Intensified Competition in AI Chips: China's 2023 AI Industry Development Implementation Plan
-
The artificial intelligence (AI) industry has witnessed new developments. On December 6, U.S.-based Advanced Micro Devices (AMD) hosted its 'Advancing AI' launch event, introducing the much-anticipated MI300 series chips. These new products are designed for training and running large language models, offering improved memory capacity and higher energy efficiency compared to previous generations. AMD CEO Lisa Su even described one of the products, the Instinct MI300X accelerator, as 'the world's highest-performing accelerator.' The rapid advancement of AI applications is fueling increasingly fierce competition in the chip industry.
The Instinct MI300X accelerator consists of eight MI300X GPUs. GPUs, or advanced graphics processing units, were initially designed for graphics rendering and processing but were later found to excel in parallel computing, making them ideal for accelerating various computation-intensive applications. Compared to CPUs and FPGAs, GPUs represent the latest generation of AI training chips. Nvidia's flagship GPU chip, the H100, has sold hundreds of thousands of units in a year, contributing to the company's record-breaking performance. In Q3, Nvidia reported revenue of $18.12 billion, a 206% year-over-year increase.
AI data models place enormous demands on high-performance, high-computing-power AI chips. Predictions suggest that the AI chip industry will exceed $400 billion in the next four years. Driven by strong demand, tech giants like AMD, Intel, and IBM are all developing AI chips, while companies such as Google, Microsoft, Alibaba, and Baidu are also investing in in-house chip development. AMD's newly launched MI300X chip, with over 150 billion transistors and 2.4 times the memory of Nvidia's H100 chip, is highly anticipated. It performs comparably to Nvidia's H100 in training large language models but excels in inference tasks. AMD's new product has already garnered interest from major clients, including Microsoft, Meta, and Oracle, who have announced plans to adopt AMD's chips.
The initial layout for artificial intelligence computing power has taken shape, with domestic AI chips, deep learning frameworks and other foundational software/hardware products significantly increasing their market share. Computing power chips and other core components have basically achieved independent controllability. The proportion of domestic hardware has risen notably, achieving full compatibility with domestic deep learning frameworks. The interconnection of AI computing power resources promotes high-quality independent development of foundational software/hardware.
2023 China Artificial Intelligence Industry Development Plan
Recently, the Beijing Municipal Government issued the "Implementation Plan for Accelerating the Construction of a Globally Influential AI Innovation Hub in Beijing (2023-2025)". The plan proposes to fully leverage Beijing's innovation resource advantages in AI, continuously enhance global influence, and further promote AI's leading development.
By 2025, the target is to achieve a core AI industry scale of 300 billion yuan, maintaining over 10% growth, with the radiation effect exceeding 1 trillion yuan. Leading AI enterprises will continue increasing R&D investment, startup numbers will keep growing, and the total number of enterprises will maintain domestic leadership, with 5-10 new unicorn companies cultivated. The depth and breadth of AI applications will further expand, with generative products becoming mainstream applications and ecosystem platforms in the domestic market, driving high-end industrial development.
It also proposes promoting breakthroughs in domestic artificial intelligence chips. To meet the demand for distributed training in AI cloud environments, the development of general-purpose high-computing-power training chips will be pursued. For low-power requirements in edge application scenarios, the focus will be on developing multi-modal intelligent sensing chips, autonomous intelligent decision-making execution chips, and high-efficiency heterogeneous intelligent chips for edge devices. For innovative chip architectures, exploration will include reconfigurable, in-memory computing, brain-inspired computing, and Chiplet approaches. Efforts will be made to actively guide large model development companies to adopt domestic AI chips, accelerating the increase in the localization rate of AI computing power supply.
Artificial Intelligence (AI) Product Market Share
Recently, investment bank JPMorgan stated in its investment report that Nvidia is expected to capture up to 60% of the artificial intelligence (AI) product market this year, thanks to its hardware products such as GPUs and networking solutions.
It is reported that due to cyclical slowdowns in its gaming division, Nvidia's revenue for the first quarter of fiscal year 2024 fell by 13% year-over-year to $7.19 billion. However, during the same period, its data center business revenue reached a record $4.28 billion, a 14% increase year-over-year, accounting for 60% of its total revenue. Gaming business revenue was $2.24 billion, down 38% year-over-year, representing 31% of total revenue.
Currently, Nvidia leads in the AI field, holding approximately 80% of the AI processor market share. Its high-end processors are widely used for training and running various chatbots. The company is highly favored by investors and is considered a key supplier for meeting AI computing power demands.
The AI industry chain typically consists of the upstream data and computing power layer, the midstream algorithm layer, and the downstream application layer. Recently, the market has paid more attention to the upstream industry chain, particularly the computing power sector. AI hardware has emerged with many new investment opportunities, as AI software applications rely on the computing power provided by hardware.
Domestic AI Computing Power Demand Will Maintain Growth Momentum
Driven by the continuous influence of ChatGPT, domestic AI computing power demand will maintain its growth momentum, and computing power server manufacturers are expected to benefit. It is estimated that ChatGPT's total computing power requires 7 to 8 data centers with an investment scale of 3 billion yuan and a computing power of 500P to support its operation. In the era of the digital economy, global data volume and computing power scale will show rapid growth.
