This week’s podcast is about SenseTime, which is a leader in computer vision in China / Asia. It’s a fascinating company for looking at the evolving business models of AI software. The two big questions are:
- What are the unit economics of large AI companies?
- What types of scale advantages and network effects do large AI companies?
Here is my new book:
- Moats and Marathons (Part 1): How to Build and Measure Competitive Advantage in Digital Businesses Kindle Edition
I break the network effects question into three sub-questions:
- Do existing AI models get better the more they are used?
- How much? Does it flatline? Does it keep increasing?
- Do existing AI models make each other smarter and better?
- Do existing AI models and the data they collect make additional AI models smarter? Do they make them faster and cheaper to train?
My working for conclusions are:
- This is mostly an integrated software bundle going for global scale.
- It’s a learning platform with a new type of network effect.
- It is building an innovation platform, with another network effect.
- SenseTime and an Introduction to AI Software Economics (1 of 3) (Asia Tech Strategy – Daily Article)
- 3 Lessons in China AI/ML from Artefact (Data Consultants and Digital Marketers)
- Adobe Inc. and the Power of Old School Software Economics (Asia Tech Strategy – Podcast 81)
From the Concept Library, concepts for this article are:
- Learning Platforms
- Integrated Bundles
- Artificial Intelligence
- Computer Vision
From the Company Library, companies for this article are:
Welcome, welcome everybody. My name is Jeff Towson and this is Tech Strategy. And the topic for today, two big questions about SenseTime and AI software business models. Now this is the, you know, this is the computer vision AI giant of China. It’s one of, you know, the new cadre of really large AI companies that have been launched in the last five to six years in China. one of the first ones to go public. There’s a little bit others, but really an important company and has struggled to go public for quite a few years now. It was supposed to be in Hong Kong and then it got delayed a couple of times. Then it was supposed to go into the US and then at the last minute, it got placed on the US Treasury Department’s list of, I think the exact title is Chinese Military Industrial Complex Companies. which has various implications including that I don’t believe US citizens are legally allowed to invest. I’m 90% sure about that. But they did go public and in Hong Kong and what we saw was a first clear look at how this new type of digital business really functions and that’s what got my interest is like. This is a new animal. and it’s one of the large ones and we’re getting a really good look at how it works as a business model, how to think about it in terms of strategy. Yeah and it’s huge. I mean this is arguably, it’s definitely the largest computer vision company in Asia. Outside of that you could probably find some stuff in the US but yeah let’s say it’s in the top two or three globally. So that’s the topic for today. The two big questions, I’ll go into that shortly. Some housekeeping stuff. My book is out, it has been updated. There was some issues with some fuzzy graphics. Those have all, in theory, been fixed. I’m talking with Amazon. Apparently there was an issue that it was updated and fixed, but Amazon doesn’t automatically push new versions of books to everybody because people mark them up and things and they don’t wanna just delete that. So. There has to sort of be an approval process that look, just push it into everyone’s iPad so that the graphics are nice and clear. That has been done in theory, but I’m still checking. Anyways, hopefully this is all behind us and the new book is coming out in a couple of weeks and the graphics are beautiful. Well, I think they’re beautiful, but I really like graphics. Okay, so there’s that standard disclaimer, nothing in this podcast or in my writing or the website is investment advice. The numbers and information. By me and any guess may be incorrect. The views and opinions expressed may no longer be relevant or accurate. Overall, investing is risky. Do your own research. This is not investment advice. And with that, let’s get into SenseTime. So as always, there are two concepts for today. Concept number one, learning platform. Concept number two, integrated bundles. Now these are in the concept library. And you know, as always, This is how I try and get smarter every day. Every time I look at a company, I’m trying to get smarter about a company because that’s where the rubber meets the road is understanding a company. But then also, so you sort of build out your list of companies you’re pretty good at. And then additionally to target the big concepts that matter. And usually I sort of, every podcast, I try and touch both of those bases. So these are, I mean, these are kind of, I think important ideas and… Integrated bundles is easier to understand. I’ve written a lot about bundling. I love bundling. Everyone wants network effects when they talk about digital companies. I like switching costs and I like bundling. Because when you have something that’s a digital good, a service, an internet service, software, eBooks, it’s really easy to tie them together in bundles. In physical products, you go to the supermarket and you can get shampoo or you can get conditioner or you can get shampoo plush conditioner. That’s a bundle. But it’s kind of hard, you can’t bundle 50 things together in the shampoo aisle, but you can do that in digital. It’s one of the unique strengths of digital economics is the ability to bundle extensively, easily, and also