OXMIQ CEO Raja Koduri on Re-Architecting the GPU Stack: From Atoms to Agents скачать в хорошем качестве

OXMIQ CEO Raja Koduri on Re-Architecting the GPU Stack: From Atoms to Agents

The cost of AI infrastructure is no longer theoretical. In this episode of the Semiconductor Leadership Podcast, Salah Nasri sits down with Raja Koduri, Founder and CEO of Oxmiq Labs, to unpack the real constraints shaping next-generation AI silicon—where bandwidth, physics, and economics collide. As Raja explains, the silicon budget for deploying modern AI infrastructure has surpassed $30 billion per gigawatt, forcing architects to confront hard tradeoffs that no standard or abstraction layer can fully solve. This conversation goes beyond FLOPS and benchmarks to explore: Why bandwidth—not compute—is the true bottleneck in AI systems The limits of standardization as systems scale The hidden cost of premature optimization in hardware-software co-design Why simplifying software by burning more transistors is a choice Raja refuses to make How rapid innovation leaves even critical architectural work behind Lessons from failure, including insights from Intel’s Ponte Vecchio project At its core, this episode asks a deeper question facing the semiconductor industry today: Do we choose freedom in design—or fragmentation at scale? For leaders building AI accelerators, data center infrastructure, or next-generation silicon platforms, this is a candid, systems-level conversation about what breaks when assumptions stop holding. 🎧 Subscribe to the Semiconductor Leadership Podcast on YouTube, Spotify, and Apple Podcasts for more conversations with global technology and investment leaders. 0:00 Intro 4:20 What problem is OXMIQ solving first 5:42 Hardware or Software 8:12 What is OXCORE 9:16 What is OXPython 18:05 Most overrated and underrated GPU metric 26:18 Freedom of Fragmentation 30:47 One design choice Raja will never compromise on 31:37 One KPI that proves TCO actually improved 33:45 A partnership that would 10x QXMIQ roadmap 36:39 What to do with 10x budget 37:37 Most overhyped and under-hyped term in AI compute 40:05 Failure that improved playbook 46:46 "Hello, world!" for OXMIQ developers 48:33 The greatest GPU software 50:07 One piece of advice 53:05 Career journey 1:00:19 The perfect timing for OXMIQ 1:03:09 Concerns over players in the game 1:07:50 Rebuilding from atom to agents 1:19:41 What success looks like 1:27:38 Philosophy behind OXCORE 1:32:14 OXPython possibilities and limits 1:35:54 What is OXCapsule 1:47:02 What's next 1:49:00 Investment capability or market access 1:51:16 Scaling with talent bottleneck 1:55:54 Top 2 technical D-risks 1:58:05 Competition, pricing, innovation 2:01:23 One piece of advice 2:04:38 Raja's vision for OXMIQ