TSMC: AI's Next Phase Demands Integrated Systems, Not Just Better Chips
TSMC says AI's next phase is industrialization, requiring integrated systems over isolated chips, as inference demand soars.
The next wave of artificial intelligence will be defined not by individual breakthroughs in chips or models, but by the seamless integration of entire computing ecosystems, according to a senior executive at Taiwan Semiconductor Manufacturing Co. (TSMC).
Speaking at the SEMICON Taiwan 2026 IC forum, April Li, TSMC's director of AI and high-performance computing business development, said the industry is entering the "true industrialization of AI." She argued that market leadership will shift to companies that build the most integrated systems rather than those that simply produce superior models.
Li highlighted that AI agents are increasingly embedded in enterprise workflows, while investment in specialized data centers and physical AI accelerates. This shift is driven by a surge in AI inference—the process of using trained models to respond to queries. Global inference token volume has grown nearly 500-fold since 2022, with complex reasoning and agentic AI consuming far more tokens than standard requests.
"Inference is no longer a low-overhead task," Li said, noting it is becoming the primary driver of system-level expansion as AI agents run continuously in the background.
The growing workloads are straining four critical areas: logic scaling, interconnect efficiency, memory performance, and power delivery and cooling. Li noted that data movement can account for up to 60 percent of system activity in typical workloads, leaving accelerators underutilized at below 40 percent.
Looking ahead, AI packages may contain over one trillion transistors by 2030, making multi-die architecture and heterogeneous integration essential. To address these constraints, TSMC is developing advanced logic, packaging, and optical interconnect technologies. Li specifically mentioned the 3DFabric platform, which includes SoIC 3D stacking and CoWoS packaging, as well as the Compact Universal Photonic Engine (COUPE) for high-speed optical data transmission.
TSMC is also enhancing high-bandwidth memory performance by applying advanced logic to memory base dies. Li emphasized the need for the company to understand the supply chain "from silicon to the data center to the token" and to validate technologies with ecosystem partners early.
"Gone are the days when we can just ship wafers across the fence," Li said, underscoring TSMC's expanded role beyond manufacturing individual chips.