COMPUTEX 2026 | NVIDIA Keynote | Extreme Co-Design: Building the AI Factory

2026-07-10 · Source: NVIDIA · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Robotics & Autonomous Systems · Depth: Advanced, long

Summary

NVIDIA's SVP Networking, Kevin Deierling, presented "Extreme Co-Design: Building the AI Factory" at Computex 2026, highlighting three major transformations: the shift to accelerated computing, the rise of generative AI and agentic reasoning, and the re-architecture of data centers into AI factories. With Moore's Law ending and AI model parameters growing 10x annually, NVIDIA focuses on driving down cost per token. This involves optimizing models from FP16 to NVFP4 and redesigning data centers as "AI factories" across a five-layer stack. Key innovations include smoothing energy demand (DSX Flex, DSX LPS) to deploy 40% more GPUs per gigawatt, co-designing seven chips like the Rubin GPU and Vera CPU for 10x token improvement, and implementing liquid cooling and co-packaged optics for infrastructure. The emergence of AI agents necessitates new AI storage architectures (CMX, STX) for 5x faster token throughput, alongside optimized models and CUDA-X accelerated applications.

Key takeaway

For AI Architects and MLOps Engineers scaling AI infrastructure, NVIDIA's "Extreme Co-Design" framework offers a blueprint for maximizing efficiency. You should evaluate your data center's energy management, chip integration, and cooling solutions, considering liquid cooling and co-packaged optics. Prioritize specialized AI storage (CMX, STX) and agentic frameworks to optimize token throughput and prepare for future AI workloads, ensuring your AI factory generates maximum revenue.

Key insights

NVIDIA's "Extreme Co-Design" optimizes AI factories across a five-layer stack, from energy to applications, to meet escalating AI demand efficiently.

Principles

• Moore's Law's end necessitates accelerated computing.
• Data centers must become token-generating AI factories.
• Co-design across the entire stack drives efficiency.

Method

NVIDIA's "five-layer cake" co-design approach optimizes energy, chips, infrastructure, models, and applications, from bottom-up and top-down, to maximize tokens per watt/dollar.

In practice

• Implement DSX Flex/LPS for energy demand smoothing.
• Utilize co-packaged optics for significant power savings.
• Adopt CMX/STX architectures for agentic AI storage.

Topics

• AI Factory
• Extreme Co-Design
• Accelerated Computing
• Agentic AI
• Data Center Infrastructure
• NVIDIA GPUs

Best for: Investor, CTO, AI Engineer, AI Architect, Director of AI/ML, MLOps Engineer

Related on AIssential

• Extreme Co-Design for Efficient Tokenomics and AI at Scale — Extreme co-design across the entire AI stack is crucial for efficient, scalable real-time reasoning and MoE models.
• NVIDIA CEO: AGI Is Already Here — NVIDIA's CEO declares AGI is here, emphasizing extreme co-design and CUDA's foundational role in scaling AI capabilities.
• The Network at the Heart of AI Factories With NVIDIA Spectrum- X Ethernet​ — AI factories demand purpose-built, multi-infrastructure networks for low-latency, jitter-free, and high-bandwidth distributed computing.
• SpaceX: The AI Orbital Hyperscaler — A massive compute lease deal between Anthropic and SpaceX highlights critical AI infrastructure supply-demand imbalances.
• Computex 2026: Are We Heading for the Agentic PC Era Yet? — Computex 2026 signaled a shift from AI PCs to agentic AI and physical AI appliances, emphasizing Taiwan's pivotal role in the electronics value chain.
• Scaling Agentic AI Factories Through Extreme Co-Design with NVIDIA BlueField — Agentic AI integrates infrastructure directly into the inference pipeline, requiring extreme co-design for optimal performance.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by NVIDIA.