Nvidia announced NVLink Fusion on Aug. 24, 2026, introducing a system architecture that connects third-party custom XPUs directly into Nvidia scale-up networking and data center infrastructure.
Nvidia said the platform addresses the integration hurdles faced by hyperscalers and AI companies designing proprietary accelerators. Building custom silicon requires scale-up fabric, rack design, cooling, power distribution, and validation software to deploy at factory scale.
Network performance
NVLink Fusion places custom XPUs into Nvidia sixth-generation NVLink scale-up domains across up to 72 accelerators. Nvidia said the end-to-end latency for XPU-to-XPU communication is three times lower than standard Ethernet alternatives, while delivering a tenfold increase in packet rates.
The interconnect also incorporates Nvidia NVLink-C2C to link custom accelerators to Nvidia Vera CPUs or third-party ecosystem processors. Nvidia stated that NVLink-C2C achieves up to six times the energy efficiency of a standard PCIe interface. Future roadmap configurations will expand support to domains of up to 1,152 accelerators alongside co-packaged optics.
“NVLink Fusion gives customers the ability to choose the CPU architecture, the performance level, the software capabilities that best meet their needs for the workloads that they care about,” said Tim Wilson, vice president and general manager of data center silicon engineering at Intel.
Rack design and software
NVLink Fusion adopters can build on the Nvidia MGX rack-scale architecture, sharing footprints, power, and cooling with Nvidia Vera Rubin NVL72 and GB300 NVL72 systems. The reference compute trays feature 100% liquid cooling without fans, cables, or hoses, and technicians can service individual trays while the rest of the rack continues running. The design supports emerging 800 VDC power architectures.
Hardware and manufacturing partners including MediaTek, GUC, Annapurna Labs, and Quanta Computer have joined the ecosystem. “With Vera Rubin [NVL72], we are looking at almost 100% automation of system builds in the manufacturing line,” said Jack Luoh, head of product and solution at QCT and Quanta Computer, noting those manufacturing investments can be leveraged by custom chips using NVLink Fusion.
Nvidia supports the hardware with its Omniverse DSX AI Factory Blueprint digital twin design tool, alongside software modules including NCCL for distributed computing, Dynamo and NIXL for disaggregation, and Mission Control for cluster telemetry and debugging.
