(Report produced by: China Galaxy Securities)
First, across the singularity, artificial intelligence ushered in the sea of stars
The fifth round of Combo is heading for depression, and the sixth round of the Combo cycle driven by AI is about to begin
Kondratievwave: As an important model to describe the intrinsic relationship between macroeconomic growth and the technological revolution, its five occurrences in human history have been accompanied by iconic technological singularities and leapfrog events. The fifth round of Combo is heading for depression: the main driving force of historical progress over the past two hundred years has been through technological-productivity breakthroughs. Since the outbreak of the Industrial Revolution, there have been four rounds of Combo cycles (Industrial Revolution - Railway Age - Electric and Heavy Industry Age - Petrochemical and Automotive Era), and the fifth round of Combo (Information Technology Internet Age) has gone through depression. It is widely believed that we are in the bust phase of the fifth Combo cycle, which is generally characterized by a lack of economic growth momentum and the technology sector waiting for new transformative outcomes. Thanks to the shortening of the cycle brought about by accelerated informatization, the fourth round and the fifth round of Combo have significantly crossed, and we have reason to think that in the late stage of the fifth round of Combo, when Internet information technology tends to decline, the sixth round of Combo led by artificial intelligence and new energy technology has also shown the dawn.
The "iPhone moment" of the global AI community, ChatGPT opens a new era
The singularity is approaching, and the era of strong artificial intelligence is coming. Ray Kurzwell, a former Google engineer and the biggest proponent of the singularity theory, believes that artificial intelligence will pass the Turing test in 2029, exhibiting intelligent behavior comparable or indistinguishable from humans, and in 15 years humans will merge with the artificial intelligence they have created, making human intelligence a million times higher. The arrival of the singularity means that human beings are about to enter the era of strong artificial intelligence, machines reach "strong artificial intelligence", IQ exceeds humans, Baichuan intelligent Wang Xiaochuan believes that this day will come as early as 2033. ChatGPT ushers in a new era, and AIGC is becoming the engine that triggers productivity change. With the technological development of artificial intelligence driving the update and iteration of GPT, ChatGPT has significantly improved its model, function call capability, text capacity and price. In the future, the new era of artificial intelligence opened by ChatGPT will continue to develop towards multimodal integration.
AIGC multi-scenario applications are on the eve of the outbreak, and the trillion market is open
Language applications of AIGC: AIGC technology has been gradually commercialized from the initial laboratory product and deeply integrated with other technologies and fields to achieve more intelligent content generation and creation. For language models such as the GPT series, AIGC can implement text generation, text classification, and text processing. Combined with NLP, it can be applied to intelligent customer service, pre-sales/after-sales consultation, real-time language translation and other fields.
At present, the fields covered by AIGC basically belong to AI software and services. The global AI software market size is expected to exceed $120 billion by 2025, with a CAGR of 38.01% from 2021 to 2025. Although AIGC has involved a variety of content generation forms such as text, images, code, voice, video, 3D, and games, there are gaps in the technical maturity and underlying technical difficulty in each field, and the penetration rate is different. In 2021, the proportion of AI production will only account for about 1%-2% of the market, and with the development of technology, we predict that in 2025, AIGC's technology in the field of online literature, text analysis, painting and pictures, and digital music will be relatively mature, with penetration rates of 70%/60%/60%/50% respectively; The penetration rate of the network video field and the game field is relatively low due to technical reasons, about 30%/25% respectively.
The big model has entered the Warring States Period, and fierce competition has arrived
ChatGPT's "cost reduction and efficiency increase" will further affect the access cost of existing generative AI models in the market. On June 13, 2023, OpenAI released a new feature update for ChatGPT. Added new feature calls and price changes to the Chat Completions API; and model updates, where GPT-3.5-turbo-16k provides four times the text capacity of the standard 4k version, supporting up to 20 pages of text in a single request. In terms of price, the price of embedded models has been reduced by 75%, and the latest price is 0.0001 US dollars per 1k tokens; GPT-3.5-turbo's token input cost is reduced by 25%; GPT-3.5-turbo-16k tokens input price is $0.0003 per 1k tokens, and output price is $0.0004 per 1k tokens. At present, the existing generative AI models in the market are highly competitive, and OpenAI's move will greatly improve the competitiveness of ChatGPT.
