition and they just kept giving me interesting I was looking for the next startup company to start and I really wanted to apply what I thought were some of the core principles of what made the cloud super successful. It was really about automation. It was about really moving to management by exception. It was by putting those structures in place which would in turn allow you to do optimization. And we all saw what happened with the cloud. It completely transformed the world. And so we're looking for areas in which those same principles could be applied, but was not sort of it automation necessarily. And we settled on industrial automation. We thought there was going to be a big convergence of they didn't have this level of transformation hadn't happened yet. But sensor technology was getting cheaper and more advanced. Wireless network technology was getting more advanced. AI was getting more advanced. It seemed like trying to time it. We said, let's go try and disrupt industrial automation. And that was really the beginning of Tegnas, and why we chose to start it, and why we chose to go after the industrial space.
[00:03:10] Speaker B: It looks like within the industrial space, your focus is on semiconductor manufacturing. Could you explain a bit about the type of problems that you're trying to chain in AI for it?
[00:03:22] Speaker C: Sure, yeah. So when I started the company, when Matt and I started the company, we didn't actually expect to be specifically targeting semiconductor, but I can tell you, as a startup person, you really have to focus. We started out not so focused, and we tried several different kinds of industries, but at some point you only have so many go to market resources, you only have so many cycles to understand your customer, to understand your market, et cetera. And that really necessitates focusing on a particular industry. It's very hard to go broad as a startup company. So in the end, we decided to pick semiconductor for a long list of reasons, macroeconomically, very going strong, lots of geopolitical pressures, creating opportunities, and also, again, similar. They had not really had this disruption that we had seen. So the challenges in semiconductor are not so different from many manufacturing situations, just much more expensive and extreme. What do manufacturers struggle with? Yield.
Not everything you build in a semiconductor like you put on a wafer, not all those chips work when they come off the end. Some have to be thrown away, they don't meet the quality standards. Sometimes you're scrapping wafers along the way, they don't make it to the end of the line. So yield is a big issue, costs are a big issue, and that they're kind of related. The two are connected, obviously, if you're throwing away wafers, your costs, but even when you are producing decent wafers, you also are trying to manage your costs on that front.
And then really the third one is time to market.
Right now, if you come up with a new kind of material that will enable some new kind of semiconductor chip, it takes about ten years to get to the point where you're manufacturing that at a reasonable scale. And this has all been seriously. So why is it so expensive now in semiconductor? Well, semiconductor is just a much more extreme version of manufacturing, particularly at the leading edge. The sensitivities are so tight, you're basically doing nanoscale construction. And when you're doing nanoscale construction, like a single hair follicle can look like a meteorite that takes out your thing. And also just how do you start trying to lay things down at the nanoscale layer? You can't even see them, they're so small. You make sure you don't overlap, et cetera. And so the complexity of these chips, the leading edge, is getting so crazy because they have to put so many layers on these chips. And so the number of manufacturing steps is going up like crazy, and yet the sensitivity is getting even more sensitive while there's more steps. And every step you take, you have the opportunity to introduce variability. And that variability then leads to yield issues, etcetera. So there's a huge opportunity now to use physics inspired AI solutions to help understand and manage all that variability, thus leading to higher yields, primarily higher throughputs, faster time to market.
[00:06:25] Speaker B: Got it. John, you mentioned that you chose semiconductor manufacturing to start with, given the tailwinds in the sector, right, from a macro perspective, and also from a geopolitical perspective, but at the same time, it's a very challenging problem from a manufacturing perspective, one of the most challenging ones within the industrial sector. So how did you weigh the pros and cons as an early stage company of then choosing this as you're starting, starting industry?
[00:07:04] Speaker C: Fantastic question. Well, I wouldn't say there's a magic answer to this. I think we as a company, we prided ourselves in having the technical skills to be top of the market, right? Like our ability to do a modeling, our ability to write software, and even to some extent our ability to understand what's going on in the physics of what's there, we believe to be sort of a superior combination. So we were sort of biased towards, we thought we were more likely to be competitively successful in the more challenging areas because we could distinguish ourselves from a technical perspective. So I think that was some of why we did it, and that was why we biased towards, let's say, that industry versus something that is less risky, food and beverage, let's say, for example. So that was part of it. And part of it was that because it's such a scary industry, there really wasn't competition. Even today, if you look to see who are the leading AI companies in semiconductor process control, there's really no other names. So it was from a positioning perspective, as I think you probably understand, positioning is very critical. Why do people, you want to create a niche area where when people think about a problem to be solved, you're the first name that comes to their head. If there's a hundred other sort of startups all clamoring for that space, it's very noisy, it's hard to stand out. But if you look at semiconductor, you get to the point where we're really the top name that comes up. If you're thinking about what's the startup doing cutting edge AI for semiconductor process, because there's really only one. Right. It's Tignus. Right. So that has an advantage. But I can say we're still in the startup stage. So it's yet to be proven. It was a good choice because all the other things are still true. This is an incredibly conservative market, and we're trying to change something very core to their main sort of business, and they're very sensitive about sharing their data and all that kind of stuff. So we took the bet, we made the risk. We took the risk, and we will see now if we can overcome the conservative aspects of the business, which was the downsides.
