After resuming trading, it rose another 15%! What is SenseTime really rising?

_Source: _ Hotspot Research (https://mp.weixin.qq.com/s?__biz=MzA3NzgwMjg2MQ==&mid=2247500116&idx=2&sn=0cadeefe3431cee6824dfc30190fd10f&chksm=9e6e533ed432c13885639fd87f953c8553f75abe3361efb2f283e926cf943a0ee408d14f2f87&scene=0&xtrack=1#rd)

After the product launch event, SenseTime surged 30% and then halted trading yesterday. Today, after resuming trading, it rose another 15%. Many people are puzzled about what happened. Is it just hype to see a rise after a product launch? From several aspects, it can be seen that SenseTime is an indispensable core coordinate in Chinese artificial intelligence.

The most criticized aspect of SenseTime is its unclear business model and the issue of losses, bearing the heavy burden of national scientific research. Its technological capabilities have been questioned to only lead in the field of academic papers. However, significant changes occurred in 2023.

I. Discussing the issue of an unclear business model

In the decision-making AI field, SenseTime's revenue once exceeded 5 billion, but faced many problems, such as severe homogenization, lack of obvious economies of scale (from financial reports, on average, one engineer develops 10 models per year, with the most typical problem being the need to develop models for each scenario, a highly customized business), and accounts receivable issues with to G.

SenseTime's technology talents are very expensive, and it is unlikely to make significant progress in this area for business similar to Hikvision.

However, under generative AI, SenseTime significantly reduced the criticized smart city business (high cost, lack of economies of scale, accounts receivable issues), accounting for only 10%, restructuring the business into generative AI + traditional AI + automotive AI, with traditional AI mainly for large customer business and facial recognition business (2 billion mobile phones).

In 2023, generative AI revenue has reached 1.2 billion, accounting for 35%, a year-on-year increase of 200%. It is expected to double in 2024 to reach 2.4 billion (possibly exceeding 50%), maintaining a 50% growth rate in the medium to long term. By 2026 and 2027, over 75% of the revenue may be from generative AI.

Main business models:

Public cloud model, calling standard APIs, can be charged by traffic or token/computing power.

Private deployment model, mainly for banks, securities firms, insurance companies, etc., equipped with exclusive models, billed based on the number of open user accounts or specified years of authorization + upgrade fees, service fees.

Model customization services, providing training tuning or custom development services according to customer needs, usually billed as service fees Introduced the product "Jueying" in 23, which includes an integrated platform of intelligent cockpit, intelligent driving, autonomous driving connection, and vehicle-road collaboration. Jueying's revenue reached 380 million, a year-on-year increase of over 30%, with the delivery of 1.29 million units of cockpit and intelligent driving products, a year-on-year increase of over 160%.

Added over 16 million designated units from over 40 racing models (at an average price of 260 yuan per unit, the revenue for designated models throughout their lifecycle is 4 billion, with typical designated contracts lasting around 5 years). NOA business made breakthroughs in landing on GAC Aion and Nezha; released the automotive voice large model function in the second half of 23.

Currently, the main customers have changed from various zf projects to NDB, Xiaomi, Honor, WPS, securities companies, banks, and major automakers, resulting in significant changes in the entire service targets. There will be significant improvements in long-term receivables, business continuity enhancement compared to installing cameras, and scale effects improvement. Gradually formed a productized, scaled commercial model, although still B2B, it now looks more like a true artificial intelligence company.

II. Financial Issues

In the current market environment, financing is relatively difficult. It is very unlikely for SenseTime to conduct equity financing at the current stock price (previously valued at over 200 billion, peaking at 300 billion).

For the first time in history, all three expenses of the company decreased in 2023, with a total decrease of 10.6% compared to 22. The accounts receivable amount in 2023 reached a historical high of 3.9 billion, exceeding our revenue for the past two to three years, a real increase of 48.5% from the revenue receivable at the end of 2022.

The total amount of quasi-cash assets at the end of 2023 exceeded 13 billion RMB. Our total bank credit limit at the end of 333 was also close to 13 billion RMB, with approximately 8.3 billion of the credit limit unused.

