AI Server Shipments Expected to Reach Nearly 1.2 Million Units in 2023
-
AI servers are data servers capable of providing artificial intelligence (AI) services. They support local applications and web pages, as well as deliver complex AI models and services for cloud and on-premises servers. AI servers facilitate real-time computational services for various AI applications. There are two primary architectures for AI servers: a hybrid architecture that stores data locally, and a cloud-based architecture that utilizes remote storage technologies and hybrid cloud storage (a combination of local and cloud storage) for data retention.
AI Server Shipments Expected to Grow 38.4% Year-over-Year in 2023
The AI industry chain typically consists of upstream data and computing power layers, midstream algorithm layers, and downstream application layers. Recently, market attention has shifted toward the upstream segment, particularly the computing power sector. New investment opportunities are emerging in AI hardware, as AI software applications rely heavily on hardware computing power.
Driven by the ongoing influence of ChatGPT, domestic demand for AI computing power is expected to maintain growth momentum, benefiting server manufacturers. Estimates suggest that ChatGPT's total computing power requires 7 to 8 data centers, each with an investment scale of 3 billion yuan and 500P computing power. In the era of the digital economy, global data volume and computing power are poised for rapid expansion.
Printed circuit boards (PCBs), often referred to as the 'mother of electronic products,' are critical components of servers. With the rapid development of the AI industry, demand for high-value PCBs used in AI servers has surged significantly.
As demand for AI servers and AI chips rises in tandem, 2023 AI server shipments (including those equipped with GPUs, FPGAs, ASICs, and other primary chips) are projected to approach 1.2 million units, marking a 38.4% year-over-year increase and accounting for nearly 9% of total server shipments. By 2026, this share is expected to rise to 15%. The institution has also revised its compound annual growth rate (CAGR) for AI server shipments from 2022 to 2026 to 22%, while AI chip shipments are forecasted to grow by 46% in 2023.
The institution noted that NVIDIA GPUs dominate the AI server chip market with a share of approximately 60-70%, followed by ASIC chips developed independently by cloud service providers, which hold over 20% of the market.
Compared to general-purpose servers, AI servers employ multiple accelerator cards and utilize high-layer HDI-structured PCBs, which are significantly more valuable. The motherboard layers in AI servers also far exceed those in general-purpose servers, with AI server PCBs being 5-6 times more expensive.
At NVIDIA Computex 2023, founder and CEO Jensen Huang announced that the generative AI engine NVIDIA DGX GH200 has entered mass production. Demonstrations revealed notable architectural changes in the new GH200 server compared to the DGX H100. The GH200 reduces one UBB and one CPU motherboard while adding three NVLink module boards, alongside substantial performance improvements in accelerator cards. This suggests that AI advancements will continue to drive value growth in the PCB sector.
2023 Demand Outlook for the AI Server Industry Chain
PCBs serve as the foundation for server chips, facilitating data transmission and component connectivity. Due to chip upgrades in AI servers, PCBs require targeted enhancements to handle more signals, reduce interference, improve heat dissipation, and enhance power management, necessitating further optimization in materials and processes. Compared to general-purpose servers, AI servers use multiple accelerator cards with high-layer HDI PCBs, significantly increasing their value. According to Prismark, the global server PCB market was valued at $7.804 billion in 2021 and is projected to reach $13.294 billion by 2026, with a CAGR of 11.2%. The per-unit value of server PCBs is expected to rise from $576 in 2021 to $705 in 2026. With the deployment of large AI models and applications, demand for AI servers is growing, signaling imminent market expansion.
Increased AI server shipments are expected to directly benefit related chip, memory, and component suppliers.
For server chips, TrendForce predicts a 46% growth in AI server chip shipments in 2023. NVIDIA GPUs dominate the AI server market with a 60-70% share, followed by ASIC chips developed by cloud providers, which account for over 20%.
AI servers will drive simultaneous upgrades in Server DRAM, SSDs, and HBM. The institution is particularly optimistic about the growth potential of HBM (High Bandwidth Memory), noting that demand for high-end GPUs like NVIDIA's A100 and H100, AMD's MI200 and MI300, and Google's custom TPUs is steadily increasing. HBM demand is projected to grow by 58% year-over-year in 2023, with a further 30% increase expected in 2024.
TrendForce stated that with more clients adopting HBM3 this year, SK Hynix, as the sole supplier of the new-generation HBM3 products, is expected to increase its overall HBM market share to 53%. Meanwhile, Samsung and Micron are projected to hold 38% and 9% of the HBM market share by the end of this year or early next year, respectively.
As global tech giants engage in the arms race for large-scale models, coupled with rapid growth in model parameters and the shift from single-modal to multimodal upgrades, the demand for AI computing power is poised for sustained growth. Vertical industry models are diversifying, downstream applications are proliferating, and user bases are expanding.
The emergence and performance upgrades of large models are driving the AI computing power boom. According to OpenAI, the computing power used for the largest AI training runs has grown exponentially since 2012, doubling every 3-4 months. From 2012 to 2018, the computing power for AI training increased over 300,000 times (compared to a mere 7-fold growth under Moore's Law). OpenAI's 2020 data shows that training a 174.6-billion-parameter GPT-3 model requires approximately 3,640 PFlop/s-day (equivalent to 3,640 days of computing at 10^15 operations per second). With continued investments in large models and growing user bases, AI computing demand is expected to remain robust.
In 2022, China's total computing power reached 180 EFLOPS, with storage capacity exceeding 1,000 EB. The one-way network latency between national hub nodes dropped below 20 milliseconds, and the core computing industry scale hit 1.8 trillion yuan. On April 19, Shanghai's Economic and Information Technology Commission released guidelines for unified computing resource scheduling, aiming to connect and schedule over four computing infrastructures by 2023, achieving over 1,000 PFLOPS (FP16) of schedulable intelligent computing power. By 2025, cross-regional intelligent scheduling will be realized, optimizing supply-demand balance and boosting industrial development. Shanghai's data center computing power will exceed 18,000 PFLOPS (FP32), with green computing accounting for over 10% of new data centers. Large new data centers in cluster areas will achieve a PUE below 1.25 and a green certification level of 4A or higher.
Globally, tech giants are accelerating computing power investments. Software companies are developing their own chips, while hardware firms are building computing platforms. Microsoft, Amazon, Google, and Meta are increasing investments in AI chips, as are Chinese internet giants like Alibaba, Tencent, and Baidu. Meanwhile, chip manufacturers like Intel are focusing on computing platforms and cloud services. At this year's GTC, NVIDIA launched DGX Cloud, an AI cloud service providing access to supercomputing clusters for generative AI training, with subscriptions starting at $37,000 per month.
As the foundation of the digital economy, computing power supports diverse applications. China's data center rack capacity exceeded 6.5 million standard racks in 2022, with an annual growth rate of over 25% in the past five years. The demand for intelligent computing is rising with AI advancements. In February 2022, China approved eight national computing hubs and ten data center clusters, advancing the "East Data West Computing" project. Data centers, as the backbone of AI development, are expected to rebound in demand, with the IDC market projected to reach 612.3 billion yuan by 2024, a 15.9% CAGR from 2022 to 2024.