The field of artificial intelligence has been experiencing an explosive development in recent years, led by generative AI large models. On November 30, 2022, OpenAI introduced an AI chatbot named ChatGPT, whose exceptional natural language generation capabilities have garnered widespread attention globally. It surpassed 100 million users within two months, sparking a wave of large model enthusiasm both domestically and internationally, with various large models such as Gemini, Wenxin Yiyan, Copilot, LLaMA, SAM, SORA, and others emerging like mushrooms after a spring rain. The year 2022 has also been hailed as the inaugural year of large models.

Artificial intelligence is thus regarded as a revolutionary technology with significant strategic importance to governments worldwide. Data indicates that this year, the adoption rate of generative AI by enterprises in China has reached 15%, with a market size of approximately 144 trillion yuan. The adoption rate of generative AI technology has seen rapid growth in four major industries: manufacturing, retail, telecommunications, and healthcare.

Advertisement

As one of the three key elements driving the development of artificial intelligence, computing power is referred to as the "engine" and core driving force of AI. Computing power refers to the computational ability of a device to process data and achieve specific output results. Chen Yuanmou, a senior analyst at the Strategy Research Institute of China Telecom Research Institute, stated that for every one-point increase in the computing power index, the digital economy is stimulated by approximately 0.36 percentage points, and the GDP is stimulated by approximately 0.17 percentage points.

The shortage of computing power has even become a key factor constraining AI research and application. In response, the United States has imposed a ban on the sale of high-end computing power products to China, with companies such as Huawei, Loongson, Cambricon, Sugon, and Hygon being placed on the entity list. Their advanced chip manufacturing processes are restricted, and the domestic processes capable of mass production are 2-3 generations behind the international advanced level, with core computing power chips lagging 2-3 generations behind the international advanced level.

01

The computing power shortage has given rise to a huge market for computing power centers.

In the 21st century, mobile computing and cloud computing have flourished. The advent of cloud computing has allowed computing power to "flow" through the network like water and electricity to every corner that needs it.

The rise of artificial intelligence has put forward higher demands for computing power. The emergence of specialized hardware such as GPUs (Graphics Processing Units) and TPUs (Tensor Processing Units) has greatly improved processing efficiency and provided strong support for the training and inference of machine learning models.The scarcity of computing power has further promoted and nurtured the emergence of a vast market for computing power centers. A computing power center refers to a computing center equipped with high-performance computing, large-scale storage, high-speed networking, and other infrastructure, aimed at providing large-scale, efficient, and cost-effective computing power services.

Taking the domestic market as an example, many regions across the country are accelerating the deployment of public computing power infrastructure. In Shanghai, the first national computing power trading platform and artificial intelligence public computing power service platform were established. In Guangzhou, the first domestic computing power resource publishing and sharing platform was built. It can be said that these public platforms have played a role in connecting supply and demand.

At present, eight regions in the country are constructing national computing power hub nodes, and ten national data center clusters have been planned to build a national computing power network system. By the end of 2023, there were 128 intelligent computing center projects in the country, with 83 projects disclosing their scale, totaling more than 7.7 million P. Additionally, in 2024, a total of 39 intelligent computing center projects have been put into production.

02

The gap in intelligent computing still exists, and the three major operators are deploying intelligent computing centers

In the past two years, AI large models have emerged one after another, and the demand for intelligent computing has been growing rapidly. Market research firm IDC estimates that by 2026, China's intelligent computing power scale will enter the petaflops (ZFLOPS) level, reaching 1271.4EFLOPS. The "Action Plan for High-Quality Development of Computing Power Infrastructure" released by six departments has clarified the construction pace of top-level computing power in the next three years. It mentioned that the intelligent computing construction gap from 2023 to 2024 is 23EFlops. The national computing power target for 2025 is to exceed 300EFlops, with the proportion of intelligent computing reaching 35%, and the target for intelligent computing power is 105EFlops.

In response to this, the three major operators have also been actively deploying intelligent computing centers in recent years and have proposed relevant strategic deployments.

China Unicom has laid out a "1+N+X" intelligent computing capability system in providing specialized intelligent computing infrastructure services, including 1 ultra-large-scale single intelligent computing center, N intelligent computing training and promotion hubs, and localized X intelligent computing inference nodes.

China Mobile has strengthened the "4+N+31+X" data center layout, achieving coverage of computing power resources around hot spots, centers, and edges, with more than 1000 edge nodes built. The "4+N+31+X" data center system, where "4" refers to the four hot business areas of Beijing-Tianjin-Hebei, Yangtze River Delta, Guangdong-Hong Kong-Macao Greater Bay Area, and Chengdu-Chongqing; "N" refers to the ten ultra-large data centers planned within the ten national hub node data center clusters; "31" refers to the ultra-large data centers planned by each province; "X" refers to the data centers and convergence rooms at the city level and below.China Telecom has proposed the concept of "cloud-network integration," forming a computing power layout of "2+4+31+X+O." This specifically refers to creating integrated resource pools in the two national cloud bases of Inner Mongolia and Guizhou, building large-scale public clouds in four major regions such as the Beijing-Tianjin-Hebei area; constructing localized exclusive clouds in 31 provincial capital cities and key cities; creating differentiated edge clouds at X nodes; and deploying along the "Belt and Road" countries, extending the computing power system overseas.

03

The United States, Europe, and Japan are making significant investments, sparking a global AI computing power "arms race"

Currently, countries around the world are formulating their own artificial intelligence strategies and policies to promote the development of the AI industry.

The United States, in its "National Artificial Intelligence Research and Development Strategic Plan" released in 2016, clearly proposed to strengthen the construction of AI infrastructure. At the same time, the European Union also clearly proposed the goal of strengthening infrastructure construction in its AI strategy released in 2018. These infrastructures mainly include computing resources, data resources, and talent resources. Japan followed the United States' pace and successively issued three versions of the "Artificial Intelligence Strategy" in 2019, 2021, and 2022. In April last year, the Japanese government established an Artificial Intelligence Strategy Group, led by the Prime Minister's Assistant Officer, Muraishi Hideki, with members including officials in charge of AI policy from the Cabinet Secretariat, Ministry of Foreign Affairs, Digital Agency, and other departments.

Under a series of strategic deployments, countries and regions such as the United States, Japan, and Europe are also competing to build computing power centers, sparking a global AI computing power "arms race."

In November last year, the National Supercomputing Center of the United States and many leading companies in the AI field jointly established the Trillion Parameter Consortium (TPC). The consortium is composed of scientists from around the world and aims to jointly promote AI models for scientific discovery, with a particular focus on giant models with one trillion or more parameters. Currently, TPC is developing scalable model architectures and training strategies, and organizing and sorting scientific data for model training to optimize the AI library of current and future exascale computing platforms.

In addition, the Oak Ridge National Laboratory, Lawrence Livermore National Laboratory under the U.S. Department of Energy, and IBM, NVIDIA have established the Supercomputing Excellence Experiment Center to jointly develop the next generation of HPC computers, using IBM's Power processors and NVIDIA's Teslak accelerator cards, with a floating-point performance of at least 10 quadrillion, and up to 30 quadrillion.

In December 2020, the European Union planned to allocate 7.5 billion euros for the "Digital Europe" plan, of which 2.2 billion euros were for supercomputing and 2.1 billion euros for artificial intelligence. The plan specifically includes: acquiring at least one exascale supercomputer by the end of 2021; establishing a Europe-wide data space and testing facilities for artificial intelligence in areas such as health, manufacturing, and energy; deploying a Europe-wide quantum communication infrastructure and supporting the establishment of a cybersecurity product certification plan; and setting up master's programs in artificial intelligence, advanced computing, and cybersecurity, etc.In March of last year, the UK government pledged to invest £1 billion ($1.3 billion) in supercomputing and artificial intelligence research, with the hope of becoming a "technology superpower." As part of this strategy, the government expressed its intention to spend approximately £900 million on constructing an "ultra-large-scale" computer capable of building its own "BritGPT," rivaling OpenAI's generative AI chatbot.

