Today's sharing is the artificial intelligence ChatGPT in-depth research report: "AIGC track will continue to lead the TMT technology revolution in external recession and internal recovery". (Report produced by: Oriental Wealth)
A summary of the study is provided below
AIGC upstream hardware layer under the ChatGPT wave
1. Chip hardware support is the core
With the expansion of the deployment of deep learning and artificial intelligence, the demand for AI chips is increasing, and the market size is expected to expand to more than 10 times the existing in the next 10 years
The core segment of the AI chip market is the CPU, and GPUs, FPGAs, etc. are focused on specific task development. There are many types of AI chips in existence, including CPU, GPU, ASIC, FPGA, and brain-like chips. All kinds of AI chips have their own advantages in the AI training and inference stages, so different AI computing tasks often need to be completed by a variety of different chips.
2. CPU chip
The CPU is the central processing unit and the computing and control core of the computer system, responsible for information processing, and the final execution of program operation. In AI computing, the CPU has the advantage of its powerful scheduling, management, and coordination capabilities, while its computing power is secondary. At present, the CPU is the cornerstone of AI computing, and the role of acceleration chips such as GPUs and DPUs can only assist part of the functions of the CPU to improve the efficiency of processing and computing. In the fields of artificial intelligence, deep learning and cloud computing, the collaboration between CPU and GPU has become the main application scenario, and the combination of CPU and FPGA, CPU and NPU is also emerging, and the combination of CPU and various acceleration chips will become the leading development trend of AI computing.
3. GPU chip
GPUs are graphics processors that were originally used to assist CPUs in graphics display tasks, especially 3D graphics displays. GPUs have a multi-core parallel computing architecture, processing large-scale parallel computing data streams is its advantage, and equipped with a large number of floating-point vector computing computing resources, so it has become one of the main chips in the field of high-performance computing. Since the data and operations of the deep neural network training process can be highly parallel and the amount of computing is huge, NVIDIA has introduced the parallel computing platform and programming model CUDA architecture, CUDA+GPU plays an important role in deep learning acceleration.
Global GPU chip manufacturers mainly include NVIDIA, Intel and AMD. According to JPR data, in 2022Q3, Intel accounted for 72% of the notebook GPU shipment market, while in the independent GPU field, NVIDIA dominated, accounting for 88% of shipments.
At present, the mainland has accumulated a number of start-up GPU new force companies, including Bicheng Technology, Tianzhixin, Muxi Integrated Circuit, Xindong Technology, Denglin Technology, Moore Thread, etc. Major domestic GPU-related companies include the leading enterprise Jingjiawei, while Haiguang Information, Cambrian and other companies have product layouts in the GPU field.
4. FPGA chip
FPGA, or field programmable gate array chip, is a product developed on the basis of programmable devices such as PAL, CPLD, GAL, etc., and the main advantages include high programming flexibility, good energy efficiency, low latency, parallel computing, reprogramming, etc.
Due to the high flexibility and scalability of FPGAs, FPGAs are mainly used in markets that need to adjust chips, including network communications, consumer electronics, artificial intelligence, automotive electronics, data centers, etc. In terms of market structure, Xilinx, acquired by AMD, and Altera, acquired by Intel, occupy nearly 90% of the global FPGA market.
5. ASIC chips
ASIC is an application-specific integrated circuit, which is the product of the close combination of integrated circuit technology and the whole machine or system technology of a specific user, and its advantages include small size, low power consumption, high reliability, performance improvement, high confidentiality, low cost, etc.
In recent years, there have been more and more ASIC chips for accelerating deep learning algorithms, most notably Google's TPU. TPUs are specifically designed to accelerate neural network inference processes (including MLPs, CNNs, LSTMs) in data center search services, with 15-30 times faster speed and 30-80 times higher energy efficiency than CPUs or GPUs in the same period.
6. Optical module
Optical modules (Optical Modules) are indispensable components in optical fiber communication, belong to optoelectronic sub-equipment, responsible for photoelectric and electro-optical conversion in the process of optical signal transmission. The optical module at the sending end converts the electrical signal into an optical signal, while the receiver converts the optical signal into an electrical signal at a transmission rate between 100Mbps and 400Gbps.
The calculation, storage, analysis, etc. of modern data are based on electrical signals, and optical fiber is mainly used as the transmission medium. By connecting optical fiber nodes, optical modules can achieve higher data transmission rates and longer transmission distances, so in AI computing, optical modules are widely used in data centers, cloud computing, supercomputers and other scenarios.