it’s almost for free, because usually when you start making copies of digital goods, the marginal production cost is zero. So, you know, that’s Spotify, that’s Apple Music, you know, that’s Netflix, these are digital bundles. So one, I love bundling. Two, I like integrated bundles. This is where, okay, I watch, I don’t know, Narcos, and then I watch, I don’t know, what was I watching? I was watching Hawkeye, which was not good. Okay, they’re a bundle, but they don’t integrate together. But when you look at something like Adobe, well, they have Premiere Pro, they have Adobe Reader, they have After Effects. You know, they have this suite of digital tools for creators, but they all work together. If you have Premiere Pro, you can make videos and stuff, but if you also have After Effects, which is a sort of editing program, you change the color grading stuff. If you have both, they both add to each other and the functions can integrate. Now Microsoft Word is like that. Microsoft Office is a bundle, but it’s also integrated. take things from Word and they sort of tie directly into Excel and PowerPoint, not extensively, but there are nice ties there. The more you can do that, the more powerful it is. So I like integrated bundles. And I think that’s what SenseTime is building, a massive integrated bundle of AI software, which is, let’s say AI models, which is different than traditional software, which is pretty much everything I just mentioned. So it’s an integrated bundle, but it’s in a new type of thing. So that’s one. And the other is a learning platform, which is one of my five platform types, marketplace payments, things like that. This is the one I’m struggling with. This is like the one that doesn’t quite fit yet. Where I look at platforms, I try and define user groups. And then I try and define the type of interaction that’s happening between them. That is the purpose of the business. A marketplace is about doing transactions. That’s the interaction that matters. There’s other stuff you can leave reviews and all sorts of things, but the interaction at the core is a transaction, so it’s a marketplace platform between buyers and sellers. Okay, and then we can have audience builders and we can have payment platforms. And that’s where those words come from. But there are certain types of platform business models. Let’s call them network-based business models. where the primary thing is happening is the more interactions are happening, the smarter it gets. The better the intelligence is, the better the answers become, the better the services. That’s kind of Google. Like the more you search on Google, the better the answers get. The more that everyone searches, the better they get, especially for long tail searching. And that is a two-sided platform business model. You have basically web pages and SEO on one side and then you have people doing searches on another. So it kind of fits that model when you talk about a learning platform and you can look at something like Waze, Google Maps, Quora, Juhu. And I think what SenseTime is moving towards is a learning platform, but they’re not there yet. But this, I don’t think this fits terribly well. I think there’s some problems. But so I’m trying to sort of figure, this is like one of my things I’m trying to figure out is how to really nail the business model for these things. And I think I’m about halfway there. So learning platform, put an asterisk by it. I think I’m halfway on this one. Okay, so those are the two ideas I wanna talk about. So the two big questions I have when I look at SenseTime, and I’ll give you some background on the company in a moment. Number one, what are the unit economics of very large AI companies? Where when I say AI companies, I’m talking about a company who is in the business of creating AI software, which is very different than traditional software like Adobe or Microsoft Word. This is AI software, it’s a different animal. Well, what did the unit economics look like when it gets to sort of global scale, which is where this company’s going? We kind of know what it looks like for traditional economics. I don’t think anybody really knows for AI software. And Andreessen Horowitz, the awesome venture capital firm in Silicon Valley, they’ve been writing about this because they’ve invested in a lot of AI companies. And they’re basically saying, look, the economics do not look like traditional software. They don’t. The amount of time and effort it takes to get data, to train models and to keep training them. and to clean the data and to tag it and to keep the model working in an acceptable accuracy rate is expensive. It looks more like a combination of a software company and a services company. There’s a big labor component thus far, but it might change. So that’s one. What is the ultimate unit economics of these large AI software companies? The other question I’m trying to figure out. What are the advantages when an AI software company gets to large scale? What kind of network effects are we seeing? You want a company that when it gets bigger, it doesn’t just get more efficient, which we would call economies of scale on the cost side, it’s cheaper, blah, blah, blah. No, we want one that has demand side economies of scale. When it gets bigger, it actually gets better. Google is better the more that people use it, right? Are we gonna see what types of advantages at scale are we gonna see with AI software companies? And again, I think the answer is we don’t quite know. I’ve got a take on it. I’ll give you my working answer. But I think if we can crack those two questions, then you can start to really project the trajectory of this company and others