Unlocking the secrets behind the company blacklisted by the US, and how it competes with Google, Facebook and Baidu.
Megvii is the first so-called artificial intelligence unicorn to try to list on the public market in China. The company provides full-stack solutions that encompass algorithms, software and IoT devices to its customers. In this article, we focus on its business model, revenue potential and competitive risks.
Megvii started in a niche area within the face recognition domain. It then launched the first computer vision (CV) open platform – Face++ – of the company in Oct 2012, one year after its inception.
Megvii derives its revenue through two standard pricing models (SaaS subscription and professional services) and one special business that it calls ‘personal devices.’ Like other Software-as-a-Service (SaaS) companies, Megvii charges by a pay-per-use model for its Face ID (a cloud-based identity authentication product) and Face++ under the SaaS business.
Megvii generates revenues for professional services primarily in two segments and charges on a project basis: government and commercial. Worth noting is that a significant portion of Megvii’s direct customers are system integrators, which provide various types of assistance in project implementation, and are not the end users. 93% of revenue in the first half of 2019 came from its city IoT solutions, whereas the rest was generated from its commercial segment supply chain IoT (retail and logistics).
Megvii productized the business and earned CNY 5.9 million in 2017, one year earlier than its SaaS offerings. The gross profit saw a sudden plummet in 2019H1 due to the firm delivering more camera modules.
Arcsoft (688088:SH), Nuance Communications (NUAN:NASDAQ) and Gracenote, among others, are competing in the niche worldwide. Arcsoft is now trading on China’s new Star market at a 122x earning level with a valuation of CNY 25 billion as of April 24. Arcsoft provides single/dual camera solutions for smartphone manufacturers, which fit various hardware configurations.
It has achieved CNY 438 million revenue from its smartphone business, with a 94% gross margin. While Arcsoft’s revenue and gross margin were on a steady growth stage from 2016 to 2018, Megvii saw a meaningful dip in 2019H1, as mentioned above. As a relatively new player in the market, it is reasonable that Megvii has entered into a new stage in delivering products. In essence, it has to work more closely with camera module manufacturers, camera sensor manufacturers and System on a Chip (SoC) platforms. Megvii might need to switch to a different type of camera module for different types of mobile phones to drive the development of camera technologies fast. As a result, the cost of sales surged 17x to CNY 18 million in 2019H1 from one year before.
The cost of sales is attributed to two elements: 1) hardware and 2) project outsourcing or technologies services. Considering Megvii’s own claims, we think the previous assumption is reasonable. In essence, the gross margin dip will jump back to normal levels when Megvii reaches the critical mass of producing deliverable camera modules.
When it comes to Megvii, its SaaS business contains a data source cost, which is unusual for an SaaS offering. China, famous for its extensive public security market (USD 80 billion) as well as significant spending on tech-enabled surveillance (USD 30 billion), is also well-known for its abundant pool of data. But data alone is not enough for building AI software – data must first be labelled, which requires much labor.
Megvii engages third-party data sources for its Face ID product development, which represented 7% and 58% of revenue and total cost of sales in 2019H1, respectively. Data from 2016 to 2019 shows that data source cost as of the total cost of sales reached a plateau of 58% after 2017.
The current machine learning technology is still in an early stage. Thus companies researching it are running algorithms on various datasets to train their models, which means large overheads for purchasing these data sources.
As such, AI businesses are very different from software businesses, whether it is in terms of cost structure or competitive advantages.
AI is early in its development, and has been witnessing sharp declines in both AI training costs and inference costs. In 2018, the cost to train a neural network like ResNet-50 was USD 358. In 2019, it dropped to around USD 20. The cost of inference has dropped meanwhile. The cost to perform inference on 1 million images went from around USD 15 at the end of 2017 to USD 2 in earlier 2019.
AI companies have been improving computing ability at five times the rate of Moore’s law. No wonder there is a massive hype around AI research and funding activities. But in practice, we found that AI businesses have limits in how their value can be built up and accessed, rented or sold – they are not like those 'builds once / sell many times' software model. It adds pressure to margins, as well as defensibilities.
This article is part I of our analysis on Megvii. Please continue to part II.