NVIDIA has agreed to buy $3.5 billion in convertible bonds issued by MediaTek, officially expanding their alliance from client hardware into artificial intelligence data centers. The deal establishes a 10 year product roadmap to integrate custom MediaTek processors directly into NVIDIA infrastructure through NVLink Fusion. While the partnership aims to speed up custom chip rollouts, it also ties both chipmakers closely together in the race to build global AI factories.
Speaking to Bloomberg, NVIDIA chief executive Jensen Huang described MediaTek as one of his largest industry partners. The core of this expanded alliance centers on NVLink Fusion. This framework allows MediaTek custom processors to plug directly into NVIDIA data center hardware. Customers can deploy custom silicon alongside NVIDIA Spectrum X Ethernet, switch systems, and Vera central processors without building an entire server platform from scratch. When MediaTek wins a custom silicon contract, NVIDIA automatically sells the surrounding network hardware and host processors.
The collaboration also covers consumer hardware. Earlier work connected MediaTek mobile silicon with NVIDIA graphics chips through high speed chip to chip connections, producing the DGX Spark desktop supercomputer. Both companies plan to use this foundation to build multiple generations of processors for Windows PCs. These chips are designed to power the next wave of AI agent software across desktop and laptop computers.
Huang firmly pushed back against suggestions that the $3.5 billion bond purchase represents circular financing. He emphasized that MediaTek is already a profitable business with strong cash flow. The investment is intended to fund long term research rather than prop up sales figures.
This is not circular financing. It is a strategic investment built on long term technical cooperation.
Huang noted that he has worked with MediaTek chief executive Rick Tsai for 25 years, pointing to personal trust as a major factor behind the deal. MediaTek plans to spend the capital on advanced chip packaging, high speed interconnect IP, and global supply chain stability. Tsai pointed out that operating in Taiwan gives MediaTek direct access to the manufacturing base required for large scale AI hardware.
Huang also shared financial metrics explaining how NVIDIA captures value in modern data centers. Every 1GW of power capacity creates rising equipment revenue across product cycles. The Hopper generation generated about $18 billion per gigawatt. Grace Blackwell pushed that figure to $25 billion. The upcoming Vera Rubin generation will exceed $40 billion per gigawatt. This growth happens because a modern AI factory requires 7 distinct types of silicon, ranging from switches to network cards. Even if a cloud provider chooses a custom MediaTek accelerator, NVIDIA still provides the rest of the server ecosystem.
To support this architecture, NVIDIA is opening its NVHBM memory platform to MediaTek. The technology uses a custom NVIDIA base die combined with high bandwidth memory on advanced CoWoS packaging. Extending NVHBM access to partners like MediaTek and Amazon helps standardize custom chip assembly while easing pressure on global component suppliers.