7.Chiplet technology
Based on Chiplet's integration technology, the traditional system-on-chip is split into multiple single-function or combined multi-functional "dies", and then connected to each other through the substrate in a package to form a fully functional complex chip.
According to Omdia's forecast, the Chiplet market size is expected to reach $5.8 billion by 2024 and may exceed $57 billion by 2035. From about $800 million in 2018 to nearly $6 billion in 2024, the global Chiplet market size grew at a compound annual growth rate of 44.20% during this period, showing a very rapid growth trend. Domestic Chiplet technology-related listed enterprises such as Xingsen Technology, Changchuan Technology, Fangbang Co., Ltd., Huazheng New Materials, Tongfu Microelectric, Changjian Technology, etc.:
8. Cloud computing has become a breakthrough in reducing AIGC hardware costs
At present, the cost of AI applications is still relatively high, and the cost of high-quality industrial-grade AI model design is high, relying on deep technical reserves and human resources. Large-scale pre-training models require a large amount of computing power support, and affected by the failure of Moore's Law and the bottleneck of computing power, providing computing power for the development of AI technology through the cloud computing industry chain and realizing the scheduling and operation of computing power has become a new breakthrough. Upstream infrastructure suppliers sell servers, routers, switches and other equipment to IDC manufacturers or cloud service providers, and ultimately serve downstream Internet, finance, government and other major traditional industries and individual users. With the rise and explosion of generative AI, the demand for cloud computing in the AIGC middle layer, AI model training, and downstream application layer will continue to increase.
9. Cloud computing equipment providers
Hardware providers for cloud computing include basic hardware suppliers such as CPUs, chipsets, flash memory, storage modules, and hard disks, as well as providers that provide network infrastructure such as storage devices and servers. The foundation of the construction of public cloud IaaS services is network hardware, and after years of development, the network infrastructure of China's Internet has become quite mature. Domestic cloud computing-related listed companies such as Unigroup, Lenovo, ZTE, Ruijie Network, etc.:
At the same time, the improvement of AI computing power has also brought about the increase in power consumption, and the computing density of AI clusters at this stage generally reaches 50kW/cabinet, and the increase in power consumption of a single cabinet puts forward higher requirements for the heat dissipation of data centers. The single-cabinet density of natural air-cooled data centers can generally only support 8 to 10KW, and the micromodules with isolated hot and cold air ducts and water-cooled air conditioners with horizontal cooling above 15KW will greatly reduce the cost performance.
According to CCID Consulting, the market size of China's liquid-cooled data centers in 2019 is conservatively estimated to be 26.1 billion yuan, optimistically estimated to be 35.1 billion yuan, and conservatively estimated that the market size of mainland liquid-cooled data centers will reach 128.3 billion yuan in 2025.
10. Cloud Computing Service Operators
Developing and training AI large models requires a lot of computing resources and data, and it also requires a lot of computing power on the product operation side. According to SimilarWeb data, the total number of visits to the ChatGPT website in January 2023 was 616 million. OpenAI says that the cost of a computing power cloud service generated by a single user interaction with ChatGPT is about $0.01.
5 Mid-range algorithm layer for AIGC requirements under the ChatGPT wave
5.1. The rapid development of generative AI underlying models has increased the requirements for algorithms
There have been four major trends in generative AI, the first two of which saw the expansion of model size and training data, the enrichment of modalities, and the convergence towards human capabilities.
5.2. Natural Language Processing NLP is a core pillar of AIGC
Artificial intelligence is based on NLP (Natural Language Processing) and mainly covers natural language understanding tasks and natural language generation tasks. It is mainly used in public opinion monitoring, machine translation, view extraction, speech recognition, automatic summary, text classification, question answering, text semantic comparison, Chinese OCR, etc.
With the application of deep learning, the development of NLP technology has accelerated, and the domestic intelligent speech recognition industry has emerged. On the one hand, deep learning can continuously learn language characteristics based on the vectorization of words or sentences, master higher-level and more abstract language features, and meet the natural language processing needs of a large number of feature engineering. Deep learning, on the other hand, does not require experts to manually define the training set, and can automatically learn high-level features through neural networks.
Affected by this, the domestic NLP industry has developed rapidly, such as iFLYTEK, Haitian AAC, Tolls, Hanwang Technology and other companies with leading product technology.
