OpenAI x Broadcom and the future of compute - Episode 8

The episode opens with a dramatic announcement of a partnership between OpenAI and Broadcom, emphasizing the scale of their joint efforts in AI infrastructure. Over the past 18 months, they've been designing a custom AI chip and a full system, aiming to deploy 10 gigawatts of this infrastructure to serve the global needs of advanced AI. The discussion highlights that this project is one of the largest and most complex industrial endeavors in history, fundamentally aiming to redefine the computing systems needed for future AI advancements and potentially for civilization-scale infrastructure.

The conversation details how OpenAI and Broadcom are working together on a full-stack, vertically integrated approach— from transistor etching to system optimization and deployment. OpenAI's focus on workload-specific chip design is paired with Broadcom's expertise in system and network design. They explore how this collaboration allows them to innovate beyond traditional methods, optimize efficiencies, and push the boundaries of AI compute hardware, which is critical for scaling AI capabilities and reducing costs.

The speakers compare the current AI infrastructure build to historic large-scale projects like the railroads and internet, underscoring the unprecedented scale and collaboration required today. They discuss the massive energy demands, noting that the new system of chips and data centers will operate on an infrastructure equivalent to a small city, with plans to deploy massive capacity over the coming years. The conversation emphasizes that such infrastructure is essential for achieving superintelligence and that it will take decades to fully realize this vision.

A central theme is how compute power is driving progress toward AGI, with the focus shifting from purely conceptual ideas to practical, hardware-optimized solutions. They address how targeted hardware design— including custom chips for training and inference— can unlock new possibilities in AI applications, increase efficiency, and enable more widespread adoption. Additionally, they discuss the importance of open standards and collaboration across the industry to build an ecosystem that serves humanity broadly, not just individual companies.

Looking ahead, the speakers reveal plans to continually enhance compute capabilities through innovations such as stacking chips in three dimensions and integrating optical switching into hardware. The timeline points to start deploying these new systems around the end of the next year, with rapid scaling over three years. They also highlight that these advances will enable AI to become more accessible and impactful across all aspects of life, emphasizing the transformative potential of AI infrastructure for economic growth and global utility.