like it. Keep in mind, when people looked at Amazon, Facebook, and Google, 2008, 2010, 2012, So many people underestimated the trajectory. People used to think these companies were really expensive back then. Oh, they’re overpriced. No, you just underestimated the trajectory. These were gonna go much bigger than everyone thought. Well, that’s the unit economics, that’s the advantages of scale. So that’s kind of what I’m thinking about. Okay, those are my two questions. I’ve put them in the show notes. And the… Ideas for today, Integrated Bundles, Learning Platform are also in the show notes and they’re in the concept library. All right, let’s go into SenseTime. Now, nice thing about SenseTime, this is not a platform business model. This is one user group, the customer. It’s closer to Adobe and Microsoft Word than it is to say Alibaba or Google, right? It’s sort of, it’s a traditional in the sense of business model. But instead of being a traditional business model selling software, it’s selling AI software, which is a different thing. You could call it data technology, but I just use AI software. So it kind of looks like Adobe in that sense. And I did a podcast on Adobe. You can see it in the show notes. What Adobe does is what Netflix does, and what a lot of these purely digital creatures do. They take what is different about digital economics, and they pull that lever as hard as they can. They lean into doing all the things that you can only do with bits and bytes, and not with traditional sort of physical products and services. And I’ve talked about digital economics before, but these are things like, you know, zero marginal production costs. The first one cost you a lot of money, the second copy cost you nothing. Low distribution costs, global scalability, often at very low cost, and usually without any limits on capacity. I mean, you can sell Netflix to the whole planet just as easily, but if you wanna open Zara stores, it’s gonna take you a long time. So these companies tend to lean in as much as possible to the growth equation. I mean, they pull that lever hard. versioning, the ability to bundle, the ability to connect with other types of software and digital goods. Your average business making shoes or whatever, they might have connections and partnerships and collaborations with their suppliers, their manufacturers, their distributors, their retailers. They might have connections with 10 companies. A digital company can have a thousand connections. Software is not limited by, so you can be far more connected and integrated, which lets you do pretty amazing things. So all of those things, you know, that’s the digital, let’s call it advantages. You know, Adobe pulls that lever hard, so does Netflix, so do a lot of these companies that aren’t platforms. Okay, definitely it looks like Alright, so here’s a little background on the company. They call themselves quote unquote industry leading full stack AI capabilities. Okay, industry leading, why does that matter? If you’re gonna create a simple little AI that helps you do, I don’t know, deep fakes and you can put Luke Skywalker’s face on a video, which some peoples have done. online is kind of funny. Okay, that’s just fun. You can do a lot of sort of AI stuff at various levels of quality. But if you are going to industry quality, if you are going to sort of, you can put this into a hospital. You can put this into a truck going down the highway at 70 miles an hour. You know, that’s sort of industry leading AI capabilities. That’s a much higher standard. And you know, SenseTime is not selling B2C. They are selling to other businesses that you can put this in your business. So they’re more of an enterprise play, but it raises the bar dramatically in terms of, you know, how hard is it to build these AI models to do these types of activities? The other quote in there was full stack AI capabilities. They’re doing the software, the models, we can call them AI models, the data, the training data, the training tools. They’re moving down further into the stack, into servers, massive data centers that they’re building that are specialized for computer vision, not just AI, computer vision. And then they’re going all the way down to chips, which are again specialized not just to AI, but to computer vision. So they’re specializing and they’re going after the full tech stack for what they do. Okay, so industry leading, and they talk about AI chips. And they’re also doing sensors, which obviously matters, cameras, I don’t know, thermal sensors, anything that’s gonna pull in the data that then their algorithms look at and tell you, oh, that’s Bob walking down the street. And it can tell who Bob is by looking at his gait. It doesn’t have to see the face anymore, which is pretty spooky. Okay, so that’s kind of their basic thing. Companies founded, came out of Hong Kong, university, well, Chinese University of Hong Kong. Tang Xiao Ao, which is now he’s a famous professor, Department of Information Engineering. I mean, basically this was a academic project in object recognition, facial recognition. And for those of you who are sort of familiar with AI out of China, he came out of Microsoft Research Asia. Microsoft has these centers they’ve opened up all over the world, Microsoft Research, I don’t know. I think there’s Mumbai, Microsoft Reacher, like Singapore. Well, there’s one in Asia, which is in Beijing, set up in 1999, actually opened by Kai-Fu Li, who is the AI guru now. He’s the one who opened it. But if you hear of someone who’s really good at AI in China, look at their bio. There’s a good chance they were working at Microsoft Research Asia. The head of AI at JD, Tao