NVIDIA NVIDIA
July 10, 2026

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

Extreme Co-Design: Building the AI Factory | Kevin Deierling, SVP Networking, NVIDIA | Computex 2026

In this keynote from the Computex Forum in Taipei, NVIDIA SVP of Networking Kevin Deierling breaks down what it actually takes to build an AI factory — and why extreme co-design across the entire data center stack is the only way to deliver the tokens the agentic AI era demands.

Deierling opens with three converging transformations reshaping computing: the decades-long shift to accelerated computing, the explosion of generative AI, and the rise of agentic reasoning. Together, they're forcing a complete rearchitecture of the data center — from a cost center into a token revenue generator.

Using Jensen Huang's "5-layer cake" framework, he walks through how NVIDIA co-designs across every layer

Energy — AI training workloads create dramatic power spikes. NVIDIA's DSX software smooths energy demand, enabling 40% more GPUs per gigawatt — and 40% more token revenue.

Chips — NVIDIA's 7 co-designed chips work as one system: Rubin GPUs and Vera CPUs connected via cache-coherent links, scaling up with NVLink and out with Spectrum-X Ethernet through BlueField-4 and ConnectX SuperNICs. This extreme co-design delivers 10x more tokens at 1/10th the cost.

Infrastructure — 72 GPUs unified into one giant computer via NVLink 72, running liquid-cooled and DC-powered in the same rack form factor. And with Spectrum CPO (co-packaged optics), NVIDIA brings the optics next to the ASIC to eliminate signal degradation — saving tens of megawatts across an AI factory. Networking scales from 800Gbps to 1600Gbps.

Agents — Agents are AI eating its own tail: instead of waiting on humans, AIs prompt AIs, closing the loop at machine speed. The Vera CPU is built specifically for agentic tools — planning, writing programs, summarizing results. Agents need memory: KV cache for short-term (CMX) and files, objects, and vector databases for long-term (STX). This new class of AI storage delivers 5x faster token throughput and 5x better efficiency.

Models & Applications — 30x performance improvement on DeepSeek using Blackwell. Open NEMATRON models for partners to customize into proprietary supermodels. CUDAX libraries — CuLitho, CuOpt, CuDF — accelerating every workload.

Deierling closes with a message for every business: agentic AI isn't a future concept, it's available now. Everything — robots, medical equipment, self-driving cars, satellites, coding — will become agentic.

#nvidia #Computex2026 #GTCTaipei #AIFactory #AgenticAI #BlueField #SpectrumX #Networking #DataCenter #ExtremeCoDesign
248

Comments 0

Replying to
5000 max

Be the first to comment!

This website uses cookies
This website uses cookies.
To maintain our services, we use cookies. They allow us to personalize ads, analyze our traffic, and provide social media features.

Customize your preferences

Select the cookies you accept: