Foxconn's AI Servers for Apple in Testing Phase - AI Model Acceleration Drives AI Server Industry Development Analysis Report
-
Foxconn's AI servers for Apple are currently in the testing phase. As Apple focuses on AI applications, Foxconn has opportunities to expand shipments. Regarding whether Foxconn has received new AI server orders from Apple, the company declined to comment on individual clients or products. Supply chain sources revealed that Apple has previously conducted numerous AI feature tests, which are highly confidential, and that Apple and Foxconn have many ongoing projects under testing.
Following the popularity of ChatGPT, major tech companies have been actively developing large AI models, leading to a shortage of general-purpose computing chips. The wave of the AI era has brought massive data that drives enormous computing power demands, resulting in increasing demand for AI servers. Interface chips for data transmission both inside and outside servers have also seen rising demand.
The "2023-2024 China Artificial Intelligence Computing Power Development Assessment Report", jointly released by IDC and Inspur Information, predicts that China's AI server market will reach $9.1 billion in 2023, an 82.5% year-on-year increase, and is expected to grow to $13.4 billion by 2027, with a compound annual growth rate (CAGR) of 21.8%.
The report suggests that AI is rapidly advancing from perceptual intelligence to generative intelligence, driving rapid growth in AI computing power demand. It forecasts that China's intelligent computing power will grow at a 33.9% CAGR from 2022 to 2027, while general computing power will grow at a 16.6% CAGR.
The global AI hardware (server) market is projected to expand from $19.5 billion in 2022 to $34.7 billion in 2026, with a 17.3% CAGR. Among this, the market share of servers dedicated to generative AI will increase from 11.9% in 2023 to 31.7% in 2026.
AI Server Industry Analysis With the accelerated iteration of AI large models, intelligent computing power has become a scarce resource. For a considerable period in the future, China's AI server market will remain in a state of supply shortage, presenting a critical growth window for the domestic AI chip market.
Currently, large model manufacturers are vigorously iterating their products. According to Wang Xiaochuan, founder and CEO of Baichuan Intelligence, the Baichuan large model currently undergoes monthly updates to its model data.
A representative from Angqin Computer stated that the explosive growth in data required for large model training has created demand for high-end AI servers capable of supporting storage and training. As a result, server manufacturers are upgrading chip specifications, expanding card group quantities, and moving toward high-end AI server solutions.
Driven by the wave of AI large models, the market scale for AI chips, servers, and data centers is expected to grow significantly. IDC predicts that China's AI chip shipments will reach 1.335 million units in 2023, a year-on-year increase of 22.5%.
Regarding AI servers, IDC forecasts that China's AI server market will reach $9.1 billion in 2023, an 82.5% year-on-year growth, and expand to $13.4 billion by 2027, with a compound annual growth rate of 21.8%.
With the accelerated iteration of AI large models, intelligent computing power has become a scarce resource. Over the past five years (2018-2022), cities including Beijing, Hangzhou, Shanghai, Shenzhen, Guangzhou, Hefei, Suzhou, and Chongqing have demonstrated outstanding performance in the artificial intelligence sector. Influenced by the emergence of generative AI and investments in intelligent computing centers, China's 2023 AI city rankings have shifted, with Beijing maintaining its top position, followed by Hangzhou and Shenzhen in second and third places respectively. Other cities in the top ten include Shanghai, Suzhou, Guangzhou, Jinan, Hefei, Chongqing, and Chengdu.
The construction of intelligent computing centers has accelerated. According to incomplete statistics, as of August 2023, over 30 cities nationwide have established intelligent computing centers.
In the era of large AI models, the supply of high-end chips like imported brand GPUs faces cyclical fluctuations, creating a crucial window for domestic AI chips to enter the market. However, the training and inference of large AI models, along with massive data storage, all require high-performance computing power, which poses challenges for the underlying infrastructure capabilities and ecosystem development of domestic computing resources.
Media research has identified that chip companies such as Inspur Information, Hygon Information, Sima Computing, Zhongke Tongliang, Vastai Technologies, Moffett AI, Moore Threads, Iluvatar CoreX, Cambricon, and Enflame Technology have launched chip accelerator cards for AI inference and training tasks across various scenarios, involving different design architectures including CPU, GPU, and RISC-V.
An AI server is a data server capable of providing artificial intelligence (AI) services. It can support both local applications and web services, as well as deliver complex AI models and services for cloud and local servers. AI servers facilitate real-time computing services for various AI applications. AI servers primarily adopt two architectures: one is a hybrid architecture that allows data to be stored locally, and the other is a cloud-based architecture that utilizes remote storage technology and hybrid cloud storage (a technology combining local and cloud storage) for data storage.
AI server costs mainly come from chips such as CPUs and GPUs, accounting for over 50%
Servers consist of several components, including power supplies, CPUs, memory, hard drives, fans, and optical drives. Among these, chip costs (CPUs, GPUs, etc.) account for a relatively high proportion, ranging from 25% to 70%.
AI servers employ heterogeneous architectures, which can be categorized into combinations such as CPU+GPU, CPU+FPGA, and CPU+ASIC based on chip types. Currently, GPUs remain the preferred choice for data center acceleration, while the application of other non-GPU chips is gradually increasing. IDC predicts that by 2025, the proportion of other non-GPU chips will exceed 20%. Generally, ASICs offer the best performance but have weaker programmability and flexibility. For training or general purposes, GPUs are the better choice.
The upstream of China's AI server industry chain consists of core components and software suppliers; the midstream includes AI server manufacturers; and the downstream comprises various application markets, such as internet companies, cloud computing enterprises, data center service providers, government departments, financial institutions, the healthcare sector, and telecom operators.