NVIDIA expands its AI factory via integrated GPU, CPU, and LPU racks for training and inference, securing its market lead.

Key Highlights

• Hardware Racks: Shifts to integrated GPU, CPU & LPU racks, mastering both AI training and low-latency agentic inference.
• Architecture: Pioneers disaggregated inference, liquid cooling, and copper/optical CPO networks to shatter bandwidth limits.
• Defense: Counters custom cloud chips via deep software-hardware lock-in, cementing its AI dominance.

Table of Contents

1. Introduction
   • NVIDIA's Product Roadmaps to Segment into GPUs, CPUs, and LPUs when Targeting the AI Market
2. Vera Rubin Rack Expected to Gradually Replace GB300 after Release in 2H26
   • NVIDIA's Projected Quarterly Shipment of GB/VR Racks between 2025 and 2026
3. NVIDIA Also Proposed Vera CPU to Strengthen Agentic AI Infrastructures
4. NVIDIA Introduces the LPU as a Core Component in Its AI Disaggregated Inference Architecture
   • NVIDIA Introduces Disaggregated Inference LPX Rack Design That Is Distinct from VR Rack
5. NVIDIA's Diverse Next-Generation Chips Will Underpin Future Memory Growth
6. NVIDIA Leverages CPO to Enhance High-Speed Interconnect Performance of Rack-Scale AI Chip Solutions and Whole Rack Systems
   • Overview of NVIDIA's Chip Generations and Rack Interconnect Architectures
7. Conclusion: Facing the Challenge of Rising Market Share for CSPs' In-House ASICs, NVIDIA Pushes Integrated Rack Solutions Across CPUs, GPUs, and LPUs to Accelerate Its Expansion from AI Training into Inference

<Total Pages: 10>