What’s Driving the Explosive Growth of AI Servers and Data Centers in 2024?

Data centers can be grouped by computing power into three categories: cloud data centers that support a wide range of applications, intelligent computing centers built on AI‑specific chips to accelerate AI industrialization, and supercomputing centers primarily for scientific and engineering calculations, largely driven by national science initiatives.

China’s data‑center development started later than the global market and is currently transitioning from a cloud‑focused phase to an intelligent‑computing phase, indicating a growth stage.

According to Trendforce, global AI server shipments exceeded 1.208 million units in 2023, a year‑over‑year increase of 37.7 %. The forecast predicts 1.672 million units in 2024, up 38.4 %. TSMC expects AI demand to grow at a compound annual rate of 50 % through 2028, sustaining high AI server demand.

In 2023, China’s AI server market reached US$9.1 billion, up 82.5 % YoY, with intelligent‑computing capacity projected at 414.1 EFLOPS (59.3 % YoY). From 2022 to 2027, the market is expected to grow at a 33.9 % CAGR.

AI servers are divided into training and inference models. Training servers require high storage, bandwidth, and compute, typically using 8‑GPU designs, while inference servers have lower requirements and may use GPUs, NPUs, CPUs, or PCIe‑based AI accelerators depending on the workload.

Server evolution follows a path from general‑purpose to cloud, edge, and finally AI servers, with GPU‑enhanced parallel processing as a core feature. The combination of CPU and GPU remains the fundamental architecture, and rack‑level solutions are emerging as a dominant form factor for future AI server shipments.

The three main challenges for AI server development are: (1) ensuring the quality and reliability of multi‑layer high‑density HDI PCBs required for advanced GPU accelerators; (2) meeting the high‑speed interconnect demands (e.g., 56G‑112G‑224G) with suitable high‑density backplane connectors; and (3) addressing the thermal design power of single high‑performance AI chips, which can exceed 1000 W, pushing the limits of traditional air cooling and driving the adoption of liquid‑cooling solutions.