In an optimistic scenario, the situation of Generative AI/total revenue is 12 billion/34 billion in 2023, 30/54 in 2024, 60/85 in 2025, and 100/125 in 2026. If Generative AI finds a good business model and achieves high growth, it will greatly transform SenseTime's current predicament. (Costs are not expected to increase significantly, as more resources are being allocated to where they are needed from the previous smart city projects).

The loss in 2023 is around 6 billion, with reduced spending on smart city business after resource focus, limited expected expenditure growth, 130+80 credit on the books, and an annual loss of 6 billion can sustain until 2026. Whether the company can turn losses around and get back on track depends entirely on Generative AI, with a high probability of turning losses if revenue reaches around 8 billion in 2025.

Moreover, once expectations change, valuation rises, enabling refinancing or negotiation for acquisition. The company can then refinance again. Valuation in the Current Market:

• Yue Zhi An Mian: 30 people, valued at 9 billion RMB
• Minimax: valued at 17.5 billion RMB
• Yue Zhi An Mian: valued at 9 billion RMB
• Wang Xiao Chuan's Baichuan Intelligence: valued at 13.5 billion RMB
• Fourth Paradigm: valued at 22 billion RMB
• Despite having thousands of R&D personnel, a large number of patents, and experience, SenseTime's valuation is not high.

3. Bearing the Heavy Responsibility of National Scientific Research

• Comprehensive support coverage exceeds 99.5%. With this deep link, various domestic chips can easily adapt to mainstream large-scale training frameworks and algorithms, such as Python, Touch Deep C, and other common open-source training environments. Additionally, the algorithm libraries of SenseTime's open-source projects like Open MM Lab, Open Lab 2, and Open GV Lab fully leverage the potential of domestic chips.
• Although it incurs some costs, SenseTime holds a high industry position, always participates in setting standards, and is showcased as the country's AI business card to the outside world. This is beneficial for project collaboration, visibility, and potential support.

4. Technological Capabilities

Hardware Capabilities:

• Computing power reserve: Achieved unified scheduling of nationwide networking. In cities like Shanghai, Shenzhen, Guangzhou, Fuzhou, Jinan, and Chongqing, new computing nodes have been expanded. The total operational computing power has reached 12,000P, with 45,000 GPUs online. The computing power is expected to double in 2024, and operational optimization capabilities are among the best in the country.
• Recently announced the rental of graphics cards (all NVIDIA cards) from Eastern Supercomputing.
• Released a code server priced at 350,000 RMB, which can assist 100 programmers in writing code locally or on a private cloud (local to prevent code leakage), likely using Huawei's graphics cards.
• On April 23, at the SenseTime Lingang AIDC in Shanghai, SenseTime Technology and Huawei jointly announced a major cooperation to launch large models for finance, healthcare, government affairs, and coding in various fields. This collaboration aims to integrate the resources and advantages of both parties in computing power, data, and application scenarios to jointly build a large model ecosystem, accelerating the deep development of AI technology and industry applications.

Software and Engineering Capabilities (details can be seen in the product release conference):

• Ririxin Big Model 5.0: The model's programming and reasoning language capabilities surpass GPD4. It has multimodal capabilities and can understand the meaning of images.
• Terminal small models in cooperation with Honor and Xiaomi, with capabilities comparable to large models. Can be deployed on smartphones, laptops, and cars. Breakthrough in decoder capabilities, able to process complex tables, in collaboration with WPS.
• Miao Hua Wen Sheng Tu model surpasses industry competitors, reaching a leading level domestically Soon to be released, Wensheng's video will be benchmarked against Sora, capable of generating movie-level 4K resolution videos in minutes.

Compared to startups still chasing after GPT-4, they have not achieved full native multimodal integration of text, images, and videos, nor do they have hardware training capabilities, hardware production capabilities, engineering practical capabilities. Looking at the valuation of 100-200 billion RMB, SenseTime, with a valuation of over 200 billion (considering 23 billion in net assets, mainly 9.4 billion in cash, a building in Lingang, machinery and graphics cards), has not actually provided any valuation.