Potential Trends And Prospects Emerging From AIGC Breakthrough
Since Mar 23, Chinese companies have been rushing to release large model products. As of today, more than 80 foundational large language models with parameter sizes of over 1b have been released domestically. The innovation and iteration of AI technology will foster further implementation of application scenarios, laying the foundation for the Chinese AI market’s long-term growth. Maintain MARKET WEIGHT.
WHAT’S NEW
• Recent regulatory updates of AI-generated content (AIGC) to foster healthy industry development. On 20 June, the Cyberspace Administration of China CAC released a list of registered domestic deep synthesis service algorithms (Jun 23), which includes Meituan’s intelligent customer service algorithm, Kuaishou’s short video generation and synthesis algorithm, Baidu’s Ernie-ViLG content generation algorithm, Tmall and Cainiao’s intelligent customer service and logistics algorithm, iFlytek’s Spark cognitive large model algorithm, Tencent Cloud’s MaaS industry-specific models, etc. We believe this implies a positive evolution in the licence issue for large language model (LLM), which will lead to eventual monetisation.
• Rapid growth potential of China’s AI market. According to IDC, the total size of China’s AI market in 2022 was US$12.2b with growth rate of 18% yoy, with US$8.13b for hardware, US$2.69b for software, and US$1.41b for services. China’s spending on the AI market will increase to US$14.75b in 2023, accounting for approximately 10% of the global total. IDC predicts that the Chinese AI market will reach a market size of US$26.9b in 2026, with a five-year CAGR of over 20% for 2021-26 with major expenditures in the AI field coming from industry users in the professional services sector (search, advertising and marketing etc.), followed by the government and financial industries, which account for over half of the total market.
• Tencent launches large-scale Model-as-a-Service (MaaS) product. On 19 June, during the Tencent Cloud Summit, Tencent introduced Tencent Cloud MaaS (Model as a Service), a large-scale model product covering model pre-training and intelligent application development. Tencent Cloud has developed over 50 industry-model solutions for companies in over 10 industries. Tencent also plans to apply AI capabilities to its SaaS products, such as Tencent Meeting, Tencent Qidian, and Tencent Cloud AI Code Assistant, to enhance various functionalities. To address substantial computational power and network stability, Tencent Cloud has developed the next-generation HCC (High-Performance Computing Cluster) for model training, equipped with advanced GPUs and high-bandwidth, low-latency network transmission.
• Baidu’s Wenxinyiyan large-scale model iterated to version 3.5 as announced on 26 June. Compared with version 3.0, the training speed has doubled, the inference speed has increased by 17x, and the model’s effectiveness has accumulated an improvement of over 50%. Robin Li believes that the digital economy, driven by large-scale models as a key factor, will deeply integrate with the real economy and strengthen, optimise, and expand the real economy. In industries such as automotive manufacturing, energy, and transportation, large-scale models can penetrate core business scenarios and innovate in areas such as intelligent customer service, supply chain, and system scheduling, promoting the digital transformation and intelligent enhancement of industries.
ESSENTIALS
• Recent progress of other China Internet companies in AIGC:
- Alibaba: The debut of the Tongyi Qianwen large-scale model along with the “Tongyi Qianwen Partner Program” was followed by the introduction of “Tongyi Tingwu” (“???? “) on 1 June. Tongyi Tingwu is an AI assistant powered by Alibaba’s Tongyi Qianwen large language model. This AI assistant offers various capabilities such as real-time speech recognition, recording, translation, summarisation, and key point extraction. Tongyi Tingwu is the first publicly accessible ToC application based on a large language model among Chinese internet companies. Furthermore, Siemens and Alibaba Cloud have signed a strategic cooperation agreement to integrate cloud computing, AI large models, and industrial applications. This collaboration aims to boost innovation and productivity in Chinese enterprises and contribute to the high-quality development of the Chinese economy.
- Kingsoft: On 26 June, Kingsoft Cloud (KC) announced that in collaboration with MiniMax, it would be launching the MaaS (Model as a Service) Mutual Trust Reasoning Zone solution to address trust issues in large models and data. A custom-built mutual trust virtual private cloud (VPC) is created for model vendors, with enhanced security features and various traffic control methods. On 4 July, Kingsoft Office announced its partnership with NVIDIA to enhance WPS AI services. Kingsoft Office is utilising NVIDIA Tensor Core GPUs and TensorRT to improve the efficiency of image document recognition and understanding. By deploying NVIDIA Triton Inference Servers, Kingsoft Office has optimised GPU utilisation and achieved high inference throughput. This collaboration has resulted in a significant reduction in pipeline latency compared with CPUs and a decrease in deployment costs.
