[Real-World Asset Management] CPUs in the Spotlight in the Age of Agentic AI... Promising Investment Picks

The flow of investment in the artificial intelligence (AI) semiconductor market is shifting once again. As AI evolves beyond the training stage to inference and agentic AI, demand that was once concentrated on GPUs (graphics processing units) is now spreading to CPUs (central processing units).

원본보기 아이콘

Recently, the upward trend in CPU company stock prices has become pronounced, following Intel's first-quarter earnings surprise. Notably, Intel's data center and AI (DCAI) revenue rose by 22% year-on-year.

The stock price also surged. Although Intel was removed from the Dow Jones Industrial Average in 2024 after 25 years, the stock has increased by approximately 170% so far this year due to expectations for CPU demand. In April alone, the share price rose by as much as 119%.

Kim Jaeseung, a researcher at Hyundai Motor Securities, explained, "With DCAI division revenue growing by 22% year-on-year, the expansion of the CPU’s role and growing demand have become clear. As AI advances from the training phase to inference and agentic AI, the role of the CPU, which coordinates tasks, is being highlighted once again." He pointed out that initial AI investments focused on training, where GPUs, HBM (high bandwidth memory), and high-speed networks were key, but now that AI is moving into the service stage, the share of non-computational tasks such as model selection, data lookup, and tool invocation is expanding.

In agentic AI, the process of "decision-making → tool selection → execution → verification" is repeated, making the CPU’s role in orchestrating tasks increasingly important. Kang Jaegoo, an analyst at Hana Securities, said, "In the inference market, the CPU serves as the 'traffic controller' of AI systems, supplying data so that GPUs do not remain idle and thereby increasing efficiency." Park Jaehwan, a researcher at Eugene Investment & Securities, also explained, "In the era of AI agents, the CPU will play a central role in coordinating the entire workflow."

The structure of data centers is also changing. Whereas the CPU-to-GPU ratio was previously between 1:4 and 1:8, in agentic AI environments, it could drop to as low as 1:1 or 1:2.

Alongside rising demand, supply issues are also coming to the fore. CPUs are more difficult to ramp up in the short term than GPUs due to manufacturing constraints, which means bottlenecks are likely to recur during supply expansion. Server CPUs rely on advanced packaging technologies such as TSMC’s cutting-edge processes and CoWoS·SoIC, which bundle multiple chips together, and these production capacities are already saturated due to demand for GPUs and semiconductors for smartphones.

Han Yonghee, a researcher at Growth Research, noted, "CPU lead times have increased from the previous 1-2 weeks to 8-12 weeks, and in some cases, there are wait times of more than six months. Both Intel and AMD are raising the prices of server CPUs by about 10-15%." Ko Junhyuk, an analyst at Shinhan Investment Corp., said, "Due to the paradigm shift in AI computation, CPUs have entered a phase of structural supply shortages. As the CPU-to-GPU ratio narrows to 1:1, price hikes are accelerating in earnest."

원본보기 아이콘

Investor interest is expanding beyond CPUs to the entire back-end value chain, including substrates, packaging, and testing. Kang noted, "Although the importance of CPUs is increasing, it is difficult to say that all the benefits will go straight to Intel. AMD is likely to have an advantage in the traditional data center AI transition, and Arm-based CPUs may spread more rapidly in large-scale AI inference data centers."

He added, "Nvidia CPUs employ Arm-based designs, and hyperscalers are also increasing their use of proprietary Arm-based CPUs. As AI inference grows within the cloud, some of the demand for x86 CPUs may be absorbed by Arm-based CPUs."

Han analyzed, "CPU supply shortages are not an issue for CPU vendors alone; the benefits are likely to extend to memory, flip chip ball grid array (FC-BGA) substrates, packaging, and testing." Server CPUs require high-specification FC-BGA substrates and advanced packaging technologies, and securing sufficient memory bandwidth to support CPU performance is also essential.

From the perspective of domestic investment, attention should be paid to value chain companies rather than CPU design firms. Specifically, Samsung Electro-Mechanics and Daeduck Electronics, which produce FC-BGA substrates, as well as back-end and testing-related companies, are cited as major beneficiaries. Expectations are also being reflected in substrate companies. Ko Euyeong, a researcher at iM Securities, commented on Simmtech, saying, "With the rising prominence of agentic AI leading to increased CPU demand and the launch of Nvidia's Vera CPU rack, there is potential for an upside in SOCAMM-related revenue."

© The Asia Business Daily(www.asiae.co.kr). All rights reserved.