NVIDIA launches NVLink Fusion for custom AI factories

NVIDIA has introduced NVLink Fusion for companies building custom XPUs, targeting hyperscalers and AI-native groups creating semi-custom AI factory infrastructure.

The launch extends NVIDIA's interconnect and rack-level architecture into systems that use processors designed outside its GPU line. NVLink Fusion links custom XPUs into NVIDIA's NVLink scale-up network and includes NVLink-C2C for connecting those processors to NVIDIA Vera CPUs or other ecosystem CPUs.

NVIDIA is pitching the product to operators that want to build large AI installations without designing every layer of the platform from scratch. It argues that many groups developing custom silicon run into delays and cost overruns when they move beyond chip design into network integration, rack engineering, cooling, power, software and supplier management.

Under the new offering, adopters can use NVIDIA's MGX rack-scale architecture and the same supply chain used for MGX-based systems, including Vera Rubin NVL72. Manufacturing partners would handle design and integration, while suppliers in the MGX ecosystem would provide components for racks, cooling and power systems.

Network design

A central part of the announcement is NVIDIA's effort to make its scale-up networking available to third-party accelerators. NVIDIA said sixth-generation NVLink can connect up to 72 XPUs in a single domain, with end-to-end latency for XPU-to-XPU transfers three times lower than alternatives based on standard Ethernet and packet rates 10 times higher.

NVIDIA also outlined a broader roadmap for larger configurations, saying future NVLink designs would support domains of up to 1,152 accelerators and co-packaged optics. It added that NVLink-C2C can deliver up to six times the energy efficiency of a PCIe interface when connecting XPUs to CPUs.

The move reflects a wider shift in the AI data centre market, where operators are increasingly mixing processor types for training, inference, reasoning, retrieval and serving. That trend has created demand for shared rack layouts, common cooling and power systems, and software that can manage mixed clusters without forcing buyers to commit early to a single chip architecture.

Ecosystem pitch

NVIDIA said the ecosystem around NVLink Fusion spans ASIC design, CPU, IP and optical interconnect partners. It cited several companies as backers of the approach, framing the product as a way to preserve flexibility for customers developing their own accelerators.

"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 Centre Silicon Engineering at Intel.

That argument is particularly relevant for cloud companies and large AI operators that begin building facilities before final silicon decisions are made. NVIDIA said power procurement, facility planning, cooling, rack layout and network architecture often start well before the accelerator mix is settled, creating risk if a site is tied too closely to one chip plan.

"The value of the NVLink Fusion program is ... [customers] can deploy their rack-level solution with the NVIDIA GPU, and then they can decouple the development of their XPU and put it at a different pace," said Vince Hu, Corporate Senior Vice President and General Manager of the Data Centre and Computing Business Group at MediaTek.

Others focused on the opportunity to build systems around a common rack blueprint while using different processor combinations. "NVLink Fusion allows the hyperscalers or the custom ASIC designers to integrate their own custom CPU or XPU and bridges the NVIDIA technology with a third-party process to create a unified rack-scale architecture," said Lie-Szu Juang, Chair and Chief Strategy Officer at GUC.

Factory model

NVIDIA also tied the product to its broader AI factory strategy. It said NVLink Fusion aligns with its DSX reference architecture, which brings together building design, power, cooling, compute and networking. NVIDIA also pointed to its Omniverse DSX AI Factory Blueprint as a way for partners to model facilities and technology before deployment through a digital twin.

At rack level, NVIDIA said the reference compute trays use liquid cooling throughout and are designed so trays can be removed while the rest of the rack continues operating. NVLink Switch trays are also liquid cooled and support continued operation during service.

Manufacturing and deployment were another theme of the launch. "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. "Most of those investments can be leveraged if the XPU leverages NVLink Fusion."

Software support is intended to tie the hardware together. NVIDIA said NCCL would support distributed workloads, while Dynamo and NIXL would handle disaggregation, and Mission Control would provide cluster management, telemetry and debugging across mixed AI infrastructure.

Amazon's Annapurna Labs also pointed to the time and supply-chain advantages of the design. "With NVLink Fusion we can use proven NVL72 rack design to have time-to-market, and we can have access to multiple suppliers to help us to deliver more into the hands of our customers," said CC Lee, Senior Hardware Development Manager at Annapurna Labs, an Amazon company.