I have listed 5 types of platform business models:
- Marketplace and transaction platforms (e.g., Taobao, Meituan, Lazada)
- Innovation and audience-builder platforms (e.g., YouTube, TikTok, Android)
- Collaboration / coordination and standardization platforms (e.g., Slack, Microsoft Teams, Adobe)
- Payment platforms (e.g., MasterCard, Alipay)
- Learning platforms
Here are examples of the first two.
But I have not really discussed #5, learning platforms. Mostly because it is not that clear to me. I think this is the platform business model that is just emerging – and it is where a lot of AI / machine learning is happening.
But my basic definition is learning platforms are business models where the more users and/or interactions, the smarter and therefore better the product / service becomes. It is the platform’s rate of learning that makes it a better service.
But this can be confusing for a couple of reasons.
First, this gets lumped in with the idea of personalization and customization. People point to Amazon and Netflix as a “data network effect” companies. The more you watch Netflix, the better the video recommendations get. The you use Amazon, the more the store becomes customized to you. I don’t consider this a big phenomenon that really separates companies. I think this type of personalization is just going to be a standard feature in a data-rich world. Amazon and Netflix got there early. But personalization and customization are table stakes in most businesses. Required but not extraordinary.
Second, this is also confused with what people call “data network effects”. People claim there is a virtuous cycle (i.e., a flywheel) where when you sell more, you get more data you get, which lets you improve your products, which makes you sell more. I don’t really buy those big leaps in logic. It’s a lot of wishful thinking.
Third, I don’t think “data” really works as a concept. Data network effects. Data competitive advantages. Data advantages. It’s all just too vague. Try to actually define “data”. Numbers in a spreadsheet? Images? Maps? Movements of cars on a street? Faces? What in life isn’t data?
But a learning platform as a network-based business model makes sense to me. If you view this as a business model (not a feature, not a flywheel), you can see a much clearer and more powerful phenomenon. A learning platform is a business that somehow (by data or whatever) becomes dramatically smarter and better the more it gets used – and mostly because of interactions between one or more user groups. That jumps out at me as something powerful I can identify and measure versus competitors.
However, one caveat. These user groups can be people AND / OR SOFTWARE. That is probably the most confusing aspect. And this is what I am struggling with. When we start talking about AI and machine learning as the centers of learning platforms, we can see that the user groups interacting don’t have to be people. It turns out software can interact with other software. And machine learning can interact with and learn from the physical world. Tesla cars do get smarter just driving around by themselves. That is the part that gets me confused in terms of platforms.
For a clearer example of a learning platform, think about Google’s search engine. There is not that much learning going on when you shop on Amazon. Just personalization and curation. But think about when you search through all of human knowledge on Google? Think of all the billions of webpages and their content that the search engine hunts through to find the answer to your question. Think about how much more robust and powerful a learning platform could be in this situation.
The other concept that comes up here is long-tail. Lots of search engines are fine for common search inquiries (what is the temperature today? What time is it?). But a smarter search engine is much better in the massive sea of long-tail searches (What were grain prices in 1932? What are the most important scientific breakthroughs in chemistry in Japan?). Google is a superior learning platform in long-tail searches versus Yahoo.
So, think about learning platforms as a powerful type of business model, with lots of sub-types and subtleties that we are just starting to discover.
Which brings me to computer vision.
Computer Vision is Surprisingly Commercializable AI.
In podcasts and articles, I have pointed to Professor Agrawal’s description of AI as “cheap and fast prediction”. And that is a useful way to think about it for strategy. That podcast is:
- Products with Personalities? My Interview with JD About Conversational AI (Jeff’s Asia Tech Class – Podcast 31).
And over the past five yeas, we have seen a slew of AI-focused start-ups. Lots of cool idea and tech. But very few that have become commercial successes. Computer vision is a AI capability that is surprisingly commercializable. It turns out computers are really good at seeing. AI is very good at understanding what is in photos, videos and on cameras. And this has lots of applications businesses will pay for.
Computer vision and natural language processing are strange to think about. Computer vision means computers are able to see the world themselves (natural language processing means computers can listen to the world and speech). And they have never been able to do this before. Historically, computers have only known what we told them. They got most all their input data from us. However, if a computer can see and listen to the real world itself, it can learn on its own. It can make predictions all day long about what is going to happen with traffic on the street. And this is what a self-driving car does. The computer is constantly seeing the road itself, making predictions and then making decisions. It’s odd to think about.