In April of this year, Japan's Ministry of Economy, Trade, and Industry (METI) allocated a total subsidy of 72.5 billion yen to five Japanese companies for the development of AI supercomputers, aiming to reduce technological dependence on the United States. The Japanese government provided government subsidies of 50.1 billion, 10.2 billion, 1.9 billion, 2.5 billion, and 7.7 billion yen to Sakura Internet, telecommunications giant KDDI, GMO Internet, Rutilea, and Highreso, respectively. Reports indicate that Japan's "National Institute of Technology and Evaluation" will develop a supercomputer as early as this year, with computing power approximately 2.5 times that of existing machines. Under the supervision of METI, this institution will provide the supercomputer to domestic Japanese companies developing generative AI through cloud services.

In addition to government-backed projects, global tech companies are also investing heavily in building computational power. Amazon plans to invest $148 billion over the next 15 years to build data centers around the world to meet the demands of artificial intelligence and other needs. Google announced a $3 billion investment to construct or expand data center campuses in Virginia and Indiana. Microsoft and OpenAI are also undertaking a five-phase supercomputer construction project, with proposed investments exceeding $115 billion, most of which will be spent on procuring the computational facilities required to drive AI.

Operators initiate large-scale procurement, AI chip market takes off.

The large-scale layout of computational centers has also led to large-scale procurement of AI chips.

Recently, China Mobile's large-scale AI chip procurement has attracted widespread attention in the industry. China Mobile initiated the procurement for new intelligent computing centers from 2024 to 2025. The tender announcement shows that the total scale of this project's procurement reaches 8,054 units. Some institutions estimate that, based on previous winning bids, the scale of this procurement may exceed 15 billion yuan.

A month ago, China Unicom also initiated the procurement of more than 2,500 AI servers, and China Telecom had already taken action before. With the three major operators initiating large-scale bidding, it is seen within the industry that domestic computational deployment has entered the "fast lane."

Just two months ago, China Mobile also released a procurement project for new intelligent computing centers (trial network) from 2023 to 2024, with 12 bid packages corresponding to a total AI training server procurement volume of 2,454 units (1-11 bid packages with 1,204 units, and 12th bid package with 1,250 units).At the end of March, China Unicom announced the prequalification notice for the 2024 China Unicom Artificial Intelligence Server Centralized Procurement Project. The notice indicates that the 2024 China Unicom AI server centralized procurement project has been approved, with the bidders being China United Network Communications Group Co., Ltd., its provincial branches, and China Unicom Digital Technology Co., Ltd., among others. This time, China Unicom will purchase a total of 2,503 AI servers and 688 key networking devices, RoCE switches, without dividing the procurement into separate packages.

In October last year, China Telecom also announced the review results of the AI computing power server (2023-2024) centralized procurement project, with manufacturers such as Super Fusion, Inspur, and H3C being shortlisted, totaling the procurement of 4,175 AI servers and 1,182 switches.

05

Construction of computing power centers benefits AI chip manufacturers

At present, the main enterprises involved in the construction of computing power centers include telecommunications operators, large cloud service companies, and large internet companies. These companies have ample funds, large scale, and can bear the huge costs of building computing power centers. At the same time, they have a huge demand for computing power and have a wealth of downstream customers to whom they can sell computing power.

On October 17, 2023, the U.S. Department of Commerce issued the export control list requirements ECNN 3A090 and 4A090 to further restrict the export of high-performance AI chips, and also added 13 Chinese companies to the entity list. The modified export control design products include but are not limited to: NVIDIA A100, A800, H100, H800, L40, L40S, and RTX 4090 products. Due to the U.S. restrictions on the procurement of domestic computing power AI chips, the computing power centers and related AI chips have formed two markets, domestic and international.

