The trajectory of AI hardware has largely been defined by the needs of LLMs, yet the most profound breakthroughs in science require silicon that natively understands more than just the Transformer. This panel explores how to identify the specific primitives of scientific discovery that can be radically accelerated by novel silicon architectures —such as equivariant symmetries, sparse graph message-passing, and high-fidelity physics constraints. We will debate how the scientific community can leapfrog current hardware limitations, defining the requirements for a new generation of accelerators designed to handle the unique stressors of scientific foundation models and autonomous agents.

Realizing this vision requires more than just novel ideas; it demands a new lifecycle for hardware validation that integrates silicon design directly into the scientific workflow. Panelists will discuss strategies for testing and validating experimental architectures within the rigorous environments of national labs and academic research, exploring how AI itself can be leveraged to accelerate the hardware design loop and automate the path from prototype to production. This this session aims to provide a roadmap for moving specialized silicon into real-world production, ensuring that the next generation of foundation models is powered by hardware as sophisticated as the science it seeks to solve.