A couple thoughts on the commercialization of computer vision:
First, computer vision has lots of applications in retail, logistics, supply chain, consumer products and security (i.e., surveillance). It’s commercializable in a way that NLP is not. We are seeing B2B business models that are selling a combination of hardware, software and services. Lots of freemium and Saas subscriptions for lots of use cases.
If search engines are about long-tail searches, computer vision is about long-tail use cases. They started with identifying someone by their face in a photo (a static 2D image). Then identifying a face walking down the street (a dynamic 3D video). Then they started identifying who they are walking with. And where they were going. And what they were wearing. And what there were looking at as they walked by. And so on. Computer vision keeps improving there are more and more use cases. I have seen computer vision start-ups that can tell when someone is lying based on micro-movements in the face. I have seen start-ups that can manage herds of cattle by identifying the face of each cow. There’s a ton of crazy stuff.
Second, the value of computer vision is growing with its increasing complements. The standard complement to AI is data. More data makes AI more valuable. And vice versa.
But there are lots of others complements in computer vision. Think of the growing installed base of IoT devices and sensors. Think of all those cameras being installed. On the streets. In retail stores. On smartphones. Yes, it’s creepy but all of these complements (IoT, sensors, cameras, data) are making computer vision more and more valuable. And vice versa.
The flywheel for computer vision is: More IoT / sensors / cameras -> more data -> better solutions -> new use cases / applications.
Finally, when you pair cheap and fast prediction by computer vision with automatic decision-making, you get automation. Think about this in industrial robots. You put cameras all over the warehouse and on all the robots. The centralized AI computer vision (via rapid fast 5g) sees everything and predicts what every robot should do next. Then it decides and the robots move themselves. And not just individually, but in coordinated activities of potentially thousands of robots. Think of all the learning and use cases in this idea alone.
Which brings me to China.
China Has Lots of Natural Advantages in Computer Vision.
Pretty much everything I just said is bigger in China. From the 2019 Megvii IPO filing, China is:
- +40% of the market for security cameras.
- Where +90% of mobile phones are manufactured.
- The #1 market for industrial robots.
Plus, it’s just a massive country with a lot of big data on the consumer side. Most people tend to point to China and the US as the big hubs of AI. And its big leader in computer vision is Megvii. But China is really surging forward in B2C, retail, logistics and surveillance.
An Introduction to Megvii, China’s National Champion in Computer Vision.
The company to watch in computer vision in China is Megvii Technology. It’s an AI-centric company that focuses on computer vision and facial recognition. And it filed to go public in 2019, so we have the financials. The company got added to the US entity list prior to its IPO so they pulled it. They have since re-positioned the IPO for Hong Kong, but this is currently delayed. I think they will go public in the near future. And we have their numbers to talk about. Note: the US political problems around this company are also a problem for Alibaba and Ant Group, which are some of the major owners of Megvii.
Here is a brief introduction to their business (from their pre-entity list, pre-Covid 2019 financials).
Their core business is B2G “Smart City and Community Management”.
At the time of their 2019 filing, they generated about 1B of their 1.4B RMB revenue from what appear to be government or quasi-government contracts. They cite +430 customers.
I think this is B2G, basically selling to government bureau, city governments and lots of police precincts. They sell cameras that tie into centralized software and AI. So it’s a hardware plus software plus services model. The core functions are probably identification and facial recognition. And this is the stuff that got them into trouble with the US government (and others).
This core business generates their still small but growing revenue. And it gives them overall gross profits of +65%. If you subtract out their big R&D spending, they were basically at +20% operating profits as a business. And mostly from this B2G business. These are fairly good financials for an AI-centric company. It’s hard not to like B2B Saas businesses.
Consumer devices are their #2 business. This is not clear to me.
This is selling software and services to the makers of smart personal devices. So smartphones, home security systems, and driver assistance. Basically any smart devices with a camera in it. And this generated about 250M RMB in revenue. This appears to be mostly AI smartphone solutions for China Mobile. The key functions appear to be computational photography and device authentication (by facial recognition).
I’m not too optimistic about this business. They are competing with lots of smartphone makers, tech giants and specialized app developers (like photo editing). And I’m not sure there is a long-tail of use cases that will get them a learning platform in B2C. But who knows?