With the simultaneous rise in demand for AI servers and AI chips, it is estimated that the shipment volume of AI servers (including those equipped with GPUs, FPGAs, ASICs, and other main chips) will reach nearly 1.2 million units in 2023, a year-on-year increase of 38.4%, accounting for nearly 9% of total server shipments. By 2026, this proportion is expected to further increase to 15%. The institution has also revised its annual compound growth rate forecast for AI server shipments from 2022 to 2026 to 22%, while AI chip shipments in 2023 are expected to grow by 46%.
The institution stated that NVIDIA GPUs have become the mainstream chips used in AI servers, with a market share of about 60-70%, followed by ASIC chips independently developed by cloud computing companies, with a market share of over 20%.
Compared to general-purpose servers, AI servers use multiple accelerator cards, and their PCBs adopt high-layer HDI structures, which are of higher value. Additionally, the number of motherboard layers is much higher than that of general-purpose servers. The PCB value of AI servers is 5-6 times that of ordinary servers.
NVIDIA's founder and CEO Jensen Huang announced during his speech at NVIDIA Computex 2023 that the generative AI engine NVIDIA DGX GH200 has now entered mass production. From the demonstration, it can be observed that the newly released GH200 server architecture has undergone significant changes compared to the DGX H100. The GH200's PCB modifications include the removal of one UBB and one CPU motherboard, while adding three NVLink module boards. Additionally, the accelerator card performance has been substantially improved, which should increase the per-unit PCB value. This indicates that AI upgrades will continue to drive value growth in the PCB sector.
Computing power resources serve as a critical foundation for the development of the digital economy. The emergence of new digital phenomena, business models, and patterns has diversified application scenarios, leading to a continuous surge in demand for computing power. According to data from the Ministry of Industry and Information Technology, the total scale of operational data center racks in China exceeded 6.5 million standard racks in 2022. Over the past five years, the total computing power has grown at an average annual rate of over 25%. As computing power is applied across various industries, different precision levels of computing power need to "adapt" to diverse application scenarios. Particularly with the rapid advancement of artificial intelligence technology, the structure of computing power is evolving, and the demand for intelligent computing power is increasing daily.
From a policy perspective, China places high importance on the development of the AI industry, and the foundation for intelligent computing power is gradually being strengthened. In February 2022, four ministries and commissions jointly issued a notice approving the launch of national computing power hub nodes in eight regions and planning ten national data center clusters. This completed the overall layout design of the national integrated data center system. With the full implementation of the "East Data West Computing" project, the construction of intelligent computing centers has also entered a new phase of accelerated development. Data centers, as hubs for data and application carriers, form the foundation for AI development. In the long term, demand for data centers is expected to recover. It is projected that the IDC market will reach 612.3 billion yuan by 2024, with a compound annual growth rate of 15.9% from 2022 to 2024, indicating that data centers will enter a new upward cycle.
The Future Direction of Artificial Intelligence
Artificial intelligence encompasses many concepts, some of which are difficult to measure and verify. For instance, while machines can output representations of societal norms or responsibilities, it's challenging to validate whether they truly understand these concepts. Therefore, focusing on verifiable and measurable concepts creates a closed-loop system, and embodied intelligence perfectly fits this framework—serving as an excellent starting point toward general AI.
The rapidly evolving large AI models hold the potential to break through limitations and endow robots with "intelligence."
Robotic large models include LLM (Large Language Models), VLM (Vision-Language Models), and VNM (Visual Navigation Models). The AI "brain" of robots isn't limited to the linguistic models used in ChatGPT. As highlighted in Google's LM-Nav research, combining LLM, VLM, and VNM enables a pathway from natural language (redundant verbal descriptions) to text (landmark strings) to images (locating objects in images based on text), ultimately generating robotic path planning. This behavioral model allows robots to interact with humans while achieving a degree of adaptability.
Recently, Professor Lu Cewu from Shanghai Jiao Tong University delivered a keynote speech titled "Embodied Intelligence" at the Machine Heart AI Technology Annual Conference. He proposed the PIE framework, suggesting that embodied intelligence consists of three modules: embodied Perception, embodied Imagination, and embodied Execution, which could accelerate the practical implementation of embodied intelligence.
Currently, the integration of AI and robotics appears to be the most viable landing point for "embodied intelligence."
Compared to non-intelligent conventional humanoid robots, embodied intelligence exhibits significantly higher operational efficiency. Its capabilities in comprehension, interaction, and task planning grant it strong practical applicability as robots penetrate diverse industries. Moreover, its natural language control functionality serves as a critical prerequisite for future large-scale assistance to human workers.
Therefore, attention should focus on hardware robot types and application scenarios currently adaptable through large model transformations, including dialogue-oriented service robots, industrial robots, and humanoid robots operating in complex environments.
Major tech companies have already begun strategic deployments in embodied intelligence: Google released PaLM-E, its largest generalist model to date; Microsoft is exploring ChatGPT's expansion into robotics; Alibaba is testing its Qwen large model with industrial robots.
Among these developments, Tesla's humanoid robot Optimus stands out remarkably.
From its initial October 2022 debut requiring human assistance to walk, to its May 17 demonstration at Tesla's shareholder meeting where it showcased agile movement, object manipulation, environmental mapping, motor torque control, AI training through human motion tracking, and partial Full Self-Driving (FSD) algorithm integration - Optimus has achieved notable technological breakthroughs.