NVIDIA and MediaTek said on August 31, 2026 that they are widening a long-running chip partnership across cloud AI factories, local PCs, and cars. The commercial hook is a $3.5 billion NVIDIA investment in convertible bonds issued by MediaTek. The technical hook is MediaTek’s adoption of NVIDIA’s NVLink Fusion platform, which the companies say gives hyperscalers and model labs a prevalidated path to design custom XPUs that still sit inside NVIDIA’s rack-scale fabric.

Santa Clara-based NVIDIA and Hsinchu-based MediaTek described three workstreams. In the data center, MediaTek will help customers build custom accelerators that connect to NVIDIA systems through NVLink. On the desktop, the companies will keep shipping generations of NVIDIA RTX Spark and DGX Spark chips that pair NVIDIA GPUs with MediaTek system-on-chip designs. In vehicles, they will keep combining MediaTek Dimensity Auto platforms with NVIDIA graphics and DRIVE AGX for software-defined cockpits.

“AI is transforming every computing platform, from the world’s largest AI factories to the PC and the car.” Jensen Huang, founder and CEO, NVIDIA

“NVIDIA’s investment strengthens a collaboration that spans cloud AI infrastructure, local AI computing and automotive in the era of physical AI.” Rick Tsai, vice chairman and CEO, MediaTek

NVIDIA said NVLink Fusion is a prebuilt foundation for multi-die XPU development, covering the path from silicon and packaging to a full rack. The stack named in the release includes an NVLink Fusion chiplet that ties XPUs into the NVLink scale-up fabric over photonics or electrical links, NVLink-C2C for connections between XPUs and NVIDIA Rosa CPUs, and NVHBM for customized memory. Partners can also use the NVIDIA MGX rack architecture.

MediaTek’s role is to take a customer’s XPU design and tune connectivity, memory, packaging, performance, and power so the part can ship into an NVIDIA-connected factory instead of a one-off cluster. The companies already worked together on the GB10 Grace Blackwell Superchip inside DGX Spark. They now say that collaboration extends to RTX Spark for consumer PCs.

The MediaTek note attached to the release told investors not to treat the announcement as a revenue or earnings forecast. The work, it said, still depends on customer tape-outs, capacity, and demand.

Decoded Take

NVIDIA is paying to keep the next wave of custom accelerators inside its own rack, not as a rival fabric. A $3.5 billion convertible bond is a soft lock-in: MediaTek gets a balance-sheet vote of confidence, and NVIDIA gets a Taiwanese design house selling NVLink as the default way to attach someone else’s XPU. The risk is that hyperscalers treat Fusion as a brochure while they finish in-house interconnects. The signal to watch is a named cloud or model lab that actually tapes out a MediaTek-packaged XPU onto NVLink, not another Spark PC refresh.