Samsung and TSMC Pivot to System Architecture to Break AI Data Bottleneck
Semiconductor giants are replacing traditional chip layouts with 3D stacking and silicon photonics to stop AI accelerators from idling.
The race for AI supremacy is shifting from raw processing power to the plumbing of the chip itself. At Semicon Taiwan 2026, industry leaders Samsung and TSMC signaled a fundamental pivot toward integrated system architecture to solve the critical data bottlenecks currently throttling artificial intelligence.
To address these constraints, Samsung unveiled its "CUBE" strategy—standing for Capacity, Utilization, Bandwidth, and Efficiency. The approach centers on 3D vertical memory stacking, which allows data to move more efficiently by stacking memory components vertically rather than spreading them across a 2D plane. This architecture is designed to increase overall capacity and speed while simultaneously lowering power consumption. Simultaneously, TSMC is advancing its Compact Universal Photonic Engine (COUPE), a silicon photonics-based optical packaging platform. This technology aims to replace traditional electrical wiring with light-based data transmission, utilizing co-packaged optics (CPO) to move data at speeds copper cannot sustain.
The Multimodal Pressure
This architectural shift is driven by the evolution of AI models. As systems move toward multimodal processing—handling text, images, video, and audio simultaneously—the volume of data traveling between memory and processors has surged. Traditional copper wiring and 2D layouts have become insufficient for the energy and speed requirements of this next-generation infrastructure. According to Korea JoongAng Daily, data movement can account for up to 60 percent of system activity in typical AI workloads, suggesting accelerators may operate at below 40 percent utilization because they spend more time waiting for data than processing it.
A New Competitive Landscape
Solving this bottleneck is essential for the next leap in AI capability. When accelerators sit idle, the efficiency of the entire data center drops, increasing costs and energy waste. By integrating memory and optics directly into the system architecture, manufacturers can ensure that the most powerful chips are utilized to their full potential. This transition changes the competitive landscape of the semiconductor industry; the winners will no longer be those who simply produce the fastest individual chip, but those who can integrate the entire system.
April Li, TSMC's director of AI and high-performance computing business development, emphasized this shift, stating that in the current era, market leaders will not be those who simply make better models, but those who build the most integrated systems.
The Trillion-Dollar Stakes
The financial incentives for this transition are massive. The global semiconductor market is projected to surpass $2 trillion by 2030, a growth trajectory fueled largely by the aggressive construction of AI data centers. As the industry watches the rollout of CUBE and COUPE, the primary focus remains on whether these integrated systems can scale fast enough to meet the exponential demand for multimodal AI. While the theoretical benefits of silicon photonics are clear, the industry is now waiting to see how these platforms perform in mass-market deployment.