At 3C AGI Partners, we pride ourselves on being at the forefront of transformative innovation. After reading an intriguing article about the world’s fastest memory, capable of writing 25 billion bits per second—10,000× faster than current technology—we set out to visit Fudan University’s Frontier Institute of Chip and System and had the privilege of sitting down with the team behind 'PoX' (破晓), a breakthrough in memory technology that could redefine the future of computing.

Our conversations with the researchers offered a unique perspective into how their work addresses the critical "memory wall," a bottleneck that has long-limited computing performance. This challenge is particularly pressing in fields like AI, where massive neural networks demand extraordinary data exchange capabilities between memory and processing units.

Here’s what we learned from our deep dive into their technology and vision.

The Memory Bottleneck: A Critical Challenge in AI

The Fudan team emphasized a key insight that resonated with us: while the semiconductor industry has achieved exponential growth in processing power, memory technologies have failed to keep pace. This gap has created a "memory wall," where processors often idle while waiting for data—a problem that becomes even more pronounced in AI applications.

Current Challenges in AI Memory:

• Existing GPUs, the backbone of AI computations, are limited by small memory capacities, which restrict the size and complexity of AI models they can run.

• While technologies like High Bandwidth Memory (HBM) offer short-term solutions, they add complexity, increase power consumption, and come at a high cost.

From Fudan’s perspective, the next generation of AI will require a fundamental rethink of how memory technology works instead of just focus on expanding raw computing power.

A Decade-Long Effort to Break the Memory Bottleneck

One thing that stood out is that the team’s work isn’t a reactive response to the recent surge in AI demand. Their research began over a decade ago, around 2010, rooted in a long-term vision to address the memory bottleneck. Over the years, their progress has been iterative and methodical:

• 2018: They achieved a 10-nanosecond programming speed but struggled with retention characteristics, which lasted only tens of seconds, published in Nature Nanotechnology.

• 2021: Improved the speed to 20 nanoseconds while increasing retention characteristics to 10 years of non-volatility, published in Nature Nanotechnology.

• 2023-2024: Demonstrated that 2D flash memory could achieve a lifespan of 8-10 million cycles—far surpassing the 1-10 thousand cycles typical of 3D flash memory—and successfully achieved integrated designs, published in Nature Nanotechnology3 and Nature Electronics.

This decades-long effort underscores their commitment to solving the memory bottleneck at a fundamental level, well before the current AI wave brought heightened attention to the field.

"PoX" Technology: A Revolutionary Approach to Memory

During our visit, the researchers shared the story behind their groundbreaking "PoX" memory technology, which has already been published in Nature5. What struck us most was both the simplicity and the ambition of their approach: reimagining memory at a fundamental, physical level.

Here are the key takeaways from their work:

Core Technical Innovation:

• Speed and Efficiency: Using a novel approach that combines 2D Dirac band structures with ballistic transport characteristics, the team has achieved a processing speed of 400 picoseconds. This is up to 10,000x faster than today’s flash memory.

• Consistency Between Storage and Processing: The speed of PoX memory rivals the internal frequencies of top-tier processors, which could eliminate the current performance mismatch between computation and storage.

Material and Industrial Feasibility:

• The team’s focus on industrially viable materials stood out. They’ve carefully selected materials that are mass-producible and compatible with existing semiconductor processes, ensuring scalability from lab to market.

• By integrating PoX with CMOS technology, they’ve already fabricated a Kb-level chip, with plans to scale to MB-level in the next 3-5 years.

Transformative Applications:

The potential applications of PoX are as groundbreaking as the technology itself. By addressing the memory bottleneck, PoX could fundamentally reshape computing systems in several ways:

• With appropriate packaging forms, all chips can use PoX storage

• Will fundamentally change electronic product storage architecture, eliminating the boundary between memory and external storage

• Greatly improve AI deployment efficiency, solving current local model memory limitation problems

• Significantly reduce system power consumption while increasing processing speed

Our Take: Why This Matters to Investment

At 3C, we see PoX as more than just a technological breakthrough — it’s a glimpse into the future of AI infrastructure. The team at Fudan is addressing one of the most pressing challenges in computing, and their work has the potential to unlock entirely new markets.

Here’s what excites us most about its potential:

• AI Deployment Revolution

With PoX, much larger AI models could run locally on devices, eliminating reliance on cloud-based processing. This could democratize access to advanced AI capabilities, bringing powerful tools to edge devices like smartphones and IoT hardware.

• Simplified Computing Architectures

PoX could eliminate the need for traditional memory hierarchies (SRAM, DRAM, flash), simplifying system designs and reducing data movement bottlenecks.

• Energy Efficiency

By minimizing energy-intensive data transfers between memory and storage, PoX could significantly reduce power consumption in computing systems—an increasingly critical factor in sustainable tech development.

• Strategic Timing

With commercialization expected in 3-5 years, PoX could arrive just as AI memory demands reach critical levels, positioning it as a timely and much-needed solution.

• Competitive Advantage for China

As a rare fundamental innovation originating from China, PoX has the potential to shift the global semiconductor value chain. It could reduce dependence on dominant players like Samsung, Micron, and SK Hynix, giving China a strategic edge in the memory technology space.

As we continue to explore the intersection of AI, semiconductor technology, and sustainability, we’re excited to see how technologies like PoX will shape the next generation of computing—and we’re proud to have had the chance to witness its development firsthand and look forward to staying connected with the Fudan team as they continue to push the boundaries of what’s possible in memory technology.

If you’d like to learn more about the Fudan team and their work, or discuss investment opportunities in foundational technologies, get in touch with us at hello@3cagi.vc. 

Reference:

1. Nature Nanotechnology, A semi-floating gate memory based on van der Waals heterostructures for quasi-non-volatile applications, Chunsen Liu, Xiao Yan,  Xiongfei Song,  Shijin Ding,  David Wei Zhang & Peng Zhou, https://www.nature.com/articles/s41565-018-0102-6

2. Nature Nanotechnology, Ultrafast non-volatile flash memory based on van der Waals heterostructures, Lan Liu, Chunsen Liu, Lilai Jiang, Jiayi Li, Yi Ding, Shuiyuan Wang, Yu-Gang Jiang, Ya-Bin Sun, Jianlu Wang, Shiyou Chen, David Wei Zhang & Peng Zhou, https://www.nature.com/articles/s41565-021-00921-4

3. Nature Nanotechnology, An ultrafast bipolar flash memory for self-activated in-memory computing, Xiaohe Huang, Chunsen Liu, Zhaowu Tang, Senfeng Zeng, Shuiyuan Wang & Peng Zhou, https://www.nature.com/articles/s41565-023-01339-w

4. Nature Electronics, A scalable integration process for ultrafast two-dimensional flash memory, Yongbo Jiang, Chunsen Liu, Zhenyuan Cao, Chuhang Li, Zizheng Liu, Chong Wang, Yutong Xiang & Peng Zhou, https://www.nature.com/articles/s41928-024-01229-6

5. Nature, Subnanosecond flash memory enabled by 2D-enhanced hot-carrier injection, Yutong Xiang, Chong Wang, Chunsen Liu, Tanjun Wang, Yongbo Jiang, Yang Wang, Shuiyuan Wang & Peng Zhou, https://www.nature.com/articles/s41586-025-08839-w