Summary
On No Priors, Walter Goodwin, founder and CEO of full-stack AI chip company Fractile, talks with host Sarah about building very fast inference chips for the world’s largest models. Fractile started in summer 2022, betting that models would need much more compute at test time and would run at ever longer contexts.
Goodwin describes the current accelerator landscape as less diverse than it looks. Google’s TPU, Meta’s MTIA, Microsoft’s Maya and OpenAI’s custom chip are largely developed with a small number of ASIC houses such as Broadcom, and they share HBM memory, tensor cores and TSMC advanced packaging. Fractile’s focus is memory bandwidth, with a path to very high bandwidth from low-cost DRAM, because Goodwin argues that data-center inference economics come down to cost per gigabyte of memory.
The conversation also covers how quickly chip architecture can adapt when new models arrive roughly every two weeks, why autoregressive low-batch LLM workloads keep asking for more memory bandwidth, and how open-source Chinese models are shifting the architecture frontier. Goodwin also weighs how Nvidia, AMD, hyperscaler in-house efforts and new accelerators may divide the market, and why a rolling set of chip bets could create a lasting advantage.
📺 Source: No Priors: AI, Machine Learning, Tech, & Startups · Published October 02, 2026
🏷️ Format: Interview
