Extreme Co-Design
Extreme co-design is NVIDIA’s practice of jointly optimizing algorithms, software, chips, systems, networking, storage, power, cooling, racks, supply chain, and data-center design because modern AI workloads no longer fit inside one computer.
Key points
- Jensen says AI scaling requires refactoring algorithms, sharding pipelines, data, and models, and solving CPU, GPU, networking, switching, and workload-distribution problems together [src-065].
- The reason is Amdahl-like: accelerating only the compute part does little if networking, data movement, power, or software becomes the bottleneck [src-065].
- NVIDIA’s organizational design mirrors this: Jensen keeps a large technical staff across memory, CPU, optical, GPU, architecture, algorithms, power, and cooling, and problems are attacked in group conversations rather than isolated one-on-ones [src-065].
- Extreme co-design is also the mechanism for improving Tokens-Per-Watt Economics as power and supply-chain constraints become central to AI scaling [src-065].
Related entities
Related concepts
- AI Factories
- Tokens-Per-Watt Economics
- GPU Supply as AI Strategy
- Scale-Up vs Scale-Out Networking
- Belief-System Shaping Leadership
Source references
- [src-065] Lex Fridman – “Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494” (2026-03-23)
Recommended next
Keep reading from this thread
From 491 indexed pages and articles.
- Wiki concept AI Factories Large-scale computing facilities that transform energy, chips, networking, storage, cooling, software, and models into AI output such as tokens, predictions, or Related by 065
- Wiki concept Tokens-Per-Watt Economics The view that AI infrastructure competitiveness depends on how many useful AI outputs a system can produce per unit of power Related by extreme
- Insight AI Beyond POCs How enterprise AI moves beyond proofs of concept through ownership, governance, measurement, adoption, and production operating models Readers have engaged with this next