Nvidia Puts $3.5B Into MediaTek to Keep Custom AI Chips in Its Data-Center Stack
MediaTek’s planned accelerators are meant to run beside Nvidia hardware, giving Nvidia’s interconnect and rack architecture a role as customers pursue their own silicon.
Listen to this story
The audio brief
Story brief
3 key pointsNvidia is using a $3.5 billion convertible-bond investment to tie MediaTek’s custom-ASIC ambitions to Nvidia’s rack-scale infrastructure. Through NVLink Fusion, MediaTek can develop customer-specific accelerators that run beside Nvidia GPUs rather than displacing them—a strategy aimed at keeping cloud and AI customers inside Nvidia’s interconnect ecosystem as they build in-house silicon. MediaTek forecasts $2...
- 01
MediaTek competes with Broadcom and Marvell on custom data-center silicon; NVLink Fusion adds Nvidia-compatible connectivity, memory, packaging, and rack integration.
- 02
The $2 billion 2026 figure is MediaTek’s forecast, not current revenue or booked business.
- 03
Nvidia says customers can standardize on its rack infrastructure while placing MediaTek accelerators alongside Nvidia hardware.
Nvidia is putting $3.5 billion into MediaTek through convertible bonds, but the deal’s central promise is technical: MediaTek will use NVLink Fusion to build custom AI accelerators for Nvidia-based data-center systems. Those chips are intended to operate alongside Nvidia hardware, not replace it.
A common rack for different chips
NVLink Fusion is Nvidia’s platform for connecting custom accelerators into rack-scale data-center systems. The companies describe it as combining NVLink connectivity, high-bandwidth memory, advanced packaging and rack-scale integration, giving chip designers a base for the surrounding infrastructure while they focus on compute tailored to a customer’s workload.
That design addresses a growing tension: major cloud providers and AI companies are investing in their own chips to reduce reliance on Nvidia GPUs. Nvidia senior director Dion Harris said MediaTek customers could standardize on Nvidia’s rack-scale infrastructure while deploying custom chips beside Nvidia hardware.
Nvidia announced the investment through convertible bonds.
MediaTek’s data-center target
MediaTek is building a custom data-center ASIC business. An ASIC is a chip designed for a particular task or customer. The company said in June that it expects that business to generate $2 billion in 2026; that is a forecast, not reported revenue.
MediaTek competes with Broadcom and Marvell in helping data-center operators design custom silicon. NVLink Fusion gives it access to Nvidia’s interconnect ecosystem as it pursues that business.
The partnership extends beyond the data center
The expanded agreement also covers consumer and enterprise PCs and automotive platforms. The companies will continue work on RTX Spark and DGX Spark, as well as AI-powered vehicles built around MediaTek’s Dimensity Auto line. MediaTek’s automotive platforms use Nvidia RTX graphics for vehicle cockpits and work with Nvidia Drive AGX for autonomous-driving workloads.
MediaTek has not publicly detailed the customer base for its custom AI chips. The next substantive measure of the partnership will be whether cloud providers or AI labs deploy MediaTek accelerators using NVLink Fusion in operating data centers.
Editorial analysis
Our Read
Our read: Nvidia is not presenting custom accelerators as a market it can prevent. The MediaTek agreement instead seeks to make Nvidia’s connection technology and rack design the shared layer around them. That makes MediaTek’s customer traction the important test, not simply its 2026 ASIC revenue forecast. The strategy resembles AWS’s planned NVLink Fusion integration with Trainium: a cloud operator can develop its own chips while retaining a common infrastructure layer. The unresolved question is whether customers adopt that common layer at production scale.
Sources
- qz.comNvidia investing $3.5 billion in MediaTek for AI chip partnership
- techcrunch.comNvidia’s $3.5B MediaTek bet reveals its plan for tackling Big Tech's AI chip buildout | TechCrunch
