Will Samsung Electronics and SK Hynix See Lower Semiconductor Demand Due to AI Efficiency? [Weekend Money]
Transitioning Toward More Complex AI
Agent AI on the Rise for Large-Scale Data Processing
Rising Demand for CPUs, Memory, and Networking Devices
Recently, artificial intelligence (AI) models such as ChatGPT and Grok have been highlighting their efficiency by achieving the same performance with lower costs. From an investor's perspective, there may be doubts that if AI can run at lower cost, demand for infrastructure could decrease as well. However, NH Investment & Securities diagnosed this as an unnecessary concern. The company analyzed that this is merely a step toward more complex AI applications and that the ecosystem will continue to expand as the importance of infrastructure such as graphics processing units (GPUs) and networks increases.
In a recent report, NH Investment & Securities emphasized that token usage and AI performance still remain proportional. Higher efficiency does not reduce overall token demand; rather, it means that more tasks can be attempted for the same cost.
In particular, they pointed to "agent AI" as the core development. Agent AI refers to systems capable of autonomously dividing user instructions into multiple steps and fetching necessary data from external sources. It can also execute multiple sub-agents in parallel. This process requires much greater computational power and network connectivity compared to simple Q&A formats.
Both GPUs and central processing units (CPUs) are vital in this process. Some operations are processed in the CPU environment outside of the AI model. Instead of having the AI read entire external datasets, the CPU pre-processes the data and only passes on interpreted results to the AI. As a result, demand is projected to increase not only for GPUs but also for CPUs and legacy memory.
NH Investment & Securities also predicted that the importance of network infrastructure will rise. Agent AI operations rely on multiple agents acting in parallel and utilize "prompt caching," where previous prompts are saved and reused. In such scenarios, not only GPU internal memory but also local storage and inter-server networks are required.
Guijin Kim, a researcher at NH Investment & Securities, believes that this trend could lead to growing demand for CXL controllers, NVMe SSDs, InfiniBand, and Ethernet switches. This indicates that the beneficiaries of AI investment will not be limited to GPUs and high-bandwidth memory (HBM), but could expand across CPUs, memory, storage, and network segments.
Hot Picks Today
"No Need for Private Academies: Saving 800,000 Won a Month and Boosting Grades... Parents Flock After Hearing the News"
- "I Didn't Expect This Much"... As Chinese Numbers Plunge, South Koreans Become Top Visitors to Japan
- U.S. Retail Sales Rise 0.2% Month-on-Month... Jobless Claims Down 8,000 From Previous Week
- "One Year from Now Will Be Worse Hell"... 9 Out of 10 Japanese Fear 'Runaway Inflation'
- "How Can So Many Fit on One Boat?"...Crowds in Red: What Is This?
Researcher Kim stated, "AI companies announce reductions in token prices only after thoroughly optimizing their software, and flagship model prices have not been lowered at all." He interpreted cost reduction as a foundation for expanding usage, adding that it is not a form of cutthroat competition.
© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.
