← Back to feed

Nvidia announced on August 31 that it is investing $3.5 billion into MediaTek, the Taiwanese chipmaker that builds custom silicon for cloud providers and AI labs alongside its consumer electronics chip business. The investment is paired with a technology agreement: MediaTek gets access to NVLink Fusion, letting custom ASIC designs communicate directly with Nvidia’s rack-scale infrastructure at full NVLink bandwidth.

The same NVLink Fusion access went to AWS the week prior — without the equity. AWS will deploy an additional 2 million Nvidia GPUs as part of its arrangement, integrating NVLink into its own custom silicon designs.

What NVLink Fusion Actually Does

NVLink is the high-bandwidth interconnect that glues Nvidia’s GPUs together inside server clusters. NVLink Fusion extends that interconnect beyond Nvidia’s own chips to any silicon a partner wants to design. An AI lab building a custom training accelerator can now run that chip in the same rack as Nvidia GPUs, sharing memory bandwidth and communicating at NVLink speeds rather than falling back to PCIe or Ethernet.

The practical effect is that choosing custom silicon no longer means leaving Nvidia’s ecosystem. Hyperscalers like AWS, Google and Microsoft have been investing heavily in their own chips — Trainium, TPU, Maia — partly to reduce dependency on Nvidia. NVLink Fusion makes that independence partial: you can build custom silicon, but your data center still runs on Nvidia’s rack architecture, with Nvidia’s interconnects and Nvidia’s software stack.

MediaTek’s Custom ASIC Business

MediaTek told investors in June that it expects $2 billion in revenue from its custom data center ASIC operations in 2026, growing the business it has been building for hyperscaler customers. With NVLink Fusion access and Nvidia’s validation, MediaTek can pitch that custom silicon as first-class within Nvidia-based AI factories rather than as an alternative to them.

MediaTek also continues its existing Nvidia partnership on DGX Spark — the desktop developer AI computer — and on RTX Spark for AI PCs, plus ongoing work on Nvidia Drive AGX for software-defined vehicles. The investment extends every one of those collaborations.

Nvidia’s Investment Pattern

The $3.5B MediaTek deal continues a spending trend that Nvidia has run throughout 2026. It has invested in AI hyperscalers, cloud providers and now the chipmaker building custom silicon for its customers. Each investment flows back into the Nvidia ecosystem in some form: more GPUs deployed, more NVLink ports installed, more software licenses running.

Dion Harris, Nvidia’s senior director of hyperscaler infrastructure solutions, framed the company’s position plainly: “Nvidia is an AI infrastructure company. We expanded beyond pure computing chips years ago.” The MediaTek deal makes that claim concrete. Nvidia is not just selling GPUs; it is buying its way into every layer of the AI factory stack — interconnects, custom ASIC compatibility, and now equity in the companies building alternatives.

The risk is that customers grow uncomfortable with an infrastructure supplier that is also a significant investor in their chip suppliers. That tension has not stopped any of these deals from closing.