Shared todayAI series in-depth research report: " 2024 AI Investment Strategy Report: AI Three elements resonate, AIGC cloud-to-end acceleration》.
(Report produced by: China Galaxy Securities Research Institute)
Total report: 103 pages
Core point of view: The industry’s popularity will continue, and we will actively grasp the six main investment lines.
(1) Industry market review and 202303 performance summary
1. Affected by the boom in the borrowing industry in the first half of the year, wages fell back after the shock in late June.
Looking back on 2023, affected by the new wave of large artificial intelligence models triggered by ChatGT, the overall performance of the industry is active.
2. The industry’s revenue growth rate declined year-on-year in the first three quarters, while net profit attributable to parent companies improved year-on-year.
In 2023, due to factors such as local financial pressure, industry performance in the first three quarters fell short of expectations. In the first three quarters of 2023, industry revenue increased by 5.00% year-on-year, but the growth rate declined (last year’s Q3 growth rate was 10.36%). Net profit attributable to the parent company increased by 12.99% year-on-year (last year’s Q3 growth rate was -59.39%). The net cash flow generated from operating activities in the first three quarters was still negative. The industry’s overall ROE (diluted) was 1.79%. The average gross profit margin of companies in the industry was 39.95%. (YoY +1.8pct), net profit margin -4.86% (YoY +11.32pct).
3. The overall performance of industry revenue and net profit in the past five years has been poor, but corporate R&D investment continues to increase.
Affected by multiple factors such as local fiscal expenditure pressure, the computer industry's revenue and net profit levels have performed poorly since 2020. The enterprise's R&D expense rate has increased steadily, with the expense rate increasing significantly to 19.22% in the third quarter of 2023.
(2) Review of overseas technology stocks market
1. Since the beginning of the year, U.S. technology stocks have generally outperformed the U.S. stock market, while Hong Kong stocks and Chinese concept stocks have been sluggish.
From the start of 2023 through Dec. 4 ET, the Nasdaq and the Philadelphia Semiconductor Index have outperformed the S&P 500. From the beginning of 2023 to early April, the TAMAMA Technology Index outperformed the S&P 500 Index and the Nasdaq Index; after early April, the TAMAMA Technology Index outperformed the S&P 500 Index, the Nasdaq Index, and the Philadelphia Semiconductor Index.
2. U.S. bond yields will continue to rise in 2023, and the valuation of technology stocks will come under periodic pressure.
In 2023, the Federal Reserve will raise interest rates by 25 basis points each time on February 1, March 22, May 4, and July 26. The federal funds rate will rise from 4.50% at the beginning of the year to 5.50%. Bond yields began to move higher in early 2022 due to a series of interest rate hikes by the Federal Reserve; in early October 2023, the benchmark 10-year U.S. Treasury bond yield peaked at 4.81%, the highest level in 16 years. The Federal Reserve's move put technology stocks under periodic pressure, and the trend of interest rate hikes was sluggish. As the Federal Reserve stops raising interest rates and is expected to enter a rate cut channel in the first half of next year, the valuation pressure on technology stocks is expected to ease.
(3) Investment suggestions: It is recommended to grasp the six main investment lines
Looking forward to 2024, with the implementation of AIGC in multiple scenarios, the demand for upstream computing power remains high. The release of the value of data elements will bring about a double improvement in the volume and quality of the data industry. A large number of overseas parameters and large models are open sourced, and the penetration rate of the domestic vertical large model industry is rapid. Ascension, we believe the computer industry is poised for a Davis Double Tap next year. Molecular side (performance side): Different from the valuation drive in 2023, the computer industry is expected to usher in the resonance of the "three elements" (computing power + data + algorithm) in 2024. Active fiscal policies are expected to bring about expected improvement in downstream demand, and data As factors enter the table, molecular performance is expected to bottom out and rebound. On the denominator side (valuation side), with the Federal Reserve expected to cut interest rates in 2024, if domestic monetary policy is expected to be loose next year, it will benefit technology growth and improve liquidity, which will help further increase valuations.