- iFlytek: On 9 June, iFlytek officially released the iFlytek Spark Cognitive Model V1.5 at its 24th anniversary celebration. The model has undergone significant upgrades in its capabilities, particularly in open-domain question-answering, multi-turn dialogue, and logical and mathematical abilities. The Spark app was launched alongside the model, featuring over 200 small assistants for various everyday scenarios. The user creation centre encourages collaboration and sharing to create a portable assistant for the era of large models. During the release event, Chairman Liu Qingfeng included demonstrations highlighting the model’s intelligence in real-time information questioning, logical reasoning, and multi-turn dialogue. iFlytek is actively deploying the Spark Cognitive Large Model in education, healthcare, industry, and office sectors. It supports products including Spark Language Companion app, healthcare rehabilitation management platform, Lingyang Industrial Internet platform, and iFlyrec Intelligent Conference System.
- Uncertainties on US chip ban continue to persist. Despite the emergence of over 80 large models in China, the shortage of high-performance GPU chips and heavy investment in hardware procurement and operation continue to raise concerns to the China AI market. The lack of computational power, algorithms, and data accumulation hinder training large language models (LLMs) and other advanced computing tasks and commercialisation.
• Venturing into B2B market to fuel new wave of AIGC breakthrough. Large models are in high demand across various industries such as retail, finance, manufacturing, and government. Companies worldwide are utilising these models to enhance their services and operations. It is widely recognised that industry-specific fine-tuning of large models yields better performance in specific tasks compared to non-optimised general large models. However, the commercialisation of large models in the B2C market has been slow due to challenges such as high computational costs and potential drawbacks such as generating inaccurate output and ethical concerns. We opine that deepening the development of large language models can have a significant and long-term impact on China’s internet industry. While China has excelled in the consumer internet sector in the past two decades and is currently stagnant, expanding applications in the ToB segment, particularly in generative AI, offers promising opportunities for breakthroughs.
VALUATION/RECOMMENDATION
• At the current juncture, copyright, content review, and data security are still pending in China AIGC field. In the near future, we believe that the strategic implementation of AIGC will lead to cost reduction and improved efficiency for internet companies. In the medium and long term, the continuous digital transformation and intelligent advancement of AIGC is expected to expand into search, retail, video gaming, education and transportation and bring about industry transformation.
• Maintain BUY on Tencent (700 HK) with a target price of HK$435.00 which implies 25.8x PE based on 2023 EPS. We believe Tencent is poised to benefit from the expansion of AIGC applications into gaming, advertising, and short-form videos. The company currently trades at 19.7x 12-month forward PE, 1.2SD below its historical mean of 27.7x.
• Maintain BUY on Baidu (9888 HK) with a target price of HK$187.00 (US$193.00). As Baidu has been investing in AI for over a decade, Baidu can optimise the end-to-end process, rapidly improving the efficiency of large-scale model training and inference. Our target price implies 3.3x 2023F P/S and 17x 2023F PE. The company is currently trading at 13.4x 12- month forward PE, below its historical mean of 18.2x.
• Maintain BUY on Alibaba (9988 HK) with a target price of HK$142.00 (US$146.00) as we expect the company to continue to deliver remarkable progress in integrating LLM with Alibaba’s products to enhance user experience and competitiveness. Alibaba is trading at 10.5x 12-month forward PE, 2SD below its historical mean, below its historical mean of 23.2x on the back of a 17% EPS CAGR from FY23-26.
• Maintain BUY on Kingsoft (3888 HK) with a target price of HK$37.00. We remain cautiously optimistic on the company due to the stable growth in individual subscription and resilient growth in institutional subscriptions. Our target price implies 42x 2023F PE. The company is currently trading at 36.5x 12-month forward PE, lower than its historical mean of 47.4x.
SECTOR CATALYST AND RISK
• Catalysts: a) Increased deployment of AIGC will promote productivity and efficiency, b) monetisation of LLM by integrating with traditional industries, and c) relaxation of the US’ chip ban.
• Risks: a) Regulatory risk (security and privacy concerns), b) heavy capex and R&D costs, c) limited access of high-performance GPU chips arising from geopolitical risks.