The huge domestic computing power market has driven the benefits for domestic chip manufacturers. Recently, China Mobile officially announced the completion of the world's largest single operator intelligent computing center—China Mobile Intelligent Computing Center (Hohhot), which is now in production and use. The intelligent computing center project is equipped with approximately 20,000 AI acceleration cards, with a domestic AI chip rate of over 85%.

China Unicom has also recently built the first "government + operator" intelligent computing center in Beijing, and the computing power center continues to use the domestically produced Ascend AI basic software and hardware.

Previously, the "large-scale computing power cluster and artificial intelligence public computing power service platform" lit up by China Telecom Shanghai Company in Shanghai is the largest operator-level intelligent computing center in the country, with a computing power cluster scale of 15,000 cards, using independently innovative AI chips. The China Telecom Central Intelligent Computing Center, which started operation at the beginning of the year, also adopted a solution architecture based on the domestic AI basic software and hardware platform.It is not difficult to observe that domestic computing power centers often utilize domestically produced AI software and hardware. Currently, GPUs have the largest market share in the AI chip market, and the procurement of domestic AI chips mainly benefits Chinese representative enterprises, including Huawei, Sugon Information, Jingjia Micro, and Suiyuan Technology. Last year, Baidu ordered 1600 Ascend 910B AI chips for 200 servers.

According to institutional estimates, affected by the upgrade of restrictions on NVIDIA, the new market space for domestic AI chips will reach over 70 billion by 2024.

Other major foreign markets face fewer restrictions in chip procurement. The global AI chip market is currently dominated by large European and American manufacturers represented by NVIDIA. Industry data shows that NVIDIA, with an 80% market share, almost "monopolizes" the AI chip market. Previously, NVIDIA's CEO, Jen-Hsun Huang, also announced that they would establish an AI factory in Japan, which would prioritize the supply of GPUs for the domestic Japanese market.

Competition intensifies, and major manufacturers begin to develop their own AI server chips.

Currently, it is widely believed that under the AI boom, the most benefited are the AI chip manufacturers who are like "selling shovels." Data shows that the cost of chips accounts for about 32% of the total cost in basic servers, while in high-performance or even higher-performance servers, the proportion of chip costs can reach 50% to 83%.

The excessively high costs have also led to an increasing number of internet and IT equipment giants starting to develop their own AI server chips.

In 2016, Google launched its self-developed AI Tensor Processing Unit (TPU). Around 2022, Google began developing server CPUs based on the Arm architecture. In April 2024, Google released its self-developed Arm architecture CPU—Axion—and announced that the chip has been in internal use.

In 2020, Microsoft began customizing chips for its Azure cloud services. In November 2023, Microsoft launched two self-developed chips—Maia100 and Cobalt100. Maia100 is a chip specifically designed for the training and inference of large language models, using TSMC's 5nm process, while Cobalt100 is a 128-core server CPU based on the Arm architecture.In early April this year, Meta unveiled its new generation of AI training and inference accelerator, MTIA, which boasts more than double the computing power and memory bandwidth of its predecessor. The latest version of the chip is instrumental in driving ranking and recommendation advertising models on Facebook and Instagram.

Previously, it was reported that American AI research company OpenAI is in talks with potential investors, including the government of the United Arab Emirates, in an attempt to promote a project aimed at enhancing global chip manufacturing capabilities and reshaping the global semiconductor industry. One informed source revealed that the plan is preparing to raise up to 5 to 7 trillion US dollars.

In addition, domestic giants are not lagging behind and have begun developing AI chips. Recently, China Mobile officially launched the Dayun Panshi DPU at its 2024 Computing Power Network Conference. The chip's bandwidth reaches 400Gbps, representing a leading level in the domestic market.