CPUs Resurge in AI Agent Era Amid Shortages

Intel and AMD, which dominate the global CPU market, raised CPU prices by 15% this year. According to Nikkei Asia, the average delivery time for CPUs has also increased from 1–2 weeks to 8–12 weeks, reflecting worsening supply shortages. Semiconductor industry sources revealed that AI cloud providers like Amazon are aggressively expanding CPU servers but still face severe shortages. Global investment bank KeyBanc noted, “Intel and AMD have already sold most of their server CPU production capacity for this year,” adding, “This strong demand could persist until 2027.” Intel and AMD’s stock prices have risen 74% and 25%, respectively, compared to the start of the year.

CPUs are core semiconductors specialized in handling complex computations. They have long been considered the most critical chips in computing systems, acting as the “brain” in PCs, data centers, and other devices. However, with the rise of AI, GPUs—better suited for parallel processing of simple computations—became essential for AI training and operations, relegating CPUs to a supporting role. In fact, AI data centers typically had a structure with 4–8 GPUs per CPU. As AI demand surged, the need for “soldier” GPUs outpaced that for “commander” CPUs.

Graphics by Baek Hyeong-seon

The emergence of AI agents has changed this dynamic. While existing large language models (LLMs) were relatively simple, providing answers when asked questions, AI agents must perform more complex tasks, such as planning operations and calling external tools. This has driven explosive demand for CPUs, the “orchestrators” of these tasks. Market research firm TrendForce stated, “While existing AI data centers required 30 million CPU cores (processing units) per gigawatt (GW), the AI agent era will demand 120 million cores,” adding, “The CPU-to-GPU ratio will shift to 1:1 or 1:2.”

◇Shifts in the Memory Landscape

Amid this trend, semiconductor companies are increasingly developing their own CPUs. NVIDIA, the GPU market leader, expanded its Grace, Vera, and other CPUs as standalone products in February. Arm, a British semiconductor design intellectual property (IP) company, launched its first data center-specific CPU, the “Arm AGI CPU,” last month—its first in 35 years.

As demand surges for both GPUs and CPUs, domestic memory companies, once seen as major beneficiaries of the AI era, are poised for even greater opportunities. Samsung Electronics and SK Hynix are achieving record-high performances due to soaring sales of high-bandwidth memory (HBM) for GPU-based AI chips. Analysts suggest that demand for high-performance general-purpose DRAM (DDR5) used in CPUs could also rise long-term. While DDR5 is already reported to have surpassed HBM in profitability due to severe supply shortages, additional demand could further boost sales. A source from the semiconductor industry remarked, “The CPU shortage and price hikes for data center construction demonstrate that companies’ server investments remain aggressive,” adding, “As the adoption of the latest CPUs increases, demand for high-performance memory will also surge, positively impacting the sales growth of domestic memory firms.”