Specifically, it is recommended to grasp six major investment main lines in 2024:
Main line one, the computing power side: The proportion of AIcapex in the spending guidance of major overseas manufacturers continues to increase. Driven by domestic policies and large model parameter benchmarking and overseas upgrades, global intelligent computing power is expected to maintain a high boom. Different from the market's view, the market is worried that future computing power demand will peak in the short term and be lower than expected next year. We believe that from a long-term perspective, the new generation of artificial intelligence has triggered the start of the sixth round of Kangbo, and the AIGC large model is becoming the trigger. The productivity revolution has just begun. NVIDIA's latest financial report shows that data center revenue increased by 297% year-on-year, and the prepayment guidance for 2025 remains high. In the next five years, the compound annual growth rate of China's intelligent computing power is expected to reach 52.3%, and the compound annual growth rate of general computing power is 18.5%. The sub-fields focus on 1) localization of computing power (focusing on Huawei chain; 2) high-bandwidth storage (HBM industry chain); 3) AI server leader: 4) liquid cooling industry chain: 5) computing power leasing.
Main line two, data side: In the first year of the three value releases of data elements, they will be included in the table to promote the "explicitness" of data value. The "data human form" will be implemented in early 2024, marking the completion of data's leap from natural resources to economic assets, and the official start of the era of data assetization. Different from market regulations, the market mainstream estimates that the market size of data elements is expected to be 10 trillion market space. We estimate that the data capital stock market size is expected to reach 35.49 trillion in 225. Data elements are fully expected to become an important support for improving the balance sheets of government and enterprises, thereby promoting the upgrading and transformation of the domestic economy, and promoting the comprehensive transformation and upgrading of local governments from "land finance" to "digital finance." It is recommended to focus on: 1) The timing of the implementation of data assetization rights confirmation, pricing and other policies; 2) State-owned cloud vendors; 3) Enterprises with data assets or data operation rights; 4) Data governance, data security, data replication and disaster service providers.
Main line three, algorithm side: The large model continues to advance from cloud to end, the cloud continues to be upgraded, and the terminal large model is about to be released. OpenAI is expected to release GPT-5 in the first half of next year, and Meta is expected to be developing Llama3, an open source large model that is more powerful than GPT4. Google released the large model PaLM2 to implement lightweight mobile terminals, and will soon release the multi-modal large language model Gemini; Qualcomm released the latest Xiaolong 8Gen3 chip model StableDiffusion and its core plug-in ControlNet, both of which can be run on the terminal. The cloud service product market continues to expand, catalyzing the implementation of large models to edge devices, including smartphones, laptops, XR, robots, headsets, smart cars, intelligent Internet of Things terminals (AIT), etc. It is recommended to focus on side operating system layer software service companies.
Main line four, final wall side: 1) The screen-free wearable device "AIPin" supported by GPT4 was born, reshaping the form of human-computer interaction. Sales will begin in 2024, and it is expected to open up a new "lightweight" terminal large model A new era of hardware: 2) Tesla’s humanoid robot 202401 is expected to start ramping up production, and the era of embodied intelligence is coming: 3) The penetration rate of domestic smart cockpits is expected to exceed the global level, autonomous driving12 is increasing steadily year by year, and smart car sales are growing rapidly. Driven by demand for smart car software, China's smart cockpit sea penetration rate is expected to reach 66% in 2023, and global autonomous driving L2/L3 sea penetration rate is expected to reach 37%/1.5% in 2024. Starting from next year, with the expected breakthrough of 13 policies, L3 The sea penetration rate is expected to usher in a rapid upward trend. 4) The 5G smartphone market will resume growth in 2024, and Huawei's mobile phone growth momentum is expected to be the strongest. Focus on investment opportunities for components and software providers in different types of terminal industry chains.
