Business Breakdowns
· Brett Larson
And it really comes down to the way Keyence goes to market and how they focus. They generally focus just on the middle to lower tiers of customers and then they're really trying to develop more standardized products, are going after the very high frequency applications and it's a very process oriented sale. So they'll hire more college graduates compared to Cognex and then they'll train them with the products, put them out in the field. And it's more of a scripted sales process where they're keeping track of more of those activity like KPIs in the CRM. So it's how many calls are you making, how many shop visits, how many demos. It's really good coverage and really relentless. If you go down the rabbit hole, it's become a meme in the community. There's some pretty funny memes out there. If you're going to download a product spec sheet up the website, you should use your buddy's email and phone number, not your own. You'll never hear the end of it. That's the two main players. Maybe I'll briefly mention a couple other competitors while we're here. The other bucket I would say is China and these are primarily for manufacturers in China. Hik Robotics, which is division of Hikvision, is probably the third largest player in the industry overall and they've grown quite quickly driven by that domestic market with the domestic manufacturers. And they're probably about half the size of Cognex in terms of sales. There's a number of other smaller Chinese players as well. And then outside of there, there's more legacy players. There's within Teledyne, there's DALSA and Point Grey, there's SICK and Basler in Germany, Datalogic, Matrox was a company acquired by Zebra. And there's even companies like MVTec that just sell the software and then you can go get your own hardware. So that's kind of a lay of the land in terms of the industry structure.
· Matt Reustle, Brett Larson
[00:00:00] S0: I did read quite a bit about Keyence and their sales process and how effective of a sales organization they built. But in the sale for someone like Cognex, where it sounds like it's much more technical, built to spec, perhaps advanced and not so much off the shelf, are they selling to an Amazon who then works to integrate it into the other hardware and equipment that they're using, or are they selling into the equipment manufacturers, whoever builds the conveyor belt? Who is the customer for Cognex? [00:00:00] S1: The answer is yes. If you think about how they're going to market, 70% of sales are direct. And that would be direct either to the factory floor where Cognex and the automation engineers at the customer are working to implement the system. But direct also includes selling to a machine builder or an OEM who would then integrate it into the machine and then they take that machine to the end factory floor. And then the remaining 30%, about half of that is going to be through systems integrators and that's primarily logistics at this point. And then the other half of that 30% would be distribution. And that's primarily just for end markets like Cambodia or something, where they just don't have any presence. That's kind of how I would separate the customer.
· Brett Larson, Matt Reustle
[00:00:00] S1: The answer is yes. If you think about how they're going to market, 70% of sales are direct. And that would be direct either to the factory floor where Cognex and the automation engineers at the customer are working to implement the system. But direct also includes selling to a machine builder or an OEM who would then integrate it into the machine and then they take that machine to the end factory floor. And then the remaining 30%, about half of that is going to be through systems integrators and that's primarily logistics at this point. And then the other half of that 30% would be distribution. And that's primarily just for end markets like Cambodia or something, where they just don't have any presence. That's kind of how I would separate the customer. [00:00:00] S0: That makes sense. It's an interesting dynamic within the overall industry structure conversation. Before we move on too far, I do want to get into a bit of the history and what came across through the work that I was doing is that there is technical expertise here and there's a focus on that technical expertise. So can you bring us back to the beginning and the origin story here, when they got into the market and some of the evolution over time?