6.ChatGPT leads the explosive innovation of AIGC terminal application layer
6.1. Innovation in the media industry
6.1.1 Advertising Media: Explosive Manuscript Production Helps Marketing ChatGPT can effectively improve the productivity of editorial production in the media industry and inject impetus into advertising promotion. In the process of massive information collection in the news and media industry, ChatGPT can efficiently extract key content, intelligently compile news reports, undertake basic editorial work, and then improve the timeliness of news release, and it can create more diversified news releases and marketing texts, reducing the repetitive labor of industry staff.
ChatGPT's capabilities extend to the field of novel and script creation, so as to serve the production of film and television works and broaden the horizons of literary and artistic creation. Taking Haima Qingfan's efficient and intelligent content creation evaluation platform as an example, AI has been introduced to create novels, scripts, short videos and other content, and since 2019, it has reviewed more than 50,000 episodes of dramas, more than 20,000 movies/online movies, and more than 8 million online novels, greatly improving the efficiency of industry content creation. According to Gartner's forecast, the proportion of text-to-video content created by generative AI in blockbuster movies will rise from 0% in 2022 to 90% by 2030.
Gartner predicts that by 2025, the proportion of marketing information published by large organizations will grow from less than 2% in 2022 to 30%. The distribution platform and media resources of Internet advertising will continue to be enriched to meet the diversified and customized marketing needs of all kinds of advertisers, and the head media and medium and long-tail media platforms in the Internet advertising market will have broad space for development. At present, potential application companies include Mango Super Media, Focus Media, Bilibili, Zhaoxun Media, etc.
ChatGPT and generative AI will revolutionize the game production industry: AI is becoming increasingly important in the development of modern games. It can create realistic scenes from scratch, greatly saving design and development costs.
ChatGPT will enhance the gaming experience and reduce the cost of game production. Take NetEase's mobile game "Against the Water and Cold" as an example, it will be connected to the first game version of ChatGPT, so that intelligent NPCs and players can freely generate dialogues and give logical behavioral feedback based on the content of the dialogue. All the dialogue text, voice, expressions, and shots of intelligent NPCs in the game are generated by AI autonomously.
AI methods applied in games include decision trees, (deep) neural networks, reinforcement learning, legacy algorithms, and more. At present, AI can be used in games for image enhancement, scene and story generation, game level generation, improving NPC intelligence, balancing game complexity, etc.
6.1.3 Film and television industry: empowering script creation, editing and user analysis
AI technology can quickly create unique content for the media and film industry. For example, AI technology can be used to automatically generate scripts, videos, and write articles automatically. AI technology can also help the media and film industry manage content more efficiently.
In digital content production, asset production is the foundation and the most onerous link in the entire film and television production process, and traditional production mainly relies on manual labor, which is time-consuming and inefficient. Including Light Media, Huace Film and Television, Lingyunkai, etc. are also actively introducing AIGC cooperation, or business expansion through subsidiaries.
6.2. Internet track innovation
6.2.1. Financial industry: Enabling scenarios such as retail finance, intelligent customer service, and intelligent investment advisory
The financial sector is entering a critical stage of digital transformation, and AI technology has played a wide range of roles in multiple financial links such as risk management, marketing, customer service, investment consulting, compensation processing, and collection.
Among them, robo-advisor is emerging, which is mainly based on investors' risk tolerance, expected income and other investment characteristics, relying on Markowitz's asset allocation model, using a series of intelligent algorithms (big data + AI) to provide investors with investment reference, and real-time tracking of the market, low-cost to achieve professional-level dynamic asset allocation services.
In the field of consumer finance, ChatGPT can rely on big data analysis, provide intelligent investment advisory services, improve the quality and efficiency of investment advisory services, and take Flush's iInfo Investment Dialogue Robot as an example, after users ask about stock-related information, ChatGPT can give investment advice, and conduct rhetorical questions and carry out multiple rounds of dialogue according to the specific situation of investors (whether to hold, holding costs).
Flush's financial investment assistant provides business functions such as selecting/diagnosing stocks, funds, fund managers, encyclopedias, and idle chat pendants through text/voice interaction, solving problems such as lack of professional investment advisory tools, weak user interaction experience, and repeated questions and answers. In the field of investment research, through AI and ChatGPT, users can easily complete the reading and writing of research reports, not only can they achieve point sorting, summarizing and translating ideas after entering existing research reports, but also generate research reports through existing industry data and dynamics.
6.2.2. Online education: improve services and increase teaching efficiency
ChatGPT can improve the accuracy of online education services and promote the efficiency of the online education industry. ChatGPT can provide personalized modification suggestions for students' English essays, design a training question bank for students, intelligently answer students' questions, and improve the intelligence, accuracy and personalization of online education services. ChatGPT can complete the preparation of personalized lesson plans and automatic statistical analysis of data, freeing online educators from heavy work. Generative AI provides educators with new tools and fosters new forms and models of education, such as video, created by text-generated avatars and voices, to improve the efficiency of course recording.