Main line five, ecological side; Huawei breaks through to drive the transformation of the information innovation ecosystem, Hongmeng industrial chain ecology continues to prosper, Huawei leads the transformation of the information innovation industry ecology, "Ni Peng + Jie Teng" creates a dual alarm computing strategy, "Hongmeng + Euler" "Together we will build an important foundation for the development of the domestic operating system industry. Hongmeng is an operating system for a world where everything is connected. The number of ecological devices equipped with Huawei's Hongmeng system has exceeded 700 million. The number of Harmony0s developers has exceeded 2.2 million, and API calls have reached 59 billion times per day. , Hongmeng has become the world's third largest operating system. In 2024, with the improvement of national upstream production capacity support, Huawei's ecosystem will continue to select generations and expand. It is recommended to pay attention to Huawei's upstream and downstream partners in the industry chain.
The main line is large, and the scene side: AI+ education, finance, office, law, and medical care are expected to be implemented first. 1) AI+ education: Artificial intelligence technology has broad application scenarios in the field of education, and is one of the best scenarios for the implementation of artificial intelligence: 2) AI+ Finance: Vertical small models have high feasibility in specific fields, and domestic financial large models are seizing the ground: 3) AI+ Office: Artificial intelligence increases both the ARPU value and payment rate of office software: 4) A+ Law, generative A1 can Replaces 44% of legal work and empowers multiple application scenarios in the legal industry; 5) AI+ medical aids the transformation and upgrading of smart medical care. It is recommended to pay attention to leading companies in various sub-sectors.
Main line 1: The penetration rate of intelligent computing power on the computing power side is rapidly increasing
(1) Computing power side: technological innovation and policy are two-wheel drive, and AI computing power is booming
1. From the supply side, computing power continues to upgrade, and the "cloud-edge-end" architecture will emerge in the future.
In terms of computing power supply, it can be divided into general computing power, intelligent computing power and super computing power. The core of computing power is various computing chips such as CPU, GPU, FPGA, ASIC, etc., which are carried by computers, servers, high-performance computing clusters and various intelligent terminals. Massive data processing and various digital applications are inseparable from computing. The processing and calculation of force. The larger the computing power value, the stronger the comprehensive computing power. The commonly used unit of measurement is FLOPS (the number of floating point operations performed per second).
Computing power is the amount of information data that a device can process per second based on changes in its internal state. The development of computing power carriers has gone through the era represented by abacus and mechanical calculator to the server based on the basic platform of the Internet. In the past 20 years, as the richness of computing power carriers has been greatly improved, it has shown a trend of diversified development.
The computing power architecture can be disassembled into chips, equipment, and software, presenting an integrated "cloud-edge-end" pattern. In the future, a ubiquitous computing power deployment model will be formed in which the cloud side is responsible for large-scale complex calculations, the edge side is responsible for simple calculation execution, and the terminal side is responsible for perception and interaction.
2. Large models drive exponential growth in demand for intelligent computing power
Large models require powerful computing power to support the training process and inference process. According to OpenAI data, training the GPT-3175B model requires computing power of up to 3640PF-days (if one quadrillion floating-point operations are performed per second, 3640 days are required). Since 2018, the parameter scale of large models has reached hundreds of billions of parameters. The physical technology and number of cores of the CPU are close to their limits. In the AI era, CPU alone can no longer meet demand. The evolution of intelligent computing power through heterogeneous acceleration chips such as GPU, FPGA, ASIC, etc. has become a trend, and will eventually become the protagonist of computing power in the era of generative artificial intelligence.
According to Moore’s Law in the AI era, computing power doubles every 3.43 months on average. Since 2012, the computing power of the underlying machine learning technology that drives AI has increased exponentially. According to the OpenAI paper, in the early stage of deep learning, the doubling time of computing power was 21.3 months, and in the deep learning period, the doubling time of computing power was 5.7 months. , in the large model period, the computing power used for AI training tasks will double every 3.43 months, far exceeding the computing power increase brought by Moore's Law (transistors double every 18 months).
