TSMC's AI design aid fails to ease 3nm and 2nm decisions

In an exclusive interview with CNA on Tuesday, Taiwan Semiconductor Manufacturing Company (TSMC) Executive Vice President and Co-Chief Operating Officer Y.J. Mii pointed out that the application of artificial intelligence (AI) in the semiconductor industry remains supplementary. Particularly in the highly specialized domain of advanced manufacturing processes, AI is still unable to replace human judgment and decision-making. Mii stated that AI can provide substantial assistance in information gathering, data organization, and pattern recognition, allowing engineers to reduce the time cost of manual searching during preliminary planning and parameter analysis. However, he emphasized that for the "foreseeable future," AI will not possess the capability to make critical manufacturing decisions for humans—a point that is especially crucial for TSMC's upcoming mass production of cutting-edge nodes such as the 3-nanometer and 2-nanometer technologies.

In the semiconductor design flow, electronic design automation (EDA) tools serve as the critical bridge transforming circuit diagrams into manufacturing layouts. Mii mentioned that AI has already demonstrated significant potential in certain subfields of EDA, such as optimization algorithms in the automated placement and routing stages, which can rapidly find near-optimal solutions among vast design variables and shorten verification cycles. Such technologies have already been adopted by several major chip design companies to enhance design efficiency and reduce error rates. Even so, AI is still incapable of directly completing extremely advanced process technologies such as the "A14" node, because such processes involve highly complex physical problems including photolithography, materials science, and quantum effects that extend far beyond the scope of pure data patterns.

Mii's perspective contrasts with prevailing industry expectations. In recent years, global wafer foundries and design companies have invested in AI-related research and development, hoping to accelerate process development and reduce trial costs through machine learning. TSMC also established an AI research and development center in 2022, cooperating with multiple AI suppliers to explore applications in process simulation, defect detection, and yield prediction. Nevertheless, from the perspective of process development, every breakthrough at advanced nodes must undergo rigorous experimental verification and equipment tuning, and any decision-making error could lead to massive losses. Mii therefore reminded that AI currently can only provide "information" and "suggestions," while the final setting of process parameters and risk assessment must still be completed by engineers with deep professional experience.

Furthermore, the limitations of AI in manufacturing decision-making also reflect the supply chain structure of the semiconductor industry. Most equipment for advanced processes is provided by a handful of U.S. and Japanese firms, while extreme ultraviolet (EUV) lithography systems remain exclusively monopolized by ASML of the Netherlands. The operation and maintenance of such equipment itself require highly specialized technicians; for AI to play a role at this level, it must first break through in its deep understanding and real-time control of hardware behavior. Mii pointed out that in the short term, AI is more suited to act as an "assistant" role, helping engineers rapidly screen massive amounts of test data to enhance the informational quality of decisions rather than directly replacing human judgment.

Finally, Mii's remarks serve as a reminder to investors and industry observers that the AI craze should not be overly mythologized. Even today, as AI technology rapidly evolves, the core of semiconductor manufacturing remains human innovation and engineering practice. Moving forward, TSMC will continue to combine AI-assisted tools with a steady research and development pace to gradually enhance process performance and yields. However, at extreme nodes such as "A14," AI remains a provider of information rather than a decision-maker. This stance holds significant reference value for assessing TSMC's risks and opportunities in the competition for advanced manufacturing processes.