Mei, who I’ve interviewed, you can find the podcast there. He was a researcher there. competitor Megvi was founded by alumni of there. Anyway, so Tang Xiaoh, he was from there as well. He spent a couple of years. He did China degree, US degree, came back, worked at Microsoft Research Asia. Then he goes down to Hong Kong, teaches there professors. He’s doing algorithms and things for object detection, optical character recognition, video analysis, medical image. I mean, he’s figuring out how to do this 2013, 2014. He’s… Founds this company with a lot of people from his department. And you know, from out of the gate, and this was about the period when the wave of AI companies got founded in China. Some of them were like, we’re going after computer vision. That’s MegVee. That’s SenseTime. Others were going after natural language processing and other sort of these verticals of AI. They went after that, founded the company. And in about four years, they took it from an academic project to one of the top AI companies in the world by, you know, unicorn rankings and all that stuff. They were on the shortlist and they were winning awards for, you know, we can recognize things with a higher accuracy than anyone. They were winning these championships for all this stuff. And, you know, it just turns out computer vision. I mean, it just turns out that AI is good at some things and it’s not awesome at other things. It’s, you know, like if you want to use AI to translate speech, like if you give a talk and then you have an automatic AI translating it into Chinese or Spanish or whatever, it doesn’t do a very good job because it sort of makes the comparisons and it puts up, but it doesn’t understand jokes. I mean, AI doesn’t understand anything. It’s just looking for correlation. So it has no understanding of the actual content of the speech. So, you know, Top tier speakers will often say, if you’re gonna use translation, it needs to be a human translator, it can’t be AI. It doesn’t understand the jokes, it doesn’t put in the pauses, it doesn’t do it very well. It’s struggling, it has been for a long time. Well, it turns out AI is incredibly good at seeing things. It’s really good at it. You can put, photographs happened very quickly. You know, it can… identify people in photos, it can identify cars and trees and all that sort of stuff, very easily. But it was really video feeds that were the big breakthrough because then you can put a camera on a street and things come on and cars and trucks and traffic and people and it can understand everything. It can predict, it can see the traffic accident, it can recognize the dog that went across, it can reroute traffic in real time because it understands the congestion. It can identify everyone on the street. It can identify every car. So if you put cameras on all the streets of town and you plug it into one central AI with an enormous amount of computer power, the AI can basically see the whole city in real time, like God’s eye, and it can see the whole city and what’s going on, and you can sit in the control booth and it can see it and then flag things for the operators. And not only can it see things in real time, it can go back in time, because everything that was recorded, they don’t keep it forever. It can look, it’s like it can see into the past. It can overlay accidents on a intersection today, yesterday, last year. You know, it can kind of see into time, which is strange to think about. So I’ve asked some of these AI companies, like what was the big breakthrough that really helped with computer vision? And they said it was the live video feed from cameras that are on the streets, that are doing security for factories. You know, you don’t need the security guards, you just put the cameras up. That are on the lapels of police officers, they all have these little cameras now. And the AI in the central hub can see it all and understand it. It can predict if you’re lying. If you talk into a camera, and you give it a baseline, so let’s say you talk into a camera for two to three minutes, and then you tell a lie, it can spot the lie. Because when you lie, your blood pressure goes up. You can hold your face pretty fixed, but there are micro movements in your face, in your blood pressure, in your eyes, and the AI can see all of it. And it can predict pretty much. I met a guy who was an AI founder out of Moscow, and he was pitching a company, and his thing was… to detect lying and he he had a degree in detecting lying from a university in moscow which apparently they had this so anyways okay so they launch computer visions very good and it’s also easily commercializable natural language processing automatic translators it wasn’t totally clear what the market for this was but there was a very clear market for this and They did a lot of contracts with local government, police, a lot of police contracts. They do contracts with lots of enterprises, companies, cameras outside your factory, cameras in the mall to study traffic patterns, cameras within the store, to see what people are looking at as they walk around and the patterns and understand, I mean, you can use this thing everywhere. Anyway, so they sold a lot of contracts, policing, security, city management in China. Lots of companies studying consumer behavior. And then the other place, and those were sort of their two big businesses, government and enterprise commercial. And the third one was they started to put it into smartphones. So, you know, smartphone makers, they have all these apps. Well, you can put these algorithms right into hundreds of apps so that, you know, whatever app these people are building or putting on their