The computing power requirements of large models are mainly reflected in the following three scenarios:
(1) Pre-training computing power requirements: The model pre-training process is the main scenario that consumes computing power. ChatGPT uses a pre-trained language model. GPT-3 has approximately 175 billion parameters. GPT-4 is more than 10 times larger than GPT-3. It has approximately 1.8 trillion parameters, distributed in 120 layers, 13 trillion tokens, and OpenAI training The FLOPS of GPT-4 is about 2.15*10^25, and the computing power of a single NVIDIA A100 is 19.5TFlops (19.5 trillion floating point operations per second). If the utilization is not considered, training with 25,000 A100s takes 52 days. In fact, The case was trained on approximately 25,000 A100s for 90 to 100 days, with MFU (Mean Functional Utilization) between 32% and 36%. If the cost of OpenAI cloud computing is almost $1/100 hours, not including all experiments, failed training and other costs, such as data collection, RLHF (optimizing language models based on human feedback in a reinforcement learning method), manpower Cost, etc., the cost of this training is approximately $63 million.
(2) Computing power requirements for daily operations: ChatGPT is expected to require computing power of approximately 4874.4PFlops-days for single-month operation, with a corresponding cost of approximately US$18 million. After completing the model pre-training, ChatGPT's demand for underlying computing power has not ended. During daily operations, user interaction brings data processing needs. According to September data from the OpenAI official website, ChatGPT currently has more than 100 million users and generates 1.8 billion visits per month. According to Fortune magazine, each time a user interacts with ChatGPT, the computing power cloud service cost is about US$0.01. Based on this, we estimate that OpenAI’s monthly operating computing power cost for ChatGPT is $18 million.
(3) Computing power requirements for model tuning: From the perspective of model iteration, the ChatGPT model is not static, but requires continuous Finetune model tuning to ensure that the model is in the best application state. During the tuning process, on the one hand, developers need to adjust the model parameters to ensure that the output content is not harmful and distorted; on the other hand, based on user feedback and PPO (proximal policy optimization), the model needs to be modified on a large or small scale. Iterative training at scale. Therefore, model tuning will also bring computing power costs to OpenAI. The specific computing power requirements and cost amount depend on the iteration speed of the model.
3. Overseas: The scale of global computing power has entered a period of acceleration, and technology giants continue to increase AI capital expenditures.
With the global wave of artificial intelligence, the scale of global computing power has grown explosively. In 2022, the total scale of global computing power will reach 906EFlops, with a growth rate of 47%, of which the basic computing power scale (FP32) is 440EFlops, the intelligent computing power scale (converted to FP32) is 451EFlops, and the supercomputing power scale (converted to FP32) is 451EFlops. 16EFlops. According to China Mobile’s forecast, global computing power will grow at a rate of more than 50% in the next five years. The total computing power of global computing devices will exceed 3ZFlops by 2025 and will exceed 20ZFlops by 2030.
Applications led by AIGC performed strongly, driving the rapid and sustained growth of intelligent computing. IDC predicts that the global artificial intelligence computing market will grow from US$19.5 billion in 2022 to US$34.66 billion in 2026. Among them, the size of the generative artificial intelligence computing market will increase from US$820 million in 2022 to US$10.99 billion in 2026, and its proportion of the overall artificial intelligence computing market will increase from 4.2% to 31.7%. Generative artificial intelligence will promote the innovative development of industries such as the Internet, manufacturing, finance, education, and medical care.
Judging from the third quarter reports of overseas cloud giants, investment in artificial intelligence drives revenue and capital expenditures. Since the beginning of the year, the rapid rise of the generative AI technology wave has brought about a significant increase in the demand for AI computing power. Overseas cloud giants Google, Microsoft, and Meta (I will not consider Amazon for the time being, Amazon has reduced capital expenditures for warehousing and logistics in the third quarter) , with a larger impact) 3Q capital expenditures were US$21.205 billion, with total capital expenditures increasing by 9.43% quarter-on-quarter, mainly due to increased investment in AI infrastructure. Each company stated at the performance meeting that they would continue to invest in the AI field in 2024.
4. Domestic: The demand for intelligent computing power continues to grow, and the chip ban leads to a mismatch between supply and demand for high-end computing power.
