The global race for artificial intelligence supremacy is currently hitting a physical wall: power consumption. As large language models grow in complexity, the energy required to move data between a computer’s processor and its memory has become a primary bottleneck, creating what engineers call the “memory wall.” This inefficiency not only drives up electricity costs for massive data centers but likewise limits the potential of on-device AI in smartphones, and robotics.
To solve this, researchers are moving toward a paradigm shift in hardware architecture. Rather than moving data to the processor, the industry is developing materials and architectures that allow memory and computation to happen in the same place. According to available source materials, Professor Woo-young Shim has been recognized with the Korea Scientist and Engineer Award for developing a next-generation semiconductor base technology utilizing a material that can perform both memory and AI operations simultaneously, a breakthrough designed to significantly reduce the power consumption required for AI calculations.
This approach aligns with a broader industry trend toward “Processing-in-Memory” (PIM), a technology that seeks to eliminate the energy-intensive transit of data. By integrating computational logic directly into the memory chip, the system can process information where it resides, drastically cutting latency and power draw. This shift is essential for the next wave of AI, moving from massive cloud-based clusters to efficient, autonomous “edge” devices.
The ‘Big Three’ of Next-Generation AI Semiconductors
The semiconductor industry is currently pivoting toward three core technologies to sustain the AI boom: High Bandwidth Memory (HBM), Compute Express Link (CXL), and Processing-in-Memory (PIM). These are often referred to as the “next-generation semiconductor big three” due to their ability to handle the explosive growth of machine learning and big data according to industry reports.
While HBM is already widely deployed to provide the massive data throughput required by GPUs, CXL and PIM represent the next frontier. CXL focuses on expanding memory capacity and connectivity, allowing processors to share memory pools more efficiently. PIM, however, represents a more fundamental change in how a chip works. By performing calculations within the memory itself, PIM targets the remarkably root of the power efficiency problem, making it a critical component for the future of sustainable AI.
Major industry players, including Samsung Electronics and SK Hynix, are aggressively expanding the ecosystems for CXL and PIM to prepare for the next market “up-turn.” The goal is to move beyond traditional architectures to a system where data processing is distributed and highly efficient, reducing the reliance on power-hungry external processors.
From Data Centers to the Physical World: On-Device AI
The development of low-power, integrated memory and computation materials is not just about saving electricity in server farms; it is the key to unlocking true autonomy in robotics. For a robot to operate in real-time without a constant connection to the cloud, it needs “on-device AI” that can perceive and judge its environment using minimal power.
A practical application of this trend was recently showcased at CES 2026, where the AI fabless company DeepX and the Hyundai Motor Group unveiled an on-device AI chip. This chip operates on ultra-low power—under 5W—allowing robots to perform cognitive tasks and decision-making locally without needing a cloud connection as reported by Asia Economy. This level of efficiency is only possible when the architectural gap between memory and processing is minimized.
The integration of such chips into real-world robotics marks a transition from AI as a software service to AI as a physical capability. By reducing the power overhead, these semiconductors enable robots to operate for longer periods and react faster to their surroundings, which is critical for safety and utility in industrial and domestic settings.
The Global Scale: Integrated Design and ‘Terafabs’
As the demand for these specialized AI chips grows, the method of producing them is also evolving. The industry is moving toward a more integrated model where design, manufacturing, and packaging happen in a single, streamlined pipeline to accelerate development cycles.
A prime example of this is the “Terafab” project, a collaboration between Intel and companies led by Elon Musk, including Tesla, SpaceX, and xAI. This project involves the construction of a massive, integrated production facility in Austin, Texas, designed to handle the entire lifecycle of high-performance AI chips from design to packaging according to Ebn News. These chips are slated for use in Tesla’s robotaxis and the “Optimus” humanoid robot, as well as AI-enhanced satellites for SpaceX.
This move toward vertical integration reflects the urgency of the AI race. When the underlying material—such as the simultaneous memory-computation materials being developed by researchers like Professor Shim—is optimized, the ability to rapidly prototype and mass-produce those designs in a “Terafab” environment can create a significant competitive advantage.
Comparison of Next-Gen AI Memory Technologies
| Technology | Primary Function | Key Benefit | Current Status |
|---|---|---|---|
| HBM | High-speed data transit | Massive bandwidth | Commercialized/Widespread |
| CXL | Memory expansion/pooling | Increased capacity | Ecosystem expanding |
| PIM | In-memory computation | Ultra-low power/latency | Development/Early adoption |
What In other words for the Future of Computing
The transition toward materials that merge memory and computation represents a departure from the Von Neumann architecture, which has dominated computing for decades. In the traditional model, the CPU and memory are separate, and the constant shuffling of data between them creates the “bottleneck” that slows down AI and drains batteries.
By breaking this wall, the industry is paving the way for several breakthroughs:
- True Edge Intelligence: Devices that can run complex LLMs locally without needing an internet connection or a massive battery.
- Sustainable AI: A drastic reduction in the carbon footprint of AI data centers by lowering the energy cost per inference.
- Advanced Robotics: Humanoid robots capable of millisecond-level reactions and autonomous navigation using ultra-low-power chips.
The recognition of researchers like Professor Woo-young Shim highlights the importance of material science in this journey. While software algorithms often secure the spotlight, the physical limits of silicon and electricity are the real boundaries of AI. Innovations in semiconductor materials are the only way to ensure that the AI revolution does not collapse under its own energy demands.
As these technologies move from the laboratory to the fab, the focus will shift toward scalability and integration. The next major checkpoint for the industry will be the full-scale deployment of PIM and CXL-based systems in commercial AI servers and the first generation of mass-produced humanoid robots powered by these integrated chips.
Do you think on-device AI will eventually replace cloud computing for most daily tasks? Share your thoughts in the comments below or share this analysis with your network.
Keep reading