phone, suddenly it has AI capabilities. So that was their third sort of thing is smart devices. It’s in, you know, hundreds of millions of apps in China have sense times algorithms in them. Okay, so today they got about 11%, 12% of the computer vision market for China, but you have to pair those down by AI, types of AI. Within computer vision, they’re number one at 11%, but these are early days. So they’re in the lead, but you know, it’s gonna change. And you know. Now Professor Tang is worth $3.4, $3.5 billion as of last week. So that’s kind of the basics for how to think about it. And let me switch from there into, okay, their business model, which is basically the mass production of AI models. And this gets us to the first question for today, which is what are the unit economics of this business model? And it’s been around for a couple years, so we have some history. And I mean, the financials, let’s say 2020, total year revenue 3.4 billion rem MB. So I mean, they’re in the $700 million range for revenue, which is good, but I mean, this is not a huge company yet. Their cost of sales, about 1 billion. So out of 3.5 billion in revenue. one billion’s cost of sales, so you get 2.5 billion, 70% basically gross profit. Okay, that’s nice. I mean, this is software type gross profits. Their revenue, I mean, I’ll go back a couple years, 2019, 3.0 billion, so that’s down from 3.4, 3.5, but 2020 obviously was a bit of a stranger. If you go to 2018 is 1.8 billion, so 1.8 billion to 3 billion to 3.4. I mean, it’s growing, you know, and basically as it’s been growing, the gross margin has been expanding as well. Gross margin used to be about 65%. Now it’s up to more closer to 70%. Okay. And then you start, then you start to look at their main sort of fixed cost and, you know, they’re just flooding money into research and development. So you know, their gross profit for 2020 was 2.4 billion. there are R&D expenses for 2020, also 2.4 billion. So I mean, they’re clearly playing the long game here, right? They’re going for scale, they’re going for technological leadership, and it’s not an uncommon strategy for this. So then you see a big net operating loss. What is kind of interesting is they’re selling expenses were pretty low, about 15% of revenue. That’s not bad for the B2B side. I mean, that’s… That’s pretty interesting actually. Okay, anyway, so at first glance, it kind of looks like the economics that we would see in a software company, traditional software company. Probably the big, and you can kind of see if you’re doing B2B enterprise and they’re selling corporate and government contracts, those are gonna be, you know, you sign the contract upfront, that is generally a good way to bootstrap an organization. The difference is if you’re doing government contracts, I don’t know, for those of you who’ve never done contracts with the government, they always pay late and they take a long time, not because they’re being mean, just because there’s a lot of procedure. So you always see these massive accounts receivables and you can see that with SenseTime. You can see like they don’t have the nice, pretty negative working capital that you would often see in a B2B enterprise company. They’ve got these big AR numbers from government contracts, but that’s kind of just how that game often works. Okay, but that doesn’t really give us any great look at what the unit economics are gonna be. So let me, let’s talk about the business model. Now, okay, it’s a linear business model. So they are producing AI models, right? Algorithms models. They are in the business of producing AI models and then licensing them. to governments and corporations, and they will either just sell the license, or they will sell them a piece of hardware, like a camera, or some server that has the software embedded in it. I mean, they’re doing both there. And I like that model, I think that’s great. I like, when I look at digital companies, I like, I like linear business models that create these types of digital goods, because you get those nice gross margins. I like when the digital goods are highly specialized, but also universally needed. And that’s really Adobe. I mean, if you think about it, Adobe is a super specialist in creator tools, making videos, making photographs, writing documents. I mean, they are very specialized in one thing, but at the same time, it is also universally required. I challenge you to find any company in the world that doesn’t use Adobe PDF. Everyone has it. I challenge you to find any company that doesn’t create content in some form. Photographs, videos, marketing materials, and what do they use? They’re all using Adobe. So it’s like highly specialized, but universally required. I like that. It has digital economics and the ability to… to grow to a global size, if not global, and let’s say regional, it can be very, very big. And then the idea of, look, you can bundle these things together and integrate your functions. I like all four of those factors. And SenseTime has those. I mean, they are a super specialist in computer vision. Computer vision is highly specialized, but it’s also gonna be universally required everywhere. They have the ability to scale, grow, they can bundle, and they are flooding money into R&D to make sure that they are the tech, the unquestioned technology leaders in this specialty technology. I mean, I like all of that. But then we get to sort of the heart of the matter, which is what is the difference between the economics of software and the economics of AI software? Now, their business model, the way to think about it I think about it like a giant factory. They call it SenseCore. They also call it