Lowering the Cost of Intelligence: NVIDIA's Ian Buck on Mixture of Experts
Ian Buck, VP of Hyperscale and HPC at NVIDIA, joins host Noah Kravitz to explain why Mixture of Experts (MoE) has become the dominant architecture for frontier AI models and how NVIDIA's hardware-software co-design drives down the cost of intelligence. Buck walks through MoE fundamentals — splitting a model into specialized "experts" with a router so only a fraction of parameters activate per query — citing concrete examples: Llama 405B activates all 405B parameters and scores ~28 on Artificial Analysis at ~$200 per benchmark run, while GPT-OSS's 120B-parameter model activates only ~5B parameters, scores 61, and costs ~$75. He credits DeepSeek R1, released about a year prior, as the catalyst that proved world-class MoE was possible at scale and triggered the current wave of MoE-based releases, including the trillion-parameter Kimi K2.
Original sources
Podcast NVIDIA AI Podcast
Episode:Lowering the Cost of Intelligence With NVIDIA's Ian Buck - Ep. 284 ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.