Episode Details

When thousands of GPUs must act as one, the network — not the chips — becomes the bottleneck. This hour-long OpenAI Podcast episode (now condensed to a 10-minute summary) unpacks how Andrew Mayne talks with Mark Handley and Greg Steinbrecher about redesigning data-center networks for large-scale AI training. Learn how Multipath Reliable Connection (MRC) sprays traffic across paths, uses packet trimming to resolve loss vs. delay, and shifts path selection to endpoints with source routing and IPv6 segment routing. You’ll get practical takeaways on reducing tail latency, improving reliability amid millions of optical links, lowering power and cost with flatter topologies, and why these networking advances matter for artificial intelligence, machine learning, and supercomputing. Hear real operational wins—fewer interrupted runs, simpler switch software, and faster reaction times—plus how OpenAI is publishing an open standard to drive broader adoption. Listen now to get the key ideas in minutes.