Universal AI Infrastructure, which, SenseCore, it’s a massive technological factory made up of a huge amount of computing infrastructure because to train these models, you need tremendous computing power, and to run them, you need huge computing power. So they have sensors, chips, edge devices, they’re building a massive server facility. You need deep learning platforms that can create these things, which means training data, training frameworks, proprietary training approaches, which is really like their trade seek is within the factory. And then the models that you produce, and they’ve produced about 22,000 models thus far. They’re getting faster at producing the models. They’re getting cheaper at producing the models. And that’s kind of what you see when a factory gets bigger and bigger. They get faster, they get more productive, they get cheaper, and they probably get better quality-wise. So it’s like this giant factory that’s just churning out these models. And it’s kind of the mass production of algorithms. And they keep just coming up with new things to look at. Okay, we have a video feed coming from a factory. Our AI models are good at looking at facial recognition, who’s coming in, who’s going out. Now maybe our AI algorithms are good at spotting cracks in the sidewalk so that we can alert security. Oh, there’s a crack in a machine. It can spot that, it can alert maintenance. We’re seeing the early stages of a crack. And you can use visual cameras, but you can also use sort of X-ray and other types. So you can start to scan the roof with drones and point to that’s the part of the roof that’s weakening. And I can give you a couple examples of these, which are in the IPO prospectus. They had a example project where they… They worked with a high speed railway maintenance company. So one of China’s big railway companies. And they put these special carts that go along the tracks and they just have a camera and the camera just looks up at the overhead electricity power lines that these high speed trains run on. And it just sort of cruises down the track and it can spot when the overhead electricity power line has a problem. Maybe it’s starting to fray, maybe a bird zapped on it or something. And they say that their AI is now at the point where it can automatically detect 2160 different types of defects among 514 components. Overhead equipment, supporting structures, suspension gear, all of that. They say as of six months ago, they had basically detected 26914 defects in these overhead power lines. Now the traditional way, you used to have an inspection crew. They say that the old inspection crew could do 2.5 kilometers. They could check that much of power lines in one day. The AI on the train can do 50 kilometers of power lines in a single day. And it can just sort of run forever. Well, I guess it doesn’t work as well as night. So that would be one example. Another one that I thought was interesting, well, they have stuff like residential management. property management, so you put the cameras up for property management companies and they oversee access to the facility, checking vehicles, checking residents, checking what parking spaces are available and which ones aren’t, anomaly detection. They look for things. The cameras can pick up smoke, fire, falling objects, assaults, crime, things like that. That’s one that they’re working on. They’re doing one in Shanghai on the Bund. Oh, actually here’s another good one. The Hangzhou International Expo Center, which is actually totally spectacular. Hangzhou is a really great city. Like used to not be that great, but you know, because Alibaba and other, it’s become a tech company, or tech city, and it’s really spectacular now. Anyways, this is a massive, multifunctional complex expo center, 850,000 square meters. hotel, catering, commercial, exhibit hall, office spaces, 400 conferences per year, not this last year, 700,000 annual visits, 40,000 visitors per day. There’s a lot going on there. So they’ve sort of equipped their AI to do a whole lot of interesting services like, okay, same one, anomaly detection and alert, security, safety. fire theft, I mean that’s going to be pretty standard across a lot of companies projects. Rental management, you know, renting out the spaces. You can basically create a virtual replica of the whole center and then you know the seventy five hundred exhibit booths and the forty one hundred parking spots and you can give you know your clients and you can give customers a virtual reality view of the whole center which comes in through the cameras. they’re really important for generating virtual reality spaces. And you can sort of explore and check on the whole center remotely. You can plan things. And in theory, it’s a better experience. You get a higher occupancy, things like that. And the one I thought was kind of cool was enhancing the visitor experience, where you provide augmented reality, navigation, tour guides. I mean, when you hear a lot about the metaverse, augmented reality, virtual reality, A lot of that requires moving the physical world into the virtual world. Well, how do you do that? You do it with computer vision. I mean, that’s what does it. Like computer vision is the bridge between the digital world and the physical world. You put the cameras in the center, it automatically then translates all that into your virtual world. I mean, it is kind of an important linkage. And SenseTime has renamed, I think, one of their businesses. They have a foundry business and a commercial business. but they’ve renamed one of them the Meta, which is, you know, they’re going after sort of Metaverse, virtual reality, and you know, they’ll