A brief analysis of future investment opportunities in AI
Application: Artificial intelligence empowers thousands of industries, and multi-field integration is possible. The ability of artificial intelligence to generate text, code, images, videos and other content continues to improve, making it not only applied in mature fields such as finance, medical care, government affairs, education, and law, but also in logistics, manufacturing, energy, public utilities and agriculture, and innovative application scenarios are gradually increasing. In the next five years, as human-computer interaction, machine learning, computer vision, and speech recognition technologies reach a more mature stage, artificial intelligence will move from single-point technology application to the integration of multiple artificial intelligence capabilities, post-analysis to pre-judgment and active execution, and from computational intelligence and perceptual intelligence to cognitive intelligence and decision-making intelligence.
Second, artificial intelligence drives the release of demand for the entire industrial chain from computing power to application
Computing power: The development of AIGC has led to the release of demand for AI servers
The demand for computing power of AI large models is higher than the iteration speed of Moore's Law, and the server demand will increase. According to OpenAI calculations, since 2012, the training computing power of the head AI model has doubled every 3-4 months, and the training computing power has increased by up to 10 times. The slowdown of Moore's Law has become an indisputable fact, but the amount of data continues to increase, the growth rate of computing power has lagged significantly, and AI servers as AI computing power carriers, demand is gradually released.
Algorithm side: the accumulation of basic algorithms
Basic algorithm: The cornerstone of AI large models, and technology accumulation starts from the foundation. Most of the common basic algorithms of mainstream AI artificial intelligence come from foreign teams, and similar algorithms in China usually appear about a few years later. The accumulation of algorithms to models is a long-term process, and each upgrade iteration relies on the strong mathematical foundation and abstract ability behind the model. China started late in terms of basic algorithms, and although the process of catching up is currently accelerating, there is still a certain gap with foreign countries.
Data side: High-quality data is a necessary element for AI large model training
High-quality and large-scale datasets are the key points for the breakthrough of AI large models. At present, domestic data resources are abundant and have advantages in terms of total volume, but due to insufficient data mining, data cannot circulate freely in the market, and the industrial chain of data cleaning and labeling is not perfect, high-quality data sets are still scarce.
High-quality datasets are inseparable from data governance, and the AI data governance market is vast. According to iResearch data, in 2021, the scale of data governance for artificial intelligence in China will be about 4 billion yuan, and the scale will exceed 10 billion yuan to 10.5 billion yuan in 2026, with a five-year CAGR of 21.3%; China's data governance market is about RMB12.1 billion and is expected to reach RMB29.4 billion in 2026, with a five-year CAGR of 19.5%.
Application: AI+Medical
AI has a wide range of application scenarios in the medical field, mainly divided into pre-diagnosis, intra-diagnosis and post-diagnosis. The information process mainly involves drug research and development, appointment visits, guidance, clinical diagnosis and treatment, electronic medical records, drug retrieval, medical payment, drug collection and report acquisition, post-diagnosis follow-up and rehabilitation management, etc.
According to Frost & Salivon data, the AI+ medical market will show high growth from 2020 to 2025, and the total market size will reach 34.8 billion yuan in 2025, maintaining a growth rate of about 40%. According to the data of the data center of Guanyan, in the AI artificial intelligence segment, imaging, data exchange and storage, and comprehensive auxiliary diagnosis accounted for relatively high, accounting for 34%, 22%, and 13% respectively.
Application: AI+ office productivity tools
Microsoft announced its blockbuster product Microsoft 365 Copilot at the March 16 conference, which integrates GPT-4 functions and appears on the right side of the product in the form of a chatbot. By giving them instructions, users can automatically generate text, tables, presentations, and more. With the empowerment of the large model, the user's learning path is simplified, and the machine can interact with the machine and issue instructions through natural language, greatly simplifying the operation process, and greatly improving the work process and efficiency.
Domestically, Kingsoft Office followed suit to release Kingsoft WPS AI, which has the functions of generating documents, multi-round dialogues, processing existing content, and generating presentations. We believe that Kingsoft Office has advantages in the competition pattern of office software, and the payment rate of users and the ARPU value of products will be greatly improved with the blessing of AI large models.
Application: AI+Finance
On March 30, 2023, Bloomberg released its large-language model, BloombergGPT, which is a symbol of the acceleration of LLM in the vertical finance scene. BloombergGPT focuses on the financial scene, and trains LLM on reporting terms in the financial field accumulated over the years, which is more efficient and accurate in its understanding of financial terminology than universal ChatGPT products.
BloombergGPT uses 50 billion parameters and 700 billion datasets to train large models, which is much smaller than GPT3's 175 billion parameters and 45TB of data. Its data set mainly contains two parts, one is a general text dataset with a data volume of 49%, mainly including general common sense datasets such as The Pile, C4, Wikipedia, etc., and the other is a financial expectation data set with a data volume of 51%, mainly including news, research reports, company financial reports, financial documents crawled by the Internet and extracted social media messages. According to the paper released by Bloomberg, BloombergGPT's ability in the general domain is almost equal to GPT3, but the text writing and Q&A capabilities are more prominent in the financial vertical. The success of BloombergGPT LLM pre-training shows the availability of LLM in vertical applications and the feasibility of controllable training costs.
Application: AI+ education
Policies have been released one after another, and large models are being applied. In February 2023, the Central Committee of the Communist Party of China and the State Council issued the "Overall Layout Plan for the Construction of Digital China", which mentioned vigorously implementing the national education digitalization strategic action and improving the national smart education platform. In May 2023, the Ministry of Education and the 18th Department issued the "Opinions on Strengthening Science Education in Primary and Secondary Schools in the New Era", which mentioned that it is necessary to explore the use of artificial intelligence, virtual reality and other technical means to improve and strengthen experimental teaching and make up for the lack of high-quality education and teaching resources.
NetEase Youdao "Ziyue", Good Future, iFLYTEK "Spark", Taoyun Technology and other education informatization companies have successively released or announced that they will release large models for education, and the application of AI in the field of education is expected to accelerate.
3. Market review and performance summary in the first half of the year
Market review: The industry performance is eye-catching, and the market is expected to be optimistic
The computer industry significantly outperformed the market, with the third largest increase in the industry in the medium term. As of June 30, the computer industry has risen 27.57% year-to-date in 2023, ranking 3rd among 31 industries in Shenwan, while the Shanghai Composite Index rose 3.65%, the CSI 300 fell 0.75%, and the ChiNext Index fell 5.61%. Since 2023, digital China and data element policies have been intensively introduced, vigorously supporting the development of the computer industry, and the emergence of AI technology and increasingly diversified application scenarios have driven the formation of a dual stimulus of policy + market, and the computer industry has performed brightly, significantly ahead of the Shanghai Stock Index, CSI 300, and ChiNext Index.
In terms of valuation, PE in the computer industry has continued to recover since bottoming out in April 2022 and has maintained an upward trend this year, reflecting the market's optimistic expectations for the computer industry. As of June 30, the overall valuation level (PE, TTM) of the computer industry was 69.69 times, which was 56.27 times higher than the central level.
Performance analysis: Revenue and profit were under pressure and expenses remained stable
The growth rate of industry revenue has slowed down, and the fundamental repair can be expected throughout the year. In 2023Q1, the computer industry achieved operating income of 221.327 billion yuan, and the quarterly revenue growth rate decreased by 2.4% year-on-year; The net profit attributable to owners was 6.218 billion yuan, a year-on-year increase of 54.8% in a single quarter, mainly due to the change in the fair value of non-recurring profit and loss, and the net profit attributable to non-recurring owners was 21 million yuan, down 99.1% year-on-year. In terms of profit, the gross profit margin and net profit margin in 2023Q1 were 25.1% and 2.9%, respectively, both slightly recovered year-on-year. Although the first quarter performance is still under pressure, with the macroeconomic recovery and the recovery of downstream B-side and G-end demand, the full-year fundamental repair can be expected.
The expense side remains stable and profit elasticity is expected to be released. The total expenses of the computer industry during 2023Q1 were 34.648 billion yuan, a year-on-year increase of 6.38%; R&D expenses totaled RMB25.212 billion, a year-on-year increase of 9.46%; During the period, the expense ratio and R&D expense ratio were 15.14% and 11.02%, respectively, which rebounded slightly compared with the same period last year. We believe that the overall cost side is stable, and it is expected that the cost reduction of fees and the recovery of industry demand are expected to further release profit margins.
Excerpts from the report:
(This article is for informational purposes only and does not represent any investment advice from us.) For information, please refer to the original report. ）
Selected report source: [Future Think Tank]. 「Link」