[00:09:15] Speaker B: Yeah, I'm sure. John, in terms of some of the early success on the use cases that you mentioned, for example, improving the time to market, improving the yield, cost optimization, can you share some of the early success stories that you're seeing with your clients?
[00:09:36] Speaker C: Yeah, it's kind of a mix. So we, there's really two sub markets that we really, three of most sub markets we work with. There's the manufacturers themselves. These are names like intel and Texas Instruments and people, Broadcom, people you might recognize who are members of that market. But there's also a very big market. Actually, one of the most profitable market submarkets are the equipment makers. These are the people who make the manufacturing tools, and they're responsible for a lot of the process control. And this is Toki, electron and Lam research and applied materials and guys like that. So we started out actually with the latter group saying, hey, we can sell our technology and get it licensed as an OEM license by these equipment makers. That's a lot easier from a sales perspective for us because they're going to go do the sales and we'll just provide the technology. And we have been somewhat successful on that front. We're getting one of the largest big equipment companies to license our autonomous controller technology, or AI autonomous controller technology for their next generation tools. So obviously this is a big deal.
They're making a bet on our software as the brain for these next generation tools, which can cost tens of millions of dollars per tool. Now, the use case there was, we enabled them to control a process which was not possible to control before because it was just too complicated. There are too many variables that need to be controlled, too many interdependencies. If you tried to model it from a physics perspective, it took hours and hours per wafer. They needed to be processing 200 wafers per hour. And so we were enablement. We enabled this new product, which hopefully will enable them to sell a billion dollars of new equipment. So that's where we started. But that was a hard business. It's hard to convince these large multibillion dollar companies to make a billion dollar bet on a, on a young startup company. And so we broadened from that to focusing, saying, okay, well, we'll still support those customers, but we don't rely necessarily on them for continued growth. Instead, we're focused on sort of packaging our software so that we can sell it to the manufacturers so that they can make their existing tool fleets more productive, et cetera. And that's kind of a new push that we're making right now, is kind of moving out into this space.
Where are the use cases there? The use cases really split into, and we'll talk more about our, and can talk more about our two products in the future. But they are either. It's really about really getting all the value out of all the sensor data that's coming off of these tools. And so there are hundreds of process tools in a semiconductor fab and a midsize semiconductor fab, possibly thousands, and a, in a large leading edge fab. Each one of those tools is generating off, has sometimes anywhere from ten to 500 sensors in it. All that sensor data today is most of it's ignored. Most of it's not used or leveraged. What we do is we provide software that can ingest all of that data.
I think of it as our solutions are really like. They're like, but there's data. So we help them store all that data in a way that was more efficient than ever before possible, and to make it more rapidly accessible at scale. Algorithms. We provide algorithms that allow them to find the needles in the haystack to explain why variability is occurring in their world. For example, interfaces.
Now you have all this massive data. How do you actually do useful things with it? We've created a low code programming environment for scientists and engineers to leverage all that data and AI together without being trained AI people, and of course, large language models. We have an agent that they can talk to to kind of enable more people to do AI analytics across all that data within the fab environment. And then finally, autonomy. We actually have packaged some of that technology that we deliver to these top tier equipment makers, made it accessible to these fabs, where we have software that will actually control their tool for them, removing the need for them to sort of manually tweak it in order to keep it within tolerances.
[00:14:07] Speaker B: Got it, got it. And John, it has been seven years now since the company was founded. How is the overall vision and strategy changed since then? I mean you talked about how to start with semiconductors was the focus, but now it's more, much more broader than that within the industrial automation space. So that's one. What else?