provide the computer vision that makes that possible, which is actually a huge part of it. Okay, let me wrap this up, because I’ve been talking for quite a long time here. So can we project the economics? I mean, it’s a big factory, but it produces AI models, very data-intensive, very computer-intensive, computational power. But this is where you get to the point where like AI software is just a different thing. You’ve gotta clean the data, you’ve gotta bring in the data, you’ve gotta input it all the time, and then these models have to get to a certain amount of accuracy to be valuable. If you have a model that’s only detecting fire 70% of the time, that’s not good. It’s gotta be 99. Other things like facial recognition in a crowd, if it’s 70%, that’s kinda all right. So you have different. thresholds for the data and the requirements to create these models and some of them are gonna be more expensive and some of them are gonna be cheaper. The more long tail use cases you go after, generally it’s harder to get that data, but it makes the model more valuable. And then the issue is okay, some are gonna be cheaper, some are gonna be more expensive to train, but then they also can degrade in their functionality. Now if you train, let’s say identifying a car going down the street, okay, that’s probably gonna stay. Once you’ve trained it, it’s probably gonna remember cars forever. But other models can start to degrade. Things can change in the real world and suddenly the camera’s on a certain factory or street that used to understand everything, now they’re not being able to do it anymore. Maybe, I don’t know, maybe there’s more traffic, maybe there’s construction. You know, things can change in the real world and you have to retrain the models. Some you don’t, some you do. Well, that can be a significant cost as well. Do they degrade in performance in three months? A year? Two years? Um, so, you know, the unit economics are going to really depend on what models you’re going after. And I think, you know, what, for sure, what SenseTime is doing is they’ve been going after the easy stuff. Cause that’s the, you know, That’s what you would do first. So these economics, we are probably looking at the low-hanging fruit. This is like when Jeff Bezos launched Amazon, he did books first, because books were a much simpler problem than doing groceries and perishable seafood and fashion. You know, he went after the easier use cases first. So we are probably looking at the economics of the easier use cases. So if you look at it like just as one model versus another, they’re each gonna have their own unit economics. There’s gonna be a spectrum of economics. But then you get to an interesting question, but okay, if I have 10 models for detecting this type of traffic, do the existing 10 models make the 11th model cheaper? Right, so it’s not like you’re training them all individually. My data repository’s gonna get bigger, and the… the intelligence of the existing models might be translated into the new model. And that’s pretty much what SenseTime is doing. And they’re pretty, they talk about this, that they build core base models for common situations and then they use those and they add additional data for long tail scenarios on top of the base models. So it’s like they have specialized models that are. long tail, but then they have common base models and that’s how they sort of tie those together. So there’s a lot of interesting economics happening here at the individual model level, at the idea of a model portfolio, and then the idea of look, as the factory gets bigger and bigger and you get more and more data, that’ll change the economics as well. So it’s complicated. Like I don’t think we’ve seen granular enough data to really understand the unit economics. I think we’re seeing the low hanging fruit. But that would be one answer. If these were all independent, I’d say we’re seeing the low hanging fruit. But the counter argument to that is, yes, but when you get a lot of those together, they are gonna combine and make the complicated ones cheaper. Anyways, let me move on to the other question then I’ll finish up here, because this has been a while. Okay. Last question, the big one. Does an AI software factory… have network effects? Is it getting smarter as it gets bigger? Now, most of the time when we talk about network effects, it’s sort of like number of drivers on the road for DD or number of merchants. I mean, it’s very simple. This is a lot more complicated. Now, they say they have network effects, but the way I’ve always described network effects is, What is the marginal value and or utility of an incremental user or usage? That’s kind of my, it’s a marginal effect. Okay, but are we really talking about incremental users and usage? Does having more customers, a company, a government body on smartphones, is that the incremental user that makes the algorithm work better? Or. is the incremental user customer, maybe it doesn’t make the algorithm better, although it really does. Maybe it makes the training process of the factory better. Maybe it makes it faster, cheaper, higher accuracy. But so that would be another way to look at the impact on the factory versus the impact on existing models. But you could also say, okay, maybe I’m not increasing the number of customers, but I’m. What if a customer is using more algorithms? Is that incremental usage? Does that matter? What happens if it’s the same number of use cases for the same number of customers, but they all just put up a lot more cameras or the cameras get better? What if we apply more and more AI models to more and more use cases? Do they sort of synergistically