Nvidia is putting $3.5 billion into MediaTek, and in return MediaTek is wiring its custom AI chips into Nvidia's networking standard.
Nvidia will buy $3.5 billion in convertible bonds from MediaTek, and MediaTek will adopt Nvidia's NVLink Fusion platform for its custom AI accelerators, local AI systems, and automotive chips. NVLink Fusion lets MediaTek connect its own compute designs to Nvidia's networking fabric, CPUs, and memory technology, so a customer can design a custom accelerator and plug it into a mostly pre-built, rack-scale Nvidia system. The companies are also extending earlier work: MediaTek will keep building the Superchips behind Nvidia's DGX Spark and RTX Spark machines, and their multi-generation deal on Dimensity Auto chips for car cockpits continues alongside Nvidia Drive AGX.
The subtext is that Nvidia can't stop Amazon, Google, Meta, Microsoft, and now OpenAI from building their own AI chips to cut their Nvidia GPU bills. So instead of fighting that shift, Nvidia is positioning NVLink Fusion as the connective tissue every custom chip needs anyway, collecting a toll and a foothold even when it loses the accelerator design itself. MediaTek gets a shortcut too: smaller customers who can't build a full rack-scale AI system on their own now get one off the shelf.
It's less a rescue plan than an insurance policy - if custom silicon eats into Nvidia's GPU sales, Nvidia wants to make sure it's still selling the wiring, the memory, and the rack around whatever chip wins.