Chips for AI, AI for chips: Redesigning AI from materials up

As AI expands into every industry—supporting billions of people and devices—the field is approaching a critical inflection point. Progress is no longer constrained primarily by algorithms, but increasingly by the fundamental laws of physics. Two structural barriers have emerged: the "memory wall," where data movement consumes significantly more energy than computation, and the "materials wall," where conventional discovery cycles remain measured in decades.

Addressing these challenges requires more than incremental improvements within individual domains. The next phase of AI demands integrated design across the entire intelligence stack—from materials and devices to circuits, architectures, systems, and models. Cross-layer co-design offers a path beyond emerging physical and energy constraints, enabling new levels of efficiency and scalability.

In this context, I will share how our company Preferred Networks (PFN) is advancing a holistic approach spanning from materials to chip design. We will discuss Matlantis, a universal atomistic simulator accelerating semiconductor materials discovery, and MN-Core L1000, a next-generation AI chip designed to confront the memory transfer bottleneck through energy-efficient integration.

Guided by PFN's mission to make the real world computable, this approach forms a reinforcing cycle in which AI helps design the materials and hardware that power the next generation of AI systems.