Analyzing a Chinese A.I. Tech Unicorn

Looking at Megvii

Author’s Note: I read an article about Megvii still trying to go IPO and harkened back to this post. That company’s IPO journey has been the IPO from hell. They keep filing and un-filing. First it was the Hong Kong stuff. Then it was the blacklisting by the United States. And then the economic strains of the pandemic. Sucks for them.

I wrote this and was unable to turn it into a video since I couldn’t put together the visuals. But I think it still makes for a great email post. The stuff below might have been written a year since the first filing in Sept 2019, but I still think it’s relevant to understanding the company.

The thing that’s really interesting to me is how the company keeps trying to again and again IPO in Hong Kong. It says something about the importance of Hong Kong’s role in Chinese tech circles. Crucial for getting money out of China and into the western system. That has been Hong Kong’s role for decades. I don’t think it is going to change. How else is Joe Tsai going to get the money to buy his NBA team?

---

China has developed a number of major AI “unicorns” and a fair amount of ink has been spilled about what they do. Much like their AI models, their actual business model and operating metrics have been largely cloaked from us.

So it has been immensely interesting to me to see Megvii drop their IPO documents in Hong Kong. Because it is Hong Kong, the documents are issued in English in addition to in Chinese.

I read through the filing and wanted to share a few thoughts.

What is Megvii

This is the tech company behind the development of products such as Face++, a facial recognition system. It has been valued by Chinese venture capitalists at over a billion dollars. Maybe even four or five billion.

Alongside Sensetime and maybe ByteDance, it is one of the most prominent AI companies in China and among the few which maybe someone outside of the country can recognize.

Megvii, the AI Project Conglomerate

I recently read a good blog post by Ben Thompson that asks the question of what makes a technology company. Jumping off this definition, I wanted to think about what makes us call Megvii an AI company.

The line between “AI tech company” and “tech company now doing AI” is getting blurry. Much of the best, cutting edge AI technologies are open sourced, making it easy to adopt. Regular tech companies are very quickly adopting AI into their workflows and products. For example, Google uses AI to power its search results. Google has a lot of AI driving their products. They own DeepMind. The AI tech helps improve their main revenue-generator, which is search advertising. They do not sell the AI itself. Are they specifically an AI company?

Megvii can be considered kind of like an AI conglomerate of sorts — 3–4 disparate companies offering totally different solutions for different industries loosely unified under a single banner of having a shared technology.

It is easy to see how this situation can develop. While in the startup phase, the company makes their way in front of prominent people in various industries and present them a vague demo of what the AI tech can do. The customer is not interested in the product on the table and immediately suggests something else. Since they got the money, you go back to the engineers and they say, “Guess we can do it if they got the money and the data”. Megvii’s executives must have thought, “Well yes we can do that” and they ran for it.

Sideline: Why no Megvii in the US?

I wonder why such an AI conglomerate like Megvii exists in China but not in America. I got theories but no answers.

Maybe current American companies very quickly adopted AI technologies on their own without any need to bring in a third party to help them. The tech is open source. You just need the data to train the models. Why bother bringing in another firm to take the benefits if you can do it yourself?

Maybe an American AI startup very quickly sticks to one customer and then grows within a single vertical — finding it profitable enough to go big. I know that there are a lot of small AI firms that operate in this way. They do AI but stay within their lane (just robots, or just filters), lacking the funds, resources, connections and developer talent to invade other strategies.

Maybe the product itself is not all that better than the current alternatives for American companies. American companies are no longer predominantly manufacturing and industrial-based like they are in China. They provide services — meaning they get something done for the customer. And in many cases the best service is driven by humans, not by computers.

Maybe the American equivalent to Megvii is not public or well-known. It is known that Palantir has a data product that is quite difficult to use without extensive customization. They leverage their workforce engineers to create special systems for their customers. Megvii seems to leverage their specialities in A.I. To create a variety of solutions for their customers.

Or maybe it is just a China thing. There are a lot of companies that exist in Asia that have no American parallel. Mitsubishi, Mitsui and other Japanese trading companies for example (sogo shosha).

Product Breakdown

Megvii makes the majority of its revenue comes from adapting its core AI technology for industry:

• Personal IoT

• City IoT

• Supply Chain IoT

IoT to me means nothing so if you break that down even further into what the actual real products for each vertical are, you get:

• Personal IoT (cell phones) — Video processing technology for better pictures

• Personal IoT (authorization) — Facial and fingerprint recognition tech for payment authorization, security and logins.

