Nvidia is allowing hyperscalers and other AI developers to use custom accelerators with its rack-scale AI infrastructure, extending NVLink technology beyond Nvidia’s own processors. As a result of an enhanced collaboration with MediaTek, individuals are able to create personalized AI accelerators (XPUs) on a pre-approved NVLink Fusion platform and incorporate them into Nvidia-linked AI factory infrastructure designed around the Modular GPU Accelerated (MGX) ecosystem.
Nvidia has also invested $3.5 billion in the form of convertible bonds issued by MediaTek. According to Dion Harris, senior director of HPC and AI hyperscale infrastructure solutions at Nvidia, this offers customers greater autonomy to develop new concepts without the need to reconstruct their entire AI factory with bespoke silicon.
The focus lies on Customer Chips and AI Factories. The revelation occurs amidst hyperscalers creating their own processors specifically designed for AI tasks. Amazon has Trainium and Inferentia, Google possesses TPUs, and Microsoft has created Maia accelerators.
Nvidia aims to create a custom silicon that works seamlessly with its rack-scale stack. Related: Data Center Hardware Highlights – August 2026. Nvidia states that NVLink Fusion offers a pre-approved basis for integrating customized XPUs into Nvidia-associated AI facilities.
The platform integrates NVLink Fusion connectivity along with NVLink-C2C, high-speed memory (HBM), and chiplet technologies. Clients have the opportunity to incorporate Nvidia’s intellectual property into their systems’ components while still maintaining authority over their unique accelerator designs. “Customers can concentrate on distinct computing,” Harris commented.
MGX, Nvidia’s modular server and rack reference design, serves as the foundation for the infrastructure encompassing these processors at a rack level.
