Nvidia is turning its NVLink interconnect technology into a licensing business, allowing rival chipmakers to build custom AI processors that still plug into Nvidia's ecosystem. On Monday, MediaTek became the latest company to adopt NVLink Fusion for its datacentre XPU chips, joining Amazon, Fujitsu, Qualcomm, Arm and Marvell, in a move that shows Nvidia remains central to the AI hardware market even as customers build their own chips to reduce reliance on its GPUs.

As part of the MediaTek deal, Nvidia is investing $3.5 billion in convertible bonds issued by the Taiwanese chip designer, while continuing to license MediaTek's own designs for future DGX and RTX Spark systems. NVLink Fusion, launched last year, lets rival XPUs tap into Nvidia's high-speed chip-to-chip and rack-scale networking, and because Nvidia's MGX rack designs are part of the Open Compute Project, partners can slot their own compute blades into existing NVL72 rack designs. This cuts engineering overheads for competitors while still generating revenue for Nvidia through NVSwitch sales and licensing fees, with Amazon's upcoming Trainium4 accelerators, for example, set to replace their in-house interconnect with NVLink Fusion.