• City IoT — Facial recognition and traffic management technology for governments (the stuff that sets privacy advocates on edge)

• Supply Chain IoT — Robots for warehouses. Kind of like the type that we see in Amazon’s warehouses

The Personal IoT product revenue, facial recognition services for phones and other platforms, is Megvii’s second largest business line (after City IoT which I will talk about next). What sort of AI services do phones need? For example they will sharpen the background of a digital image so that it can appear to be a better pic. Also they help make “prettier” selfies (I think) (page 159). I have seen these “beautification” filters and their effects on your face all over China. Man, they are weird.

Facial recognition is used on other services for authentication and safety reasons. For example, Taobao, China’s shopping site, would use Megvii’s facial recognition to log you in without needing a password. Or the Chinese version of Uber (Didi) would use facial recognition technology to identify that you got the right rider. These apparently have a per-use paid model (page 157) and so I presume it’s their SaaS business.

Like I wrote before, Personal IoT is Megvii’s second largest business line but it is also the company’s most profitable business with a 73% gross margin for the SaaS side and 97.3% gross margin for the phone selfies features. The latter makes sense as it appears to be simply a license fee based on the number of phones sold. Strangely though in June 2019, the margin for the phone-side product fell from 94% the year before to 78%. The notes say that this is due to them having to engage third-party data source providers for their Face ID service (page 225). No clue what that means.

Okay let us talk hardware. While the AI is the reason people come to Megvii, much of the company’s revenue comes via the sale of AI-powered hardware. Megvii in many ways appears to be more of a hardware company than a SaaS company like SalesForce or Workday.

Much of this hardware is going to their city IoT segment. 73% of their revenue appears to come from this segment — which is about selling video surveillance hardware and such. This is a massive segment and appears to be incredibly varied. It is not just all surveillance cameras with facial recognition:

• Smart public safety — this is what most people will think about when they think about city IoT — surveillance cameras that track and remember you

• Traffic — Tracking the flow of cars to see when a car accident occurs or (and this is more interesting to municipal authorities) when someone breaks a traffic law so the city can send them a ticket

• Urban resource planning — They say this is for landlords to make sure that their apartment tenants do not AirBnb out their homes. Hm, guess there is a market for that.

• Community — This is for making sure that no one unauthorized enters a communities. For example, if someone wants to go into an office building who isn’t an employee

• Commercial outlets — We get asked a lot about this one too. Mall and retail shop operators want to know if their shops are crowded or not, and when those periods tend to happen so that they can properly staff.

• Schools — School safety. For example, in America the big conversation for school safety is weapons detection so you can be alerted when someone brings a rifle to a school. In China it is more if to detect if a kiddo is late

They offer a variety of hardware embedding these chips for their city IOT segment revenues:

• AI cameras — “MegEye-series”, which are good for urban roads and commercial outlets

• Edge servers — “MegBox”, which is a NVR (network video recorder) of some sort that can collect video streams from regular cameras

• Cloud servers — full on servers

Their logistics and supply chain IoT product, third largest and growing, consists of a software platform and robots. Hetu, launched in 2018, helps guide robots around warehouses. In a warehouse, you have a lot of items in stock and when someone buys something, you as the seller need to get them off the shelf and into the box. Hetu helps guide the robots to do that, optimizing routes so that the robots can get around with ease. They charge their customers for the robots themselves as well as a licensing fee for the use of the Hetu software.

If you are an investor coming to Megvii thinking that you are going to be investing in a scalable cloud software business then you should know this. Hardware looks like it will be a fundamental part of these integrated AI solutions. In some ways this is beneficial. You are selling something that people can actually hold and touch. Yet at the same time, revenues are going to be always boxed within the boundaries of selling something physical. You need to deal with inventory risk, manufacturing delays, and more. Worse yet, because much of this hardware revenue is project-based, that hardware piece tends to have to be customized in some particular way. There’s real marginal cost here.

So the summary of my description of Megvii is that it is a project-based company competing for large system integration projects from governments, commercial outlets and more. There is a cloud — based product too but it appears to be smaller.

The dynamics of project-based revenue

The nature of the business above is that revenue is going to be lumpy. That means that there are going to be good years and bad years.

Between the three product lines, just one of them can be considered to be consistent, subscription-type “service revenue”. That is the Personal IoT line, which has many different lines and sells to many different customers. Personal IoT royalty revenue is partially based on shipments. So if a customer does not ship as many phones in 2020 as they do in 2019, revenue there will drop (page 157).

