NVIDIA introduces NVLink Fusion to enable companies building custom AI accelerators (XPUs) to integrate them with NVIDIA

NVIDIA's announcement addresses a fundamental challenge in AI factory economics: delivering tokens per second, tokens per watt, and cost per token at scale requires not just custom processors but a complete AI platform encompassing networking, rack architecture, production software and supplier ecosystem. Building this full stack is complex and costly, delaying market entry. NVLink Fusion combines custom XPUs with proven NVIDIA infrastructure, allowing builders to innovate on processors while leveraging established technology for the supporting platform.

The scale-up networking solution must excel across three dimensions: delivered performance through end-to-end network capability and mature software integration; factory resiliency with uptime, continuous health monitoring and serviceability during operation; and platform maturity that reduces operational risk through demonstrated large-scale deployments. NVLink Fusion brings XPUs into NVIDIA's NVLink scale-up domain, with sixth-generation NVLink providing high-bandwidth, low-latency networking across a 72-XPU domain. End-to-end latency for XPU-to-XPU transfers is 3 times lower than alternative Ethernet solutions, and packet rate is 10 times higher. NVIDIA GB300 NVL72 systems deliver significantly higher throughput compared to configurations without NVL72, with future roadmap configurations including domains of up to 1,152 accelerators and co-packaged optics. NVLink Fusion also includes NVLink-C2C for connecting XPUs to NVIDIA Vera CPUs or ecosystem CPUs, delivering up to 6 times the energy efficiency of PCIe interfaces.

Teams developing custom XPUs typically underestimate the complexity of translating innovation into data center deployment, requiring integration of high-speed interfaces, validation of scale-up network solutions, design of compute and switch trays, rack architecture including cooling and power, security and storage integration, and management of complex supplier ecosystems. NVLink Fusion is supported by an ecosystem spanning ASIC design, CPU, and IP and optical interconnect partners. Tim Wilson, vice president and general manager of data center silicon engineering at Intel, noted that NVLink Fusion gives customers the ability to choose CPU architecture, performance level and software capabilities for their workloads. Adopters can use NVIDIA MGX rack-scale architecture and the supply chain used for MGX-based systems such as NVIDIA Vera Rubin NVL72, with manufacturing partners managing design and integration while MGX suppliers provide building blocks for rack, cooling, power and emerging 800 VDC designs. Jack Luoh, head of product and solution at QCT and Quanta Computer, stated that Vera Rubin NVL72 enables almost 100 percent automation of system builds in the manufacturing line, with those investments leverageable if the XPU uses NVLink Fusion.

NVIDIA's AI infrastructure platform is vertically integrated and horizontally open. NVLink Fusion adopters can optionally incorporate NVIDIA Rubin GPUs, Vera CPUs, co-packaged optics switches, ConnectX SuperNICs, BlueField DPUs, Mission Control software and full-rack solutions including NVIDIA Vera Rubin NVL72, Vera CPU Rack, LPX, STX and SPX. AI factory planning begins before silicon finalization, with power procurement, facility design, cooling, rack layout and network architecture requiring early decisions. A data center locked to one chip type becomes a schedule risk, and different workloads may favor different accelerators including XPUs, GPUs, CPUs and LPUs. Vince Hu, corporate senior vice president and general manager of the data center and computing business group at MediaTek, emphasized that NVLink Fusion lets customers deploy rack-level solutions with NVIDIA GPUs while decoupling XPU development at a different pace. NVLink Fusion addresses these challenges through unified architecture: XPU- and GPU-based systems such as Vera Rubin NVL72 can share rack footprints, networking, cooling, power delivery and management systems, allowing operators to proceed with buildout while deferring precise silicon mix and reprovisioning capacity as workload demand, silicon supply and business priorities change. Lie-Szu Juang, chair and chief strategy officer at GUC, explained that NVLink Fusion allows hyperscalers and custom ASIC designers to integrate their own custom CPU or XPU while bridging NVIDIA technology with third-party processes to create unified rack-scale architecture.

Infrastructure must be validated before construction, and NVLink Fusion aligns with NVIDIA's DSX reference architecture for AI factories: codesigning buildings, power, cooling, compute and networking. The NVIDIA Omniverse DSX AI Factory Blueprint provides a digital twin and open reference design for gigawatt-scale AI factories, enabling partners to model facilities and technology together before deployment. At the rack level, serviceability is part of performance. Reference compute trays feature 100 percent liquid cooling with no fans, cables or hoses, allowing tray removal while the rest of the rack remains operational. NVLink Switch trays are also liquid cooled and support continued operation during service. CC Lee, senior hardware development manager at Annapurna Labs, an Amazon company, stated that NVLink Fusion enables use of proven NVL72 rack design for time-to-market while providing access to multiple suppliers to deliver more to customers.

Software completes the factory ecosystem. NVIDIA NCCL for distributed workloads, NVIDIA Dynamo and NIXL for disaggregation, and NVIDIA Mission Control for cluster management, telemetry and debugging help operators run mixed AI infrastructure as a coordinated system. With NVLink Fusion, XPUs can integrate with a world-class AI platform, enabling hyperscalers and AI-native companies to build unified, semi-custom AI factories.