NVIDIA NVLink Fusion Brings Custom XPUs Into Unified AI Factories
NVIDIA NVLink Fusion links custom XPUs to NVLink, MGX and DSX infrastructure, promising faster deployment, lower latency and flexible AI factories at scale.
Summary
On August 24, 2026, NVIDIA positioned NVLink Fusion as a route for hyperscalers and AI-native companies to connect custom XPUs with established scale-up and scale-out networking, rack architecture, software and suppliers. AI factories are measured by tokens per second, tokens per watt, cost per token, utilization and uptime. For trillion-parameter models, mixture-of-experts systems and agentic AI, sixth-generation NVLink spans 72 XPUs, with NVIDIA claiming 3x lower XPU transfer latency and 10x higher packet rates than off-the-shelf Ethernet. GB300 NVL72 delivers higher per-GPU throughput and interactivity than B300 in NVIDIA benchmarks. The roadmap reaches 1,152 accelerators with co-packaged optics, while NVLink-C2C connects XPUs to Vera or ecosystem CPUs at up to 6x PCIe's energy efficiency.
NVLink Fusion adopters can use NVIDIA MGX architecture and the Vera Rubin NVL72 supply chain for compute and switch trays, security, storage, cooling, power and emerging 800 VDC designs. Jack Luoh of QCT and Quanta Computer expects almost 100% manufacturing automation for Vera Rubin NVL72, with investments reusable for Fusion XPUs. Intel's Tim Wilson emphasized CPU, performance and software choice. Manufacturing partners handle integration while ASIC, CPU, IP and optical partners support deployment.
XPU and GPU systems can share rack footprints, networks, cooling, power and management, letting operators build facilities before fixing the silicon mix, then reprovision for workloads or supply. MediaTek's Vince Hu, GUC's Lie-Szu Juang and Annapurna Labs, an Amazon company, manager CC Lee highlighted decoupled XPU schedules, third-party integration, multiple suppliers and faster delivery. NVIDIA DSX and the Omniverse DSX AI Factory Blueprint model gigawatt-scale facilities before construction. Reference trays use 100% liquid cooling without fans, cables or hoses and remain operational during servicing. NCCL, Dynamo, NIXL and Mission Control coordinate distributed workloads, disaggregation, management, telemetry and debugging.
Positives
- Sixth-generation NVLink supports 72 XPUs with 3x lower transfer latency and 10x higher packet rates than off-the-shelf Ethernet alternatives.
- NVLink-C2C offers up to 6x PCIe's energy efficiency when connecting XPUs with NVIDIA Vera or ecosystem CPUs.
- The NVLink roadmap expands scale-up domains to 1,152 accelerators and adds co-packaged optics.
- MGX and Vera Rubin NVL72 infrastructure lets custom XPUs reuse established racks, cooling, power, suppliers and manufacturing investments.
- NVIDIA DSX and Omniverse provide digital twins and reference designs for validating gigawatt-scale AI factories before construction.
- Reference trays provide 100% liquid cooling and permit component servicing while the remaining rack continues operating.
Risks & concerns
- Custom XPU deployment requires complex integration across CPU interfaces, networks, trays, racks, cooling, power, security, storage and suppliers.
- Slow scale-up fabrics reduce accelerator utilization and increase cost per token for trillion-parameter, mixture-of-experts and agentic workloads.
- AI factory construction begins before accelerator choices are final, making facilities tied to one chip a schedule risk.
- Factory mistakes can force expensive rework, increasing the importance of validating infrastructure before construction.
- The cited latency, packet-rate and GB300 performance comparisons rely on NVIDIA's own configurations and benchmark results.