Supply chain and smart retail are their best potential B2B businesses.
This was quite small in 2019, only 100M RMB of revenue. But this is the China B2B story I find the most interesting.
Retail is already a major user of computer vision. You see it everywhere. It is a key tool for gathering consumer behavior information inside stores. Amazon knows exactly what you looked at online. But Walmart knows nothing about what you do in their stores. Computer vision can analyze where people are walking, what they are buying, who they are, and their overall behavior. Smart retail is a big proponent of computer vision. Consumers are not so sure how they feel about this.
Computer vision is also happening upstream in logistics, warehouses and the supply chain. Think about all the cameras and sensors in whorehouses, trucks, packages and robots. There are great opportunities for labor reduction, increased accuracy, increased flexibility, transformed productivity. The coordination of robots to do both simple and complex tasks all depends on computer vision.
Overall, Megvii is fascinating company doing computer vision at massive scale.
Their top 5 customers were 34% of revenue. And these main customers were 3 system integrators for smart cities, 1 system integrator for telecommunications and 1 system integrator for smart logistics. That’s a pretty good summary of their business (circa 2019). Note: their top 5 suppliers were 37% of spending and that was mostly cloud services, data services, and contact manufacturing.
Which brings us back to learning platforms.
Megvii is Basically Building a Computer Vision Learning Platform.
Mevii is an AI company at its core. They had 1,432 employees in R&D in 2019, which was 61% of their total workforce. Plus they had 405 employees in data, which was 17% of the workforce. So of their 2,349 total employees (2019), 77% were in AI and data.
Remember my SMILE marathon? This is a company that is all in on machine learning.
The company’s IPO filing had a lot of great detail on what they are focused on in technology and R&D. They say their focus is “deep learning frameworks” but I think it is a massive computer vision learning platform. Here is their description:
- They are building a “semiautomatic algorithm production line”. They repeatedly talk about how their system can self-improve. How algorithms will get created and improve automatically over time. With less and less human involvement.
- They are building a “growing portfolio of algorithms”. This is them going after the long-tail of use cases for computer vision. They are targeting fragmented needs in different verticals.
I think all of that is about a learning platform. With less and less human involvement. This is all under their core technology, which is called Brain++. The goals are to:
- Improve the effectiveness and efficiency of image and video processing. So lots of image classification, object detection, object / scene segmentation, and video analytics.
- Improve automatic machine learning. This is semi-automatic algorithm production. A system that self-improves. And becomes more automatic over time.
- Allocating their capacity to multiple users. How to run tasks and dispatch computing power.
Their other core technology is called Data++, which is their data management system:
- To securely and efficiently store data for algorithm training.
- To support semi-automatic data processing and labeling.
- To have high quality datasets to improve the quality and training speed for algorithms.
So basically, one tech system is for algorithms and one is for data.
Doesn’t this all sort of sound like Google search?
Basic computer vision may be able to do the common things for a video, like identifying people and telling people vs. dogs. But a learning platform in computer vision could excel at the long tail. It could see and then do amazing things related to videos. Like seeing which crops are not growing as fast. Like seeing cracks in sidewalks. Like seeing structural weaknesses in buildings via infrared. Like driving cars. Like coordinating the activities of hundreds of robots in a warehouse. Like seeing and tracking city-wide traffic.
Anyways, I’m still developing my thinking around learning platforms. But Megvii is worth keeping an eye on. We may have another Google-type company emerging in the computer vision space. That would be something valuable to identify early and invest it.
That’s it. Cheers, jeff
Related podcasts and articles are:
- Products with Personalities? My Interview with JD About Conversational AI (Jeff’s Asia Tech Class – Podcast 31).
From the Concept Library, concepts for this article are:
- Platform Types: Learning Platforms
- SMILE Marathon: Machine Learning, AI Factories and Zero-Human Operations
From the Company Library, companies for this article are:
- Megvii Technologies
- 38: Learning Platforms
I write, speak and consult about digital strategy and transformation.
My book Moats and Marathons details how to measure competitive advantage in digital businesses.
I also host Tech Strategy, a podcast and subscription newsletter on the strategies of the best digital companies in the US, China and Asia.
This content (articles, podcasts, website info) is not investment 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. Investing is risky. Do your own research.