NVIDIA NVLink Fusion Pushes XPU AI Factories to New Heights
NVIDIA has unveiled NVLink Fusion, a groundbreaking solution aimed at integrating custom XPUs into world-class AI infrastructures. The technology promises to accelerate time to market for semi-custom AI factories, reduce operational risks, and maximize performance for next-gen workloads like trillion-parameter models and agentic AI systems.
At its core, NVLink Fusion connects XPUs—an umbrella term for heterogeneous processing units like CPUs, GPUs, and AI-specific ASICs—to NVIDIA's mature AI infrastructure. This combination allows hyperscalers and AI-native companies to focus innovation on chip design while leveraging NVIDIA's proven networking, rack-scale architecture, and production software for the rest.
Why It Matters: Scaling AI Factories Efficiently
AI factories are massive, continuously running facilities where efficiency is measured in terms of tokens per second, tokens per watt, and cost per token. However, building these systems from the ground up is a costly and complex endeavor. NVIDIA's NVLink Fusion addresses these challenges by streamlining the integration of custom chips into standardized, scalable systems.
For example, NVLink Fusion uses NVIDIA’s sixth-generation NVLink for ultra-fast XPU-to-XPU networking. This setup delivers 3x lower latency and 10x higher packet rates compared to traditional Ethernet solutions. Through its GB300 NVL72 systems, NVIDIA claims significantly better throughput and interactivity, critical for demanding AI workloads.
Moreover, NVLink Fusion improves energy efficiency with NVIDIA NVLink-C2C, offering up to 6x better performance than PCIe interfaces when connecting XPUs to CPUs like NVIDIA’s Vera processors. This level of integration is a game-changer for agentic AI systems that require seamless coordination between compute and control.
Proven Ecosystem Reduces Risk
Building custom XPUs is only part of the equation. Deploying them into operational AI factories demands robust infrastructure, including high-speed networking, rack design, power management, and a reliable supplier ecosystem. NVLink Fusion simplifies this process by providing a platform validated for rapid development and deployment.
The ecosystem includes NVIDIA MGX rack-scale architecture, which supports modular configurations and automation in manufacturing. Notably, manufacturing partners like QCT and Quanta Computer benefit from near-total system automation, reducing time-to-market for hyperscalers.
“Most of those investments can be leveraged if the XPU leverages NVLink Fusion,” said Jack Luoh, head of product and solution at QCT and Quanta Computer.
Unified Architecture for AI Flexibility
A key feature of NVLink Fusion is its flexibility to support heterogeneous systems. AI workloads today often require different accelerators—GPUs for training, XPUs for inference, and CPUs for orchestration. By enabling these systems to share common infrastructure, NVLink Fusion allows operators to adjust silicon mixes and workloads based on availability and demand without locking into a single chip design.
“NVLink Fusion allows hyperscalers to integrate their custom designs while bridging NVIDIA technology with third-party processes to create a unified architecture,” said Lie-Szu Juang, chair of GUC.
Takeaway: Setting the Standard for Scalable AI
As AI workloads scale, the industry faces increasing pressure to optimize infrastructure costs while maintaining high performance. NVIDIA’s NVLink Fusion offers a path forward, enabling customized hardware to coexist with standardized, proven systems. This approach not only accelerates deployment but also future-proofs investments by allowing flexibility in silicon and workload configurations.
For hyperscalers and AI-native companies aiming to build next-generation AI factories, NVLink Fusion could become a critical enabler, combining specialized innovation with infrastructure standardization. With its focus on efficiency, scalability, and risk mitigation, NVIDIA is setting the benchmark for the future of AI infrastructure.