In May 2023, Ape Tutoring, a leading online education company, launched the Little Ape Learning Training Machine. Different from other learning machine equipment, the equipment is characterized by the integration of learning and practice, covering more than 100 practice scenarios of students, and digitizing the whole process of students' learning data, process and results. In terms of AI technology accumulation, Ape Tutoring is the first company in the education industry to establish an AI laboratory, with more than 300 patents, including speech laboratory, visual laboratory, natural language understanding laboratory, and audio and video laboratory. iFLYTEK, Hao Future, Guoxin Culture, Century Tianhong, Zhonggong Education, Chuanzhi Education and other enterprises have also released the latest product and technical information related to ChatGPT, including AI lecture robots, AI teaching and research system management and other directions. Companies such as New Oriental Online, Chalk and NetEase Youdao Cloud have also begun to be involved in this regard.
6.2.3. Internet e-commerce: enabling advertising and user search
ChatGPT has the ability to learn big data of advertising delivery, realize intelligent creation and delivery of advertising, reduce advertising costs, and expand the reach of advertising. ChatGPT can create and optimize product titles and descriptions, improve the operational efficiency of product selection, delivery and social media marketing, analyze user historical data and predict future user behavior, provide personalized recommendations, improve customer service intelligence, and maintain user stickiness.
At present, many domestic manufacturers have begun to use AI large models to optimize online e-commerce. For example, JD.com's ChatJD adopted the "125" plan as a roadmap for application implementation, and Tencent built an advertising multimedia AI technology system with the "mixed element" model as the core. In addition, e-commerce platforms such as Pinduoduo and Taobao have also begun to explore.
6.2.4. Internet live broadcasting: Boost live streaming talent to fill the gap
The live streaming industry in China has developed into an important area to drive economic recovery. According to data from iMedia Consulting, the size of China's live streaming e-commerce market reached 1,201.2 billion yuan in 2021 and is expected to reach 2,137.3 billion yuan in 2025. For Internet live broadcasting, the core of ChatGPT is to be able to quickly generate copywriting, which greatly reduces the work burden and entry threshold of anchors. iMedia predicts that the demand for live streaming anchor talent in China continues to grow, and is expected to reach 6 million in 2022.
If you want to break the influence of the head anchor, you also need to cultivate more middle-waist anchors with the help of AI.
In the future, AIGC is expected to be first applied on live streaming platforms with large market shares, such as Douyin, Kuaishou and Taobao Live, etc., to improve the live broadcast efficiency of anchors and make the marketing of multiple products more professional and scientific. Among them, leading companies such as Douyin, Kuaishou and Taobao Live are expected to take the lead in relevant landing.
6.3. Innovation in traditional industries
6.3.1. Traditional office: optimize personal and enterprise office processes
ChatGPT can catalyze significant improvements in personal and corporate office efficiency and optimize the interactive experience of OA software. For Word documents, ChatGPT can provide basic grammar suggestions, give paragraph summaries or further modifications and expansions; For Excel, ChatGPT can interpret data, execute formulas to complete tasks, and use VBA code to process data, and the processing method only needs to be described by users in natural language; For a given topic, ChatGPT can generate an overview of the page content to provide guidance for PPT production and design.
On March 16, 2023, after OpenAI released the multimodal pre-training large model GPT-4, Microsoft announced the launch of the artificial intelligence service Copilot, and embedded it in Office office software such as Word, PowerPoint, and Excel, and Copilot is driven by the GPT-4 model, the large language model will be combined with the actual business scenarios of Office, such as writing, editing, summarizing and authoring with the operator. Convert ideas into designed PPTs through natural language commands, create professional-form data visualization tables in a short time, and more. In addition, Microsoft released BusinessChat, which brings together data from documents, presentations, emails, calendars, notes, and contacts to help users summarize chats, compose emails, find key dates, and even plan against other project documents.
Domestic core leaders such as Kingsoft Office explore the integration of mature artificial intelligence technology into daily office software. In June 2023, Kingsoft Office said that the WPSAI function will be gradually launched. WPSAI functions have been fully integrated into light documents, WPS text, WPS presentations, WPS forms, PDFs and other components. For example: directly ask the WPS form, you can get the result, and the formula is automatically generated; According to the instructions, WPS text can automatically generate content, understand the meaning of the text, and focus on the point; According to keywords, PPT content and speech notes are automatically generated. WPSAI has high strength in content generation, assisted reading, data analysis, etc.