Computing power has a significant effect on promoting the growth of GDP and digital economy. Relevant data shows that for every 1-point increase in the computing power index of the fifteen sample countries, the country's digital economy and GDP will grow by 3.6‰ and 1.7‰ respectively. This trend is expected to continue from 2023 to 2026.
China ranks second in the computing power index and is a leader. According to the "2022-2023 Global Computing Power Index Assessment Report", the first tier countries include China and the United States; the second tier countries include Japan, Germany, the United Kingdom, France, Canada, South Korea and Australia; the third tier countries include India, Italy, Brazil, Russia, South Africa and Malaysia. In 2022, China's computing power index increased by 1.4% year-on-year, reaching 71 points. In 2022, it was repeatedly impacted by the epidemic, and the annual GDP growth was lower than expected. In such a general environment, China's computing power index continued to grow.
The scale of China's core computing industry has increased rapidly, becoming an important driver of domestic GDP growth. According to data from the Ministry of Industry and Information Technology, as of the end of 2022, the scale of my country's core computing power industry has reached 1.8 trillion yuan, and the total scale of computing power has reached 180EFLOPS, with an annual growth rate of nearly 30%; the total scale of storage power has exceeded 1,000EB; among national hub nodes The one-way network delay is reduced to less than 20 milliseconds. It is estimated that the scale of China's computing power core industry in 2023. Every 1 yuan invested in computing power will drive 3 to 4 yuan in GDP economic growth.
2023 is the turning point for enterprises’ digital transformation, and capital expenditures are expected to continue to increase in 2024. Starting from 2022, global enterprises will begin to accelerate the digitalization process under the wave of digital transformation. 2023 will be the turning point of enterprise digital transformation. Enterprises will enter the digital business era from the digital transformation era and gradually enter a new stage of digitalization. According to IDC research, by the end of 2023, global digital transformation spending will account for 52% of overall enterprise ICT spending, and 52% of global software application spending will also be in the SaaS model. Global spending on digital transformation technologies is expected to grow by 16.9% in 2023. Digital transformation has begun to show results in reducing costs, increasing efficiency, improving innovation capabilities, and transforming and upgrading business models, and has become the core development strategy of enterprises.
The overall scale of the domestic computing power industry is expected to maintain a CAGR of around 30% in the next three years, and the penetration rate and proportion of intelligent computing power are increasing rapidly. In recent years, my country has continuously increased its investment in infrastructure such as computing, network and storage, and attached great importance to the high-quality development of computing infrastructure such as data centers, intelligent computing centers, supercomputing centers and edge data centers. In the past five years, The average annual growth rate of my country's computing power industry exceeds 30%.
The penetration rate of intelligent computing power is gradually increasing. Intelligent computing power is growing rapidly, and intelligent computing power has become a new engine of growth in the new computing power. By the end of 2022, the total scale of China's computing power has reached 180EFLOPS, of which the scale of intelligent computing power has increased by 41.4% compared with last year, exceeding the global overall intelligent computing power. Growth rate (25.7%), of which the general computing power scale is 137EFLOPS, accounting for approximately 76.7%, and the intelligent computing power scale is 41EFLOPS, accounting for approximately 22.8%. According to the "2022-2023 China Artificial Intelligence Computing Power Development Assessment Report", the compound annual growth rate of China's intelligent computing power scale will reach 52.3% in the next five years, and the compound annual growth rate of general computing power scale will be 18.5%. It is expected that China’s intelligent computing power will reach 145EFLOPS by 2026, accounting for 36.7%. With the rapid development of large AI models, the demand for intelligent computing power is showing explosive growth, and the penetration rate will increase significantly.
(2) Main lines of investment on the computing power side: localization, high-bandwidth storage, AI servers, liquid cooling, computing power leasing
1. Export ban forces localization to accelerate, Huawei Ascend VS Nvidia parameter comparison
The export ban affects overseas supply and forces domestic substitution to accelerate. On October 17, 2023, the U.S. Department of Commerce’s Bureau of Industry and Security (BIS) issued new export ban regulations on chips, which more strictly restricts China’s purchase of important high-end chips. On the one hand, since the launch of ChatGPT, domestic companies and research institutes have launched more than 130 large models in just over half a year. Among them, leading players have begun to apply large models to specific scenarios to create popular applications. . On the other hand, in order to build a computing power base, local governments have launched the construction of intelligent computing centers to lay the information highway in the big data era, promote industrial innovation and upgrading, and reduce the cost for enterprises to use scientific and technological achievements represented by large models.