[00:14:32] Speaker C: Well actually it's kind of been almost like a bell. Right? So we started out not focused on semiconductor. We did work and we were actually deployed in an oil refinery. Right. We in oil fields and we were working with food and beverage companies, we're working with pharmaceutical companies. And so we kind of started broadly and then we realized that it was very hard to be, there's just lots of noise. Industrials wanted to talk to someone who understood their language and because we were going to everybody, we didn't really understand anybody's language well. And so we made two decisions. One was to focus on one particular vertical semiconductor, the other decision was really to focus on process control, because what we found was that there are like possibly thousands of AI companies saying they do industrial AI, but nobody was saying, almost nobody said they could actually control a tool. Like they could provide autopilot for a tool. That was just too hard of a technical problem for 99% of the startup companies out there in this space. And so we decided to focus on those two things again, focus on the technical problem that everyone else thought was just too hard, and then focus on the technical industry that most people, because they thought it was too hard. And again, we were gluttons for sort of punishment, I guess. But we felt that we were more likely to be successful at this and so we made that decision to focus, I would say now we've reached a point where we deeply understand the semi place we've built strong relationships, we have many kind of customer conversations. We know the people in this market and those accounts are either onboard or they're working slowly through the process. Right. And so now we're starting to look back up again and say, all right, while we wait for those accounts to slowly work their way through their slow sales cycles, who else can we help? And so we're starting out by looking at peripheral companies. So vendors to the semiconductor ecosystem. So semiconductors consume a massive amount of soaked manufacturing consumes massive amount of chemicals, right? So there's a whole ecosystem of chemical companies out there that supply chemicals to the semiconductor industry. And we're talking to a bunch of them about helping optimize their manufacturing process. So there's really nothing specific about our technology to semiconductor manufacturing. It was really more of a go to market focus. And so now we're starting to relax that go to market focus a bit more, but still trying to focus, keep things tight. So we're primarily talking to companies who also go to Semicon. Right, which is the main semiconductor conference. And then for the most part, again, trying to stay local in the US because having international operations is very taxing for a very early stage company. But we're also starting to tiptoe a little bit there as well, into other countries as well.
[00:17:27] Speaker B: So John, what are some of the things that you have done differently as a third time founder?
[00:17:33] Speaker C: Yeah, that's a great question. Well, for me it's been a journey. Every time I've kind of gone out and started a company, I've had a different set of objectives.
Initially it was just start a company. I didn't really know what I was doing and just got thrown into the fire and learned a few things along the way. With the third company, I had seen it a lot. I'd kind of been through several startups and also spent some time acquiring startups when I was at VMware. So I saw a little bit from the other side. And so I had very specific goals. One was that I wanted this to be kind of a venture funded endeavor. I wanted to go out, raise a bunch of venture money, use that venture money to do something big, I guess was the plan.
So that was something I hadn't done before in the past. I had went and built the product and then got some users, and then maybe tried to raise a little money to get going. This was, know what, I'm just going to raise the money up front, go start the company, build it. It was a very specific strategy, and I wouldn't necessarily recommend it to everybody, or I'm not sure I'd even do it again. But that was the plan was I wanted to see if I could do this where the risks in some sense was managed by the fundraise happening upfront, hiring a bunch of people upfront, building a product, taking it to market, doing something big, etcetera. And that's not something I even would have known how to do before. Right. I didn't have the confidence that I could raise money before, but in this thing it was also something I was able to do based on my credibility from previous success. So I'd had a couple of successes in the past. I'd been CTo of Vmware. This gave me a network of people, of network, of trust that folks were willing to put some venture money behind me to allow me to invest. So there were sort of the things that we did explicitly. I think tackling a really big problem was something that was not like before. I was trying to sort of solve small, targeted stuff. This was, I'm going to go change the world of industrial automation. So I think that the two previous successes gave Matt and I the confidence to tackle something really big. I'm always learning. So I have my list of things I'm going to do different next time. And if there's time we could discuss that, I could say, here's my new set of learnings of what I want to do. I want to be trying something different to learn new things. And so that was, I guess we intentionally went into industry that we didn't understand, right. We didn't know anything about semi, about semiconductor management or industrial, but that was intentional because we wanted to learn, right. We want to do something different.
[00:20:04] Speaker B: And John, from that list of things that you would do differently the next time, any insights that you can share with, with people who are looking to launch a new business venture, any wisdom for them?
[00:20:16] Speaker C: Absolutely. I'm happy to share. By the way, I enjoy advising startup companies. I try to reserve some time every week to do some free advising. But absolutely the next time. The number one thing I would do is basically have the first buyer lined up before I actually start building the product.
I think this is a critical insight. When we built tickness, it was a bit of, we know that there has to be, it's like, we know that the world's going to need a transformation. We're going to be there to provide that transformation. And it's a broad transformation across the whole industry, et cetera. So we'll just get going and we'll figure out what our incremental use cases were going to be. I don't recommend that strategy. I mean, it is a go big strategy and sometimes you can go very big doing it, but it also creates a lot of risk. I think it's much better to have the first, not just one, because you should be sure there's more than one possible customer here. But before you really launch and start building your product, kind of have already a set of customer advisors who for which what job to be done? Like what's the job that you're solving the job to be done? How much you know that they have the money to pay, they have the authority to pay. It's a classic sort of sales qualification, right? You know that there's going to be a budget that you've talked to people who have authority to buy this, but their timing and their pain is pretty bad and that they're doing this like figure all that stuff out before you start building anything. I think is, is my number one piece of advice for the next round. And that's probably independent of like, whether you decide to bootstrap or venture fund or whatever.
