Nvidia puts $3.5B into MediaTek chips

Nvidia said Monday it is buying $3.5 billion of convertible bonds in MediaTek, the Taiwanese chipmaker best known for smartphone processors. The bonds convert into MediaTek shares, and the deal folds MediaTek into NVLink Fusion, Nvidia’s platform for connecting custom accelerator chips into its rack-scale data-center systems.

The timing explains the move. Hyperscalers and frontier labs are increasingly designing their own AI chips to cut their dependence on Nvidia GPUs. Rather than fight that trend, Nvidia is selling the plumbing: any custom silicon MediaTek helps a customer build still needs to talk to the rest of a Nvidia-based rack, and NVLink Fusion is how it does that. “Nvidia is an AI infrastructure company. We expanded beyond pure computing chips years ago,” Nvidia’s Dion Harris, senior director of HPC and AI hyperscaler infrastructure solutions, told TechCrunch.

The agreement also extends the companies’ existing work on Nvidia’s RTX Spark and DGX Spark chips for consumer and enterprise PCs, plus AI-focused automotive platforms built around MediaTek’s Dimensity Auto line.

Why this matters beyond one deal

The custom-silicon wave, Google’s TPUs, Amazon’s Trainium, Meta’s and Microsoft’s in-house accelerators, is the biggest structural threat to Nvidia’s margins. This deal is Nvidia making sure that even chips it didn’t design still run on rails it controls. For anyone architecting around a specific accelerator today, from Nvidia’s own Vera Rubin generation to the wave of hardware-focused capital chasing the same buildout, like a16z’s new Machine Age Fund, the lesson is the same: the interconnect layer, not just the chip, is where the next round of lock-in gets built.