Conditions for XPU to Build World-Class AI Factories

Large-scale AI factories must optimize cost per token and utilization rates, requiring design at the full platform level rather than just for individual accelerators. Hyperscalers and AI-native companies must consider networking, rack architecture, and software when developing custom XPUs, but this poses a major obstacle to time-to-market.

NVLink Fusion addresses these challenges by connecting custom XPUs to NVIDIA's proven AI infrastructure. This allows innovation to focus on core components while leveraging mature technology for the rest of the infrastructure, thereby reducing risk.

In terms of scale-up performance, 6th-generation NVLink delivers 3x lower latency and 10x higher packet throughput across 72 XPU domains compared to existing Ethernet solutions. Additionally, NVLink-C2C provides 6x higher energy efficiency for connections between XPUs and CPUs compared to PCIe, removing barriers between control and compute in agentic AI systems.

During the development and deployment phases, leveraging the same supply chain as the MGX rack-scale architecture increases manufacturing automation rates and enables the management of complex supply chain ecosystems. This reduces schedule risks associated with dependency on a single chip during AI factory planning and ensures stability through infrastructure standardization.