[00:22:02] Speaker B: Now, John, I would like to also talk a little bit about the advancements that we are seeing in AI and ML these days, right. Especially with generative AI, the pace of progress, especially what we've witnessed since the launch of Chad GPT, is unprecedented. What does that mean for tickness? And how are you making sure that with such a rapidly changing world, competitive edge with your clients?
[00:22:30] Speaker C: Yeah, well, this is very challenging because on one hand, semiconductor marketing, manufacturing is a very conservative industry, right? And so the kind of errors that generative AI make are exactly the kind of things that alienate a very conservative market, right? Semi manufacturing is all about like precision and tolerance and large generative AI agent give you bad advice about how to respond to an excursion in your tool is, could kill your product almost instantly. So I think it's sort of an open question. So anyway, it's not like you just drop it in. It works kind of thing with generative AI in this space.
But I think as a startup folks, we're also very aware that the last thing we want is to spend seven years becoming the premier AI company in this space, only to kind of lose our position because some young startup goes all in on large language models and that turns out to be the disruptor. Right? So we make sure that we're not there by investing some amount of our engineering cycles and sort of staying up to date on what's going on with generative AI, but also implementing it and deploying it. So we've actually deployed generative AI solutions to one of our customers, one of our early customers already, right, which allows. So I guess the point is that we are staying on implementing with along with the market. But the reality of how they're actually going to successfully be deployed I think is still an open question. I think that one of my observations is that the actual large language model part is becoming commodity model. We're just going to use something off the shelf that's the same for almost everybody and that there are multiple models out there. It's basically a commodity and everybody has them and everyone who wants to compete with prints? Well, the difference really is going to be think at the a, do you have the data, do you have the data sources to feed into this? And in semiconductors, very specific. And two is how do you have the infrastructure to tie those generative AI's into the broader available non textual. How do you really integrate all that massive sensor data coming off of these tools with these generative AI solutions and effective means? You're going to need an AI. You're going to need an AI ready data infrastructure to support these generative AI capabilities. I think that's the way we look at it right now, is that there's really no provider besides us in this space. And we think of the generative AI models really as just being interfaces. They consume that data. They're commoditized interfaces that will consume that data and support specific task based use cases. Now, will we provide those interfaces for some tasks? Yes. Will other customers? Will other cooperation, maybe the large language, generative AI type things, and they'll consume our data platform and work together. The customer. I think it's a lot of open questions on that space. I could go on probably for a whole hour talking about them.
[00:25:10] Speaker B: Final question, John, with all the advancements and the way things are changing in the AI, what do you envision as the future of AI based process control?
[00:25:21] Speaker C: Yeah, absolutely. Well, I think the number one thing is that it's more data driven than what happens today in the fab, right? A lot of what happens today in the fab is run by human intuition, right? You have guys with PhDs wandering around looking at small amount subsets of data that are not complete, that don't account for all the factors, and then just using their best trained intuition about what to do to resolve an issue. I think that the primary thing that AI process control is going to do is to really move from intuition driven decisions to more data driven decisions. The AI solutions, such as those provided by Tegnas, enable you to look at a much more of the data and much deeper into the data to understand much more the interacting factors, not just the surface level factors. And that in turn allows you to, with more data driven decisions, you have more predictable outcomes, which is really what yield is about, is you want the same thing to happen every time.
And then of course with automation, because you can automate these decisions, you're taking humans out of the loop. That also creates more predictability and less latency and that kind of stuff. And so I think that's really the key of the future, and particularly if you look at the need to onshore more in a desired onshore more semiconductor manufacturing within high cost regions like the United States, there's really no choice but to automate more.
We need to have more productivity, right. We have to be more productive in order to be condensed. So how do you become more productive? That you take some human processes, you automate them two is you actually make the equipment more productive. So a little known fact is that these tools are there number of productive hours of these manufacturing tools that can be actually surprisingly low. Right. Once you take into account when something goes wrong, they have to turn it off. When something, maybe it breaks, maybe it's breaking the product that's coming out of it. And so in order to do that, they do lots of preventive maintenance. In order to do that, they have to take the tool down. Right. Preventive maintenance sometimes creates its own problem. Anyway, the point is that we have to make the tools more productive. We have to make the people more productive. And that's really the future of AI. Process control is this sort of massive increases in productivity through better tool productivity, through better human productivity. And it's really just about leveraging the data in a more extensive way.
[00:27:56] Speaker B: All right, thank you so much, John.
Pleasure to have you here with us today and thank you for sharing your journey of building table.
[00:28:06] Speaker C: Yeah, well, thanks for having me. It was an enjoyable conversation.
[00:28:09] Speaker B: Sounds good. Thank you, John.
[00:28:11] Speaker C: Thank you.
[00:28:18] Speaker A: Thanks for listening to Ina Insights. Please visit Ina AI for more podcasts, publications and events on developments shaping the industrial and industrial technology sector.