2. The “memory wall” restricts the release of computing power, and the volume and price of HBM high-bandwidth storage are rising.
Currently, GPUs have an absolute advantage in training and inference among AI chips. AI chips, also known as AI accelerators or computing cards, are modules specifically designed to handle a large number of computing tasks in artificial intelligence applications. AI chips are the core components of AI servers, accounting for nearly 70% of the value of AI servers. Currently, mainstream AI computing chips mainly include CPU, GPU, FPGA, ASIC, etc. Among them, GPU is a relatively mature general-purpose artificial intelligence chip, while FPGA and ASIC are semi-customized and fully customized chips based on the characteristics of artificial intelligence needs. GPU, FPGA, and ASIC serve as acceleration chips to assist the CPU in large-scale calculations.
Main line 2: On the data side, data elements release value three times and enter the table to promote the "explicitness" of value.
(1) Data human form: Implemented in early 2024, the era of data capitalization has officially begun
Data as intangible assets are not exclusive, while data as inventory are exclusive. In most scenarios and enterprises, enterprises use data resources repeatedly. The usage scenarios include but are not limited to self-use data to generate product sales to generate economic value, sharing data to customers to generate economic value, etc. Data resources are used by multiple parties. , does not have the exclusive right to use and should be an intangible asset. Inventory scenarios mainly include data collection and processing manufacturers. The purpose of generating data resources is simply for sales and the transfer of data ownership. After the sale, the data no longer belongs to the original enterprise and is exclusive.
(2) There is still room for refinement of data element policies, and the policy catalytic effect is expected to be highlighted
The system formulation in the Twenty Data Articles runs through the entire life cycle of the data industry chain, and proposes four core data basic systems, including the data property rights system, the data element circulation and transaction system, the data element income distribution system, and the data element governance system. Among these four systems, the data property rights system is the foundation, the circulation and transaction system is the core, the income distribution system is the driving force, and the governance system is the guarantee. These four basic systems are the "four beams and eight pillars" of the sustainable development of my country's data economy and the cornerstone of the development of data elements.
(3) Data elements are ushering in three value releases, and the data capital space is expected to reach 30 trillion
The three value releases are gradual, with the latter one based on the previous one. At present, my country's first two data value release environments have gradually matured. We believe that the third value release is different from the past, focusing on the circulation of data from inside the enterprise to the outside, making data flow better from the supply side to the demand side, maximizing the production efficiency and value of data as a production factor. In addition, the three value releases will generate new data-related technologies, and the industrial chain will be further enriched and improved.
(4) Analysis of segmented investment opportunities: “explicitness” of value throughout the entire industry chain
Data governance is inseparable from high-quality data collection and annotation. From the perspective of the AI data governance industry chain map, the upstream is mainly the provider of data, and the downstream is mainly the final application party of data. The data used for training and inference of AI models are mainly collected and annotated by midstream basic data service providers, and the data governance platform optimizes and manages the data.
At present, there has been corresponding exploration and implementation on the industrial side. People's Data, a subsidiary of People's Daily Online, issues the "three certificates" of data elements in accordance with the separation of three rights mentioned in the "Twenty Articles on Data" - "Data Resource Ownership Certificate", "Data Processing Use Rights Certificate" and "Data Product Operation Rights Certificate" . The "Three Certificates" are based on the People's Chain Baas service platform (version 2.0), which performs rights confirmation, chaining, certificate storage, and transaction services. The "three certificates" are mainly aimed at unifying and connecting relatively scattered data between party and government agencies at all levels and big data exchanges, forming a national data trading service platform, and at the same time solving the problem of unclear data rights.
Total report: 103 pages