6.3.2. Software service providers in traditional industries: ERP and enterprise cloud are becoming more intelligent
For traditional manufacturing, ChatGPT can also improve its intelligence level, thereby reducing costs and increasing efficiency. ChatGPT can read ERP data, help employees find the required information in natural language, and automatically complete routine business such as business inquiry and order data entry; In addition, ChatGPT can read the company's production and sales data, provide optimization solutions for manufacturing employees and manufacturing process parameters, improve production line efficiency, and use data to make better business decisions; At the same time, ChatGPT can deeply understand customer preferences, classify and manage customers, provide customized services for customers, and provide intelligent services.
Based on 30 years of enterprise knowledge base and service experience, Kingdee has independently developed the assembled enterprise-level PaaS platform Kingdee Cloud Firmament. The platform has similar intent recognition and language understanding capabilities to ChatGPT, provides intelligent cognitive services, and is also the first enterprise-grade intelligent search function in China's ERP field that supports language query business documents. Compared with ChatGPT, Kingdee Cloud Firmament is more secure and controllable. Based on the in-depth understanding of domestic enterprise business scenarios and data, Kingdee Cloud Skydome intelligent search function can provide a series of services from key word matching, semantic understanding to data retrieval, etc., to meet the needs of diversified intelligent scenarios of enterprises in an all-round way, helping enterprise customers reduce costs and increase efficiency.
Enterprise cloudification is the general trend of future development, and hybrid cloud and multi-cloud have become the main strategies for enterprise digital transformation. Among the possible participants include Yonyou Network, Kingdee International, Inspur International, Dingjie Software, Zhiyuan Internet, etc.
6.3.3 Smart City: Building a Digital Economy Application Carrier
Smart cities use digital technologies such as cloud computing, Internet of Things, big data, and spatial geographic information integration to promote the intelligence of urban planning, construction, management and services, which is an important scenario for the development of the mainland's digital economy. According to IDC's forecast, the investment scale of the smart city ICT market led by the mainland government will reach 38.9 billion yuan by 2026, with a CAGR of 17.1% from 2022 to 2026.
At the 16th China International Public Safety Products Expo held in June 2023, SenseTime shared the application of AI models in smart cities, bringing a new path to efficiently solve diversified urban management problems such as transportation and public safety, and effectively lowering the application threshold to build a new generation of AI infrastructure for smart city development. The AI model will build digital transformation infrastructure and create a "AI+ one-network unified management" urban governance model.
Among them, smart city manufacturers such as E-Hualu, Hytera, and Rotary Pole Information can be empowered.
6.3.4 Smart Healthcare: Build an AI platform for serious and precision healthcare
Analysys pointed out in its "2022 China Smart Healthcare Industry Research Report" that China's smart healthcare industry is expected to reach a market size of RMB 376.6 billion in 2022. ChatGPT can provide more accurate personal medical solutions, and artificial intelligence technology will be widely used in scenarios such as diagnosis, treatment, and patient services, providing strong support for medical personnel; Intelligent robots and smart devices will also be widely used in medical and health services; Emerging technologies such as the Internet of Things will enable online services for patients, doctors, and medical resources, and promote the online nature of the medical industry. It is expected that with the continuous innovation and development of technology, it is expected that the patient's allergy history can be automatically recorded and analyzed, and the intelligent diagnosis of the cloud can be realized, and the intelligence of the whole process of medical treatment and drug purchase will become the main advantage of the future development of smart medical care.
For example, Runda Medical and HUAWEI CLOUD officially signed a comprehensive strategic cooperation agreement in Shanghai, under which the two parties will build an AI model for the medical field based on the Huawei cloud platform to realize smart medical services. According to the agreement, Runda Medical and Huawei Cloud Computing Technology Co., Ltd. will cooperate closely in smart healthcare, enterprise digital transformation, and ecosystem construction based on Runda Medical's profound industry experience and service resource advantages in the field of smart inspection and smart medical services, combined with HUAWEI CLOUD's leading technologies and resources in AI, cloud computing, big data, and Internet of Things, and its own digital transformation practices.
(This article is for informational purposes only and does not represent any investment advice from us.) For information, please refer to the original report. ）
The report totals: 91 pages. Due to space limitations, only some of the contents are listed.
Selected report source public: [Phantom Studio], reply to the keyword "202307" to get the full PDF electronic version