OpenAI x Broadcom — The OpenAI Podcast Ep. 8

Quick Overview

OpenAI and Broadcom announced a major partnership involving the design of a new custom chip and a full system approach, culminating in the planned deployment of 10 gigawatts of computing infrastructure starting late next year to serve global advanced intelligence needs.

Key Points: OpenAI and Broadcom are partnering to design a new custom chip specific for OpenAI's workloads, which evolved into designing a whole custom system over 18 months of collaboration. The partnership plans to deploy 10 gigawatts of these new systems and chips starting in late next year to serve the world's need for advanced intelligence. Sam Altman stated that optimizing across the entire stack, from transistors to the final token output, allows for huge efficiency gains, leading to faster and cheaper models. Greg Brockman noted that OpenAI applied its own models to chip design, achieving optimizations that human designers would have taken much longer to reach, accelerating the schedule. Hock Tan characterized the collaboration as natural, positioning the resulting infrastructure as critical utility over time for 8 billion people, like the industrial revolution of a different sort. The scale of the AI infrastructure build-out is described by Sam Altman as potentially the biggest joint industrial project in human history. Charlie Kawwas detailed the technical progression from custom AI accelerators (XPUs) to integrating networking and ultimately incorporating optics for 100 terabits of switching integrated into the same chip.

Context: The OpenAI podcast episode featured Andrew Mayne hosting Sam Altman and Greg Brockman from OpenAI, alongside Hock Tan and Charlie Kawwas from Broadcom, to announce a significant strategic partnership focused on scaling AI compute infrastructure. The discussion centered on the necessity of vertically integrating the design process—from transistor-level chip design to system architecture—to meet the exponentially growing demand for advanced AI models like GPT-6 and future iterations.