help each other? Does one model make another model better? So in that case, it’s not an increase in users or usage, it’s just an increase in models. Is it about real-time performance where the amount of data coming in from all the cameras in real-time makes the performance of the algorithms better? Is it a real-time thing where we want to look at the amount of data coming in in any minute? Or is it about the cumulative data that has come in over the last five years? You could argue one of those scenarios might be much more valuable for one type of model than another. So when you start to break down network effects, it’s like there’s a lot more dimensions to this than your typical network effect for Adobe. The more people that use Adobe PDF, then the more valuable it is to everybody because we can all share files. That’s nice and simple. This is much more complicated when we start to take apart network effects. So I put this under the category of advanced network effects, which is, I’ve talked about briefly, but I haven’t really written. I have a whole bunch of… notes to myself on this, but I haven’t quite figured it out. So this idea of a feedback loop and a network effect, it’s really interesting. I boiled it, okay that was a lot of talk. Here’s the three questions I have and I’ll write these in the notes. These are the three I look at. Do existing AI models get better the more they are used? And how much? Does it go up linearly? Does it go up for a while in flatline? And how does that differ with different types of models? Long tail, common. That’s one type of feedback loop network effect. Question number two, do existing AI models make each other smarter and better? So it’s not just usage, data usage, it’s having a portfolio of models makes them all better. Separate question. Last question, do existing AI models and the data they collect make newly created AI models smarter, do they make them faster and cheaper to train? So there’s sort of these network effects between existing models, between usage and trained models. I mean, a couple of dimensions, but I take it down to those three questions. Okay, and I think that is more than enough for today. So let me just give you an opinion since I’ve been giving you a lot of… questions more than answers today. My working explanation, the two questions, what are the unit economics of AI software at scale and what kind of network effects do we see in AI software businesses? My takeaway, I think this is what I have described as a learning platform with at least one new type of network effect and maybe two. That’s where I put that one. I think they are also gonna build an innovation platform on top of the learning platform, which is when they get these widely used standardized model, they will then get more and more developers who write on top of them. So that’s separate. So that would have an indirect network effect. So there’s at least two network effects. I think there’s gonna be three. I call it a learning platform with an innovation platform being built right now, but it’s quite small. And third point on this, I still think this is 51% about an integrated bundle going for massive scale. An integrated software bundle going for massive scale. That’s 51% of the story for me, which sort of makes it more like Adobe than anything else, but instead of software, it’s AI software. That’s kind of where I fall. That’s kind of my conclusions at this point. I’ll put those in the notes if that helps. And that is it for today. Kind of dense today. Two concepts again, learning platforms, integrated bundles. As for me, it has been, like it’s been one of my most productive weeks ever. I have just been kind of a content machine. I’m teaching at the SEABs in Shanghai tomorrow, although online this time. So I had to put together some lectures. Book number two got finished. I gave a talk this afternoon on smart manufacturing. Podcasts, I mean how many pieces of content I’ve done this week must be nine or ten. It’s basically from 7 30 in the morning until about midnight for a weekend. A week straight so I’m kind of mentally beat. I feel good though, I always feel good when I feel like I’m really sort of operating at capacity. So anyways I’m flying out to Sri Lanka in middle of next week. That’s good, I’m gonna look forward to that. I love the kind of safari stuff. I really enjoy that, so I’m gonna go see some elephants and things like that. Looks like I’m gonna meet the folks from Huawei in Sri Lanka, maybe Duras, the e-commerce company, I think I might chat with them. That’ll be kinda fun. That’s literally like my favorite thing to do. Travel, see some elephants, and then talk with tech companies. I mean, that’s pretty awesome for me. Anyways, I hope everyone is doing well. I hope this is helpful. And you know, it’s a bit of an interesting month, so everyone stay safe. And I will talk to you next week. Bye bye.
I write, speak and consult about how to win (and not lose) in digital strategy and transformation.
I am the founder of TechMoat Consulting, a boutique consulting firm that helps retailers, brands, and technology companies exploit digital change to grow faster, innovate better and build digital moats. Get in touch here.
My book series Moats and Marathons is one-of-a-kind framework for building and measuring competitive advantages in digital businesses.
This content (articles, podcasts, website info) is not investment, legal or tax advice. The information and opinions from me and any guests may be incorrect. The numbers and information may be wrong. The views expressed may no longer be relevant or accurate. This is not investment advice. Investing is risky. Do your own research.