The Future of Frontier Model Architectures with Walter Goodwin, Founder & CEO of Fractile

Building high-performance AI inference chips requires a fundamental shift toward maximizing memory bandwidth to support the rapid scaling of frontier models. Fractile, led by founder and CEO Walter Goodwin, addresses this by integrating full-stack design capabilities to create chips that handle multi-trillion parameter models at significantly higher speeds than current industry standards. Moving beyond traditional SRAM-based architectures, the company leverages high-bandwidth DRAM to accommodate growing context lengths and model complexity. By compressing the chip design cycle, organizations can maintain a critical three-to-six-month technological advantage, which is essential for winning deployments in a competitive market where model architectures evolve every few weeks. This approach emphasizes the need for specialized hardware that balances computational speed with economic efficiency, ultimately enabling the next generation of long-running, high-speed AI agents.