We don’t know much about the supply chain revenue so we cant say much. It’s just about a two year old business. But what I am curious about is the city IoT revenue. It’s highly dependent on winning projects from governments and large clients. And winning these jobs is a matter of brutal competition, delivering on features and matching prices. It’s a rough gig. There are a few other AI companies out there and I wonder if they’re doing the same things and in the same deals as Megvii. Then how do you win the deal? My best guess on how is below.

Top Customer Breakdown

Another part of Megvii’s customer arrangement with its biggest partners appears to be selling them a share of the company. Megvii’s biggest name customers include Lenovo, China Mobile and CapitaLand (Singaporean real estate group). With the exception of CapitaLand they hold some portion of Megvii shares (page 181).

Megvii’s top 5 customers accounted for over a third of their revenue (34%, page 182), which feels like a whole lot to me considering that they tout that they say that they have 339 customers in their city IoT division, 1,102 for their SaaS and just 34 for their personal IoT. The top five are not named but are described:

1) a system integrator in China for smart city management solutions (single handedly responsible for nearly 14% of total 1H2019 revenue, or $18 million USD) 2) a system integrator in China for smart city management solutions, 3)a system integrator in China for smart logistics solutions, 4) a provider of telecommunications service in China (likely to be China Mobile) 5) a system integrator in China for smart logistics solutions

I saved Alibaba for its own section because that connection is on a whole another level.

Friend of Alibaba

Megvii shows deep ties to Alibaba beyond just their investment. It’s complicated.

You got the Alibaba Group itself holding some 20.5 million shares. Then you got Taobao China — Taobao is the name of Alibaba’s flagship marketplace — owning another 20.5 million. Alibaba’s financial arm Ant Financial owns some another 21.6 million shares.

Ant Financial as you recall runs Alipay, which is deeply integrated into Taobao too (they also got their own IPO issues too), but that company isn’t directly owned by Alibaba. Instead it’s held by Jack Ma through shell companies. These Ant Financial parent shell companies (“Hangzhou Yunbo Investment Consulting”, “Hangzhou Junao Equity”, and “Hangzhou Junhan Equity”, which own 70+% of Ant) also have another ~70 million shares.

There are also some really deep trade ties. Megvii likely sells to and buys from various arms of the Alibaba octopus. Megvii uses Alibaba cloud for their computing cloud. This makes sense as Alibaba has one of the largest cloud computing businesses in the world and certainly the single biggest vendor in China. They do not mention any other vendor used so this is likely an exclusive relationship. Ant Financial also uses Megvii’s facial recognition services for the same user authentication Taobao uses.

Something I noticed too is that Jack Ma, former Alibaba CEO and Chairman, personally owns 21.6 million shares of Megvii, which is a shareholding size on par with Taobao China, the company. Add to the fact that Jack Ma entirely owns one of Ant Financial’s three parent companies (“Hangzhou Yunbo” if anyone cares) which owns yet another 21.6 million shares, then I think well, it concludes that Jack Ma is rich.

(Also, no wonder Jack Ma talks a lot about AI, he owns a mega-chunk of one of the companies doing AI — maybe a 20–30% economic interest. Depending on how the IPO goes - if it goes at all - then that is worth maybe a billion.)

Strangely, Jack Ma is not on the board — but I am guessing his connections are partially how Megvii gets its foot in the door for all these competitive government projects. City government deals are not easy to chase and close.

Why IPO in Hong Kong?

All of the news items repeatedly emphasize a listing in Hong Kong. A question that I thought about is why the founders are so determined to IPO the company in Hong Kong rather than in mainland China. After all, Hong Kong is in the news right now for having all sorts of horrible political issues.

And then on the other side we got the Chinese government trying their best to get their own companies to IPO in their own country. The government has long wanted its big private companies to IPO at home rather than outside of China. Alibaba, Baidu, and Tencent all IPO’ed outside of mainland China.

With great fanfare, China has opened exchanges with goal of getting their flagship companies to IPO at home. Shenzhen has the ChiNext bourse, which is an abject disappointment. And then Xi announced with great fanfare the Star Market, which opened with tons of money flowing and well-known names such as China Railway Signal & Communication Corp.

The answer of course, I think has to do with Hong Kong’s special place in the world. The goal of an IPO is to give its early investors liquidity — to turn their shares into cash so that they can cash out and buy islands and NBA teams (the Joe Tsai Life) like the rest of the billionaire folk. The founders of Megvii are no different. I think there are a lot of folks out there on both sides of the divide who still want this unique situation to work out in the future.