NVIDIA is putting $3.5 billion behind a deeper MediaTek relationship that reaches from custom data-center accelerators to AI PCs and software-defined vehicles. The companies announced the expanded collaboration on August 31, with MediaTek adopting NVIDIA’s NVLink Fusion platform as a foundation for custom XPUs.
The joint announcement published by MediaTek says NVIDIA invested in convertible bonds issued by MediaTek. It also describes a plan to connect customer-designed accelerators into NVIDIA’s NVLink-connected, rack-scale AI factories, rather than asking every customer to build the surrounding interconnect and packaging stack alone.
That distinction matters because the custom-chip market is moving beyond the question of who can design an accelerator. A useful way to read the announcement is as a bet on the infrastructure around the accelerator: high-speed links, memory, packaging, manufacturing and the software needed to operate a complete system. Neon Control has previously explained why memory bandwidth has become a central AI hardware constraint, and its NPU, GPU and CPU explainer shows why different compute engines increasingly share one system. NVLink Fusion is aimed at that same system-level pressure from a different angle.
What MediaTek is adopting
MediaTek will offer NVLink Fusion to customers developing custom AI accelerators. NVIDIA describes the platform as a prebuilt, prequalified and system-prevalidated foundation for multi-die XPU development. The company says it combines an NVLink Fusion chiplet, NVLink-C2C and NVHBM. The chiplet connects an XPU to the NVLink scale-up fabric through photonics or electrical interconnects. NVLink-C2C is the high-bandwidth link between XPUs, NVIDIA Rosa CPUs and compatible processors, while NVHBM is intended to integrate customized memory capabilities.
Those components do not turn a customer design into a finished product by themselves. The companies say customers can tailor connectivity, memory, packaging, performance and power characteristics to their workload while MediaTek and NVIDIA handle more of the engineering around a production-ready system. The pitch is less about replacing NVIDIA GPUs than about making other compute blocks easier to place inside an NVIDIA-centered rack.
NVIDIA’s announcement also says the wider NVLink Fusion ecosystem is intended for hyperscalers, cloud providers and frontier model developers. The company presents its MGX rack-scale architecture as another part of that path, with the stated goal of reducing development complexity and shortening time to market for semi-custom AI factories. Those are vendor claims about the platform’s benefits, not an independent performance benchmark.
The same partnership reaches the edge
The deal is not limited to data centers. NVIDIA and MediaTek say they will continue work on multiple generations of RTX Spark and DGX Spark PC chips, combining NVIDIA GPUs with MediaTek system-on-chip designs for consumer PCs, developer systems and enterprise workstations. MediaTek previously collaborated on the GB10 Grace Blackwell Superchip used in DGX Spark, according to the announcement.
The companies also describe ongoing work on AI-defined vehicles. MediaTek’s Dimensity Auto platforms integrate NVIDIA technologies for AI and RTX graphics, and can work alongside NVIDIA DRIVE AGX. The statement gives no launch schedule, product specifications or customer list for the expanded work, so the immediate news is the financing and the platform relationship rather than a new chip available for purchase.
For builders, the signal is clear but narrower than a new GPU launch. NVIDIA is trying to make its rack-scale fabric the place where custom accelerators, CPUs, memory and future edge systems meet. MediaTek brings custom silicon, SoC, connectivity and packaging expertise to that effort. The practical test will be which customer XPUs reach production, and when, not simply how many partners join the ecosystem.
