How AI and “East‑West Computing” Are Reviving the Server Market

Since OpenAI released ChatGPT in 2022, large‑language models and AIGC have pushed global tech giants to adopt AI‑driven compute, turning servers into the primary engine for massive processing power and driving rapid growth in AI‑focused server demand.

The AI server ecosystem comprises CPUs (x86, ARM, MIPS, RISC‑V), GPUs, memory (DRAM and high‑bandwidth memory HBM), local SSD storage, NICs, PCIe slots, and cooling solutions. AI chips—GPU, FPGA, ASIC, and NPU—are the core compute units for AI workloads. HBM, now standard on high‑end GPUs, integrates memory closely with the processor, dramatically increasing bandwidth and alleviating the memory‑wall problem.

Enterprise‑grade SSDs consist of NAND flash, a controller chip, and DRAM, running sophisticated firmware that turns the drive into a small system capable of processing, caching, computing, and security protection. As a result, SSD penetration in data‑center servers is expected to rise steadily.

Cost composition differs sharply between conventional and AI‑optimized servers. A typical 2‑socket Intel Sapphire Rapids server priced at $10,424 allocates roughly 17.7% of its cost to CPUs and over 50% to memory and storage. In contrast, Nvidia’s DGX‑H100 AI server costs $268,495, with CPUs accounting for only 1.9% while GPUs dominate at 72.6%; memory value increases but its cost share drops to about 4.2%.

These figures illustrate that chips and storage are the decisive factors for AI server performance and bandwidth, and the shift in cost structure reflects the growing emphasis on AI workloads in modern data centers.