Nvidia is putting $3.5 billion into MediaTek, and the deal cuts straight to the tension defining AI infrastructure right now. Hyperscalers and AI labs — Amazon, Google, Microsoft, OpenAI, Anthropic — are all pouring money into custom silicon to reduce their dependence on Nvidia GPUs. Rather than fight that trend, Nvidia is positioning itself as the connective tissue those custom chips will need anyway.
MediaTek gets access to Nvidia’s NVLink Fusion ecosystem, including the high-speed interconnect technology that lets any chip — Nvidia’s or otherwise — communicate quickly within a data center. The pitch, as Nvidia’s senior director of HPC and AI hyperscaler infrastructure solutions Dion Harris put it on a Monday press call, is that customers can “standardize on the rack-scale infrastructure across their AI factories” and deploy custom ASICs right alongside Nvidia GPUs using the same platform. “Nvidia is an AI infrastructure company,” Harris said. “We expanded beyond pure computing chips years ago.”
The financial stakes for MediaTek are concrete. The Taiwanese chipmaker said in June it expects its custom data center ASIC business to hit $2 billion in revenue by 2026. Nvidia clearly wants a piece of that trajectory — and more importantly, wants whatever MediaTek builds for customers to plug into Nvidia’s ecosystem rather than someone else’s. The deal follows last week’s announcement that AWS will deploy an additional 2 million Nvidia GPUs and integrate NVLink Fusion, suggesting a deliberate pattern: Nvidia cedes the actual silicon design to partners while tightening its grip on the infrastructure layer surrounding it.
Beyond data centers, the partnership extends to consumer and automotive markets. The companies will keep collaborating on DGX Spark, Nvidia’s desktop AI developer machine, and RTX Spark for AI PCs. They’re also building software-defined vehicle platforms combining MediaTek’s auto systems with Nvidia’s RTX graphics and Drive AGX autonomous computing. Jensen Huang framed it as building platforms that “bring NVIDIA accelerated computing to new markets.” The unspoken logic: if you can’t stop customers from designing their own chips, make sure those chips still speak Nvidia’s language. That’s the real bet here — and it’s a clever one.