· Brett Larson
[00:00:00] S1: It's really important to understand for Cognex the DNA of them stacking S curves essentially over 40 to 50 years. The company was founded in 1981 and they came to market with what was the world's first industrial optical character recognition system. So again, reading numbers and letters. And at the time the system looked more like a big video camera with external processing. And the first application was actually to read Serial numbers on semiconductor wafers for IBM. That was the DataMan product family that still exists. From there they got a call from Johnson & Johnson to do all optical character recognition for labels. They did that application. But then J&J asked them, hey, can you do some of these other at the time, what would have been novel applications like verifying the caps are on the bottles and the labels were present and the like. So that's when Cognex added more of these inspection type of applications. The first 20 years or so of the business was pioneering machine vision for all these different applications. But it was very much for semiconductors and the electronics capital equipment industries primarily. It was as high as 80% of sales going into the dot com bubble were to those industries. And that went to 54% following the hard way. Around 2000 was when we saw smart cameras really come onto the scene. So where you embedded the processing directly into the camera. Smaller footprint, more ruggedized now that enabled the product to be ready for the factory floor as we more often think about it. So we saw machine vision get adopted in more heavy manufacturing industries like automotive, consumer electronics, food and beverage packaging. So by the end of 2010 that semiconductor and electronics capital equipment was down to 15% of sales as they really brought in the end markets that they could serve all those new s curves. Around 2010 was when Cognex came out with their first ID product for reading barcodes. It was primarily aimed at displacing laser based scanners in logistics facilities because it could do it faster and then with much higher read rates. If you go back in 2010, they were talking about they'd be happy if one day, long term target $75 million of sales from this line of business. And then they worked very closely with Amazon to develop the technology.
· Brett Larson, Matt Reustle
And in 2021, the peak, it was about $300 million in sales and 30% of Cognex's overall business. So it turned out to be just an enormous new S curve for them. At present we're really in the early innings of the next tech evolution for the industry. And it began in 2017 when Cognex acquired ViDi. And then in 2019 they acquired a company called SUALAB. And those were both essentially IP and acqui-hires in the field of deep learning or the application of AI into machine vision. ViDi's co founder and CTO is Cognex's current VP of AI technology. So after those acquisitions there was a big development effort and they came to market with these new Deep learning and Edge learning products. And the way I conceptualize this is moving from rules based programming of vision systems to teaching by example, which has large implications. And I think we're probably in the first inning of this next chapter behind the scenes of a long cyclical down cycle. [00:00:00] S0: Could you paint that picture in terms of an example rules based? Let's say a package is over 100 pounds with these dimensions, send it in the left lane. If it's under that, send it through the middle. If it's super small, send it to the right. And it's a pretty standard tree of logic and decisioning. Where would this new example or this new training come into play? Do you have any real world applications that would paint the picture?
· Brett Larson
[00:00:00] S1: There's deep learning and edge learning, as I mentioned, deep learning first. So traditional rules based vision is great for a lot of things, but not for things very subtle or nuanced tasks with a great deal of variation. Those sorts of tasks, you still usually have a human performing them because frankly, it's just hard to quantify them to then program them. Maybe an example of that would be inspecting a phone case for very subtle defects. There's obviously a wide variety of colors and there's just an endless list of potential defects. So scratches, dents, blemishes in the paint, and they're very small, very nuanced. And then there's also usually a scale of what's acceptable and what's not, where a human would be able to look at it and pretty quickly discern that. But it's very hard to program a machine to be able to do that, especially at obviously high speeds. But with deep learning, instead of programming it, you teach it with very large data sets and very large image sets. These are good phone images. These are defects. These are acceptable defects. And the machine learns what is acceptable and what is not and is able to accomplish that task, just to quantify this a little bit. So there's currently still 30 million people in the world doing visual inspection. And that's something that humans actually aren't great at. They can do it quickly, but they get fatigued and they miss things. That's a big opportunity. Another example would be something like deboning chickens, where every chicken's different, it comes down the line a different orientation, the wings are somewhere different. It'd be impossible to quantify and program a robot to be able to grab the right places. But with deep learning, that's an application that they can now do deep learning. The implication would be potentially new applications for machine vision. Edge learning is the other end of the spectrum and it's probably more financially tangible for the investor community right now. These products, they come pre-programmed for specific applications but then they're trained with just as few as five to 10 images. So it's essentially very easy to deploy relative to traditional rules-based vision. You can usually get an application up and running in a few hours. So the major implication here is that Cognex can now sell to less sophisticated customers.