AI data bottleneck drives race to reinvent chips

With the global semiconductor market on track to surpass $2 trillion by 2030, surging demand for AI infrastructure is transforming the chip race from a battle over individual chip performance into a competition over who can best integrate memory, logic, optical communications and advanced packaging.

Samsung Electronics presented its “CUBE” strategy at Semicon Taiwan, Asia's largest semiconductor exhibition, on Sept. 1 to overcome the technological limitations of memory chips.

CUBE stands for Capacity, Utilization, Bandwidth and Efficiency, referring to power and thermal efficiency.

The name, which evokes a three-dimensional cube, reflects a strategy of building memory upward rather than spreading it outward — moving from a two-dimensional structure to a three-dimensional one.

The idea is to maximize capacity by stacking memory vertically without expanding the area it occupies on a printed circuit board. Optimizing memory layout can also increase data processing speeds. The approach is designed to improve performance per watt by reducing the energy required to move data.

The chip industry is also working to fundamentally change how chips connect with one another. SEMI, the global semiconductor industry association that organizes Semicon, said this year's event will focus on silicon photonics and co-packaged optics (CPO). Silicon photonics creates pathways for light directly within chips to transmit data instead of relying on electrical wiring, while CPO integrates optical communication components and semiconductors into a single package.

At a silicon photonics forum held on Monday ahead of Semicon Taiwan, industry participants also highlighted the need to shift from conventional copper wiring to light-based transmission to overcome bottlenecks in data movement. Taiwan's TSMC, the world's largest contract chipmaker, has also been promoting its silicon photonics-based optical packaging platform, known as its Compact Universal Photonic Engine (Coupe).

Chip giants are racing to develop technologies ranging from vertically stacked memory to light-based data transmission because of one major bottleneck created by AI: moving data.

As AI becomes more sophisticated and capable of processing text, images, video and audio simultaneously, the amount of data moving through computing systems has surged.

April Li, TSMC's director of AI and high-performance computing business development, said data movement alone can account for up to 60 percent of system activity in typical workloads, leaving expensive accelerators operating at below 40 percent utilization, according to Taiwan’s Central News Agency.

"In this new era, market leaders will not be the ones who simply make better models, but those who build the most integrated systems," Li said at the Semicon Taiwan 2026 IC forum.

Investment in equipment and technologies designed to break through these data bottlenecks is also driving growth in the semiconductor market.

SEMI forecasts that the global semiconductor market will surpass $2 trillion by 2030, fueled by the boom in AI data center construction.

Against this backdrop, this year's Semicon Taiwan is also set to be the largest in the event's history. More than 1,300 companies will participate, operating 4,300 booths, while attendance is projected to exceed 100,000.

BY KIM SU-MIN [kim.jiye@joongang.co.kr]