@Scale Conferences podcast

Powering the Frontier, Inside OpenAI's Networking Infrastructure | Jitendra Padhye

0:00
46:10
Rewind 15 seconds
Fast Forward 15 seconds

What does it take to build the networking infrastructure powering frontier AI?

In this episode of the @Scale Podcast, Jitendra Padhye, Networking Architect at OpenAI, joins the conversation to explore the evolving challenges of designing and scaling high-performance networks for large AI training clusters.

Drawing on his experience at AT&T, Microsoft Research, Azure, and OpenAI, Jitendra shares how networking requirements change when moving from traditional cloud environments to infrastructure built around frontier AI workloads.

The conversation explores Multi-Rail Congestion Control (MRC), congestion management, packet reliability, and the engineering innovations needed to keep massive AI training workloads running efficiently. He also discusses the challenges of detecting network failures, optimizing data movement across multiple paths, and navigating the power and space constraints shaping the next generation of AI infrastructure.

In this episode:

- The networking challenges behind scaling frontier AI training clusters

- How Multi-Rail Congestion Control (MRC) helps manage network traffic at scale

- Why packet-level reliability is critical for large AI workloads

- The differences between building infrastructure for public cloud and AI-specific workloads

- How power, space, and compute demands will shape the future of AI infrastructure

Tune in for an inside look at the engineering challenges and innovations powering the next generation of AI.

Subscribe for more conversations with the engineers and technology leaders building systems at scale.

More episodes from "@Scale Conferences"