Introduction

On June 4, 2025, NVIDIA unveiled its new AI Factory reference design at Computex, pitching it as the manufacturing floor for next-generation AI. The blueprint combines Blackwell Ultra GPUs, NVLink Switch 5, and DGX GB200 systems into a vertically integrated platform for training and serving trillion-parameter models.

“Think of it as a modern assembly line—inputs are tokens, outputs are intelligence,” said Jensen Huang, NVIDIA’s CEO.¹

The AI Factory stacks GPU pods on Grace-based CPU complexes and disaggregated NVLink memory. A 3 TB/s optical mesh links every rack, while an on-prem “RTX Pro” tier handles real-time inference for humanoid robots and digital twins. NVIDIA claims you can scale from one cabinet to an exaFLOPS super-cluster without rewiring.

Why it matters now

• Hyperscale clouds dominate AI; enterprises crave on-prem sovereignty.
• Energy and floor-space caps push demand for dense, tightly coupled compute.
• NVLink Switch 5 now lets third-party accelerators join NVIDIA fabrics.

Call-out: An AI assembly line for every enterprise

NVIDIA says an AI Factory cuts GPT-5-class training time by 45 % while trimming energy per token by 28 % versus siloed GPU clusters.

Business implications

• Cloud & telco operators can launch regional AI factories to satisfy data-locality laws.
• Automotive & robotics firms gain micro-second latency for humanoids and self-driving stacks.
• CIOs get turnkey blueprints—power, cooling, interconnect, and software validated end-to-end.

Reference racks ship in Q3 to Amazon, Foxconn, and Siemens. The software stack—DGX OS 5 with NeMo Agent and Fleet Command—can already be simulated on standard x86 servers.

Looking ahead

Next iterations will add liquid-cooled NVLink blades and Grace Blackwell Ultra chips, pushing per-rack performance beyond 600 petaFLOPS. NVIDIA also teased “Mini AI Factories” for midsize firms needing private LLM fine-tuning.

Gartner forecasts that by 2028, 40 % of Fortune 500 data centers will adopt factory-style AI layouts, up from 6 % today.

The upshot: NVIDIA’s AI Factory elevates infrastructure from ad-hoc GPU farms to industrial-grade production lines. Intelligence isn’t merely computed anymore; it’s manufactured.

––––––––––––––––––––––––––––
¹ Jensen Huang, NVIDIA Computex Keynote, June 4 2025.