AIGC (AI Generated Content) is generative AI, which is considered to be a new type of content production method after professional production content (PGC) and user generated content (UGC), which refers to the use of artificial intelligence technology to generate content, including text, pictures, audio, video, etc. ChatGPT is an intelligent chatbot program, and ChatGPT is a manifestation of AIGC. AIGC involves unsupervised and semi-supervised learning algorithms, and its development process is mainly divided into three stages: statistical machine learning method stage (before 2010): first manually annotate the data, then build its important features, and finally build a probabilistic model and optimize the parameters, so as to take the output with the highest probability as the result; Neural network model based on deep learning (2010-2017): Deep learning algorithms were introduced, essentially training neural networks through a large amount of data, mainly in the form of: CNN (convolutional neural network), RNN (recurrent neural network), etc. Compared with statistical learning methods, complex and manual feature construction is eliminated. Pre-trained model based on Transformer structure (2017-present): self-supervised learning with a large amount of unlabeled data, and then fine-tuning downstream tasks (i.e. transfer learning) using a small amount of labeled data. Data, computing power, and algorithms are the core elements of AIGC.
Where does AIGC come from
At present, most of the achievements in AI development are closely related to big data. Through data collection, processing, and analysis, valuable insights can be obtained from massive data from all walks of life, providing materials for more advanced algorithms, and at the same time, the emergence of artificial intelligence has also improved the breadth of available data, through the combination of computing power and algorithms. As an emerging factor of production, data owners and processors are the foundation of industrial development. Leading domestic Internet manufacturers such as Tencent, Alibaba, and Baidu have the ability to continue large-scale investment in R&D and computing power investment, and on the other hand, they are also owners of massive data, and are expected to launch global Chinese language models in the future.
Computing Power as infrastructure, large language models and AIGC will bring a sharp increase in the demand for underlying computing power, is the main beneficiary of AIGC capital expenditure, core players NVIDIA, AMD has significant competitive advantages. Computing power-related manufacturers include chip manufacturers, server manufacturers, data centers and cloud service vendors, and from the perspective of market demand and supply chain security, chip manufacturers benefit from the forefront, including CPU, GPU, FPGA, ASIC and other segments.
Compared with ASIC chips, GPGPUs have stronger versatility. In the mainstream AI acceleration chip market, GPGPU accounts for 90% of the market share. The traditional small-volume model relies on CUDA, so GPGPU is more suitable, while the large model has less dependence on the CUDA ecosystem, so the gap between GPGPU and ASIC is not obvious. However, domestic large model training can only be completed by GPGPU, and ASIC is not mature enough. The first echelon of Haiguang chips (Shenshuan No. 1) can run a general large model, but the efficiency is relatively poor. The second place should be Huawei's Ascend 910, but it can only run Huawei's own optimized large model. Cambrian can only run inferences for large models. Jingjiawei does not belong to this market. Muxi's C100 is expected to perform against NVIDIA's H100, and Wall's BR100 is limited by the US Department of Commerce. At present, the most anticipated manufacturer is Mu Xi. The hardware threshold of AI chips is not high, and the software threshold is high. Several core patents are not accumulated domestically, so they will be restricted by the United States.
In addition to the strong demand for computing power hardware such as CPU/GPU, the network side has also spawned greater bandwidth requirements to match the increasing traffic. There may be some variations in the network architecture of AI data centers compared to the network architecture of traditional data centers. In traditional data centers, the network side mainly includes traditional three-tier architecture and leaf ridge architecture, and in AI data centers, non-blocking networks have become one of the important requirements due to large internal data traffic. In NVIDIA's AI data center, a fat-tree network architecture is used to achieve non-blocking functions, in which the storage side is independently networked, which is separate from the computing side network architecture, and also requires a certain number of switches and optical modules. As a result, the number of switches and optical modules in AI data centers has increased significantly compared to traditional data centers.
The development of artificial intelligence will put forward higher requirements for computing power, and the demand for computing power network infrastructure is expected to continue to increase. According to data from the China Academy of Information and Communications Technology, the total computing power of global computing devices reached 2021EFlops (floating point operations per second) in 615, a year-on-year increase of 44%, of which the basic computing power scale is 369EFlops, the intelligent computing power scale is 232EFlops, and the supercomputing power scale is 14EFlops, and it is expected that the global computing power scale will reach 2030ZFlps with an average annual growth of 56% in 65. The scale of China's intelligent computing power continues to grow rapidly, and the scale of intelligent computing power has exceeded that of general computing power in 2021. According to data from the China Academy of Information and Communications Technology, the total scale of computing equipment computing power in China has reached 202EFlops, accounting for about 33% of the world, maintaining a high-speed growth trend of more than 50%, and the growth rate is higher than that of the world, of which intelligent computing power has grown rapidly, with a growth rate of 85%, accounting for more than 50% of China's computing power.
Algorithm is the technical barrier of AIGC, the current general AI is led by GPT, and in the segment, the main players in the industry include Google, Meta, Anthropic, Hugging Face and Baidu and other companies. As subdivision leaders compete to develop innovative algorithms and optimize existing technologies, as well as the rapid expansion of the demand for data and computing power under model iteration, the technical barriers of the AIGC industry will continue to increase, and the existing excellent participants have a deep moat.
AIGC has huge market potential, and the application field ushered in productivity liberation. The global AI software market size will reach $2025 billion by 1260, with a compound annual growth rate of 2021.2025% from 41 to 02. The hot secondary market also reflects the certainty trend of AIGC development. Driven by the rapid iteration of large models, industries with applications such as search engines, office software, automobiles, media, AI painting design, AI advertising and marketing, and intelligent work assistants will have strong commercialization opportunities. Among them, whether there is overseas mapping, whether APIs can be accessed, and whether the scenario is fault-tolerant will become key considerations.
Where AIGC is going
The fire of ChatGPT has brought AIGC technology and related applications "strong out of the circle". In addition to lamenting AI's super content generation and output capabilities, people from all walks of life have also begun to think about the potential risks that AIGC may generate. On March 3, according to CCTV, the Italian Personal Data Protection Agency said that it had launched an investigation into the alleged violation of data collection rules by OpenAI chatbot ChatGPT, banned the use of ChatGPT from now on, and temporarily restricted OpenAI's processing of Italian user data. It is reported that OpenAI, through its representatives in Europe, must inform the Italian Personal Data Protection Agency within 31 days of the measures taken by the company to implement the requirements of the Protection Office, otherwise it will be fined up to 20 million euros or 2000% of the company's annual global turnover. In addition, Europol, the European Union's law enforcement agency, warned that ChatGPT could be misused for phishing, disinformation and cybercrime; The nonprofit Center for Artificial Intelligence and Digital Policy (CAIDP) also complained to the Federal Trade Commission (FTC) that GPT-4 is "biased, deceptive, and poses a risk to privacy and public safety." The formation and improvement of AIGC models rely on a large amount of data training, and the data used for training often contains legally protected content. How to standardize the development of AIGC is the next step of "strong out of the circle", and optimizing and improving model design is one of the important ways for AIGC to avoid potential risks.
