Nvidia's $3.5B MediaTek Bet Exposes Its Answer to the Custom-Chip Revolt
Instead of fighting the migration to custom silicon, Nvidia is buying into it — making sure its interconnect and rack-scale architecture stay at the center of everyone else's AI factory.
Published: 2026-09-01 Category: Quick Take Sources: TechCrunch
A Countermove Against the GPU Exodus
Nvidia is investing $3.5 billion into Taiwanese chipmaker MediaTek. In exchange, MediaTek will adopt Nvidia's technology to design custom chips for AI companies and hyperscalers that can plug directly into Nvidia-based data centers. The deal comes as Amazon, Google, Microsoft, OpenAI, and Anthropic all invest in building their own chips to reduce dependence on Nvidia's GPUs.
Nvidia's framing is unambiguous: it's no longer just a chip company. "Nvidia is an AI infrastructure company," said Dion Harris, Nvidia's senior director of HPC and AI hyperscaler infrastructure solutions. "We expanded beyond pure computing chips years ago."
The Move: License the Rails, Not Just the Train
The core of the arrangement is NVLink Fusion, Nvidia's ecosystem that lets any chips — even non-Nvidia ones — communicate with each other quickly. MediaTek gets access to that stack, including Nvidia's "proven scale-up and scale-out technology stack and ecosystem" and "rack-scale architecture." Cloud companies or AI labs can keep designing their own tailored chips via MediaTek, but those chips will live inside Nvidia-standard racks, speaking Nvidia-standard interconnect.
Harris framed it as opening up the platform: "Basically, every cloud, every model builder is deploying our platform in some shape, form, or fashion... by MediaTek being able to offer this extension to its customers, it allows them to standardize on the rack-scale infrastructure across their AI factories." In other words, Nvidia cedes the silicon itself but retains the plumbing — and the control. Last week Nvidia also announced a similar partnership with AWS (without a direct investment), which will see AWS deploy an additional 2 million Nvidia GPUs and integrate NVLink Fusion.
Why This Is a Savvy Retreat
MediaTek has been building its custom data-center ASIC operations and expects that business to generate $2 billion in revenue in 2026. Nvidia is effectively hitching its interconnect standard to that growth. This is the classic "rails over trains" strategy: let rivals and customers build their own compute, as long as they all route through infrastructure Nvidia owns.
The strategic reading is clear. Nvidia cannot stop the custom-chip trend — the hyperscalers are too large and too motivated. What it can do is ensure that even the chips that replace Nvidia GPUs still live on Nvidia-designed scaffolding, keeping Nvidia's rack-scale platform as the default substrate of every AI factory. It's a graceful concession that converts a competitive threat into a toll road. The bet: being the standard layer of AI infrastructure is worth more in the long run than being the only supplier of GPUs.
Based on reporting by Rebecca Bellan at TechCrunch.