Nvidia is investing $3.5 billion in Taiwanese chipmaker MediaTek, a deal designed to give custom AI processors a place inside Nvidia’s data-center architecture rather than leaving them entirely outside its ecosystem.
MediaTek will use Nvidia technology to design custom chips for AI companies and hyperscale cloud providers. Those processors will be able to connect with Nvidia systems through the company’s high-speed interconnect and rack infrastructure, giving customers more specialized compute without requiring them to abandon the surrounding platform.
The investment arrives as Amazon, Google, Microsoft, OpenAI and Anthropic develop their own chips to reduce dependence on standard Nvidia GPUs. Nvidia’s response is not to resist every custom processor. It is to make the connections, networking and rack design around those processors part of its business.
Custom silicon enters Nvidia’s architecture
MediaTek will gain access to NVLink Fusion and related scale-up and scale-out technology. NVLink is designed to let processors communicate at high speed, a requirement for large AI workloads spread across many chips. The arrangement means non-Nvidia silicon can operate alongside Nvidia equipment inside a common rack-scale system.
MediaTek can tailor processors to a customer’s workload while using Nvidia’s rack architecture and communications layer. The companies have not disclosed which customers may buy these custom AI chips, so the immediate commercial pipeline remains unknown.
MediaTek said in June that it expected its custom data-center ASIC business to generate $2 billion in revenue during 2026. Nvidia’s capital and infrastructure access give that operation a larger route into a market where cloud providers increasingly want designs optimized for their own models and services.
The companies will also continue their existing DGX Spark collaboration and work on RTX Spark consumer AI PCs. Their relationship extends to software-defined vehicles, including systems using Nvidia RTX cockpit technology and Drive AGX for autonomous-driving workloads.
Nvidia protects the layers around compute
Nvidia executive Dion Harris said the company moved beyond defining itself as a pure compute provider years ago and now sees itself as an AI infrastructure company. The MediaTek deal illustrates that broader claim: even when a customer chooses a custom chip, Nvidia can supply the interconnects, networking and system design that make a large deployment function.
That strategy also appeared in Nvidia’s agreement with Amazon Web Services last week. AWS committed to deploy an additional two million Nvidia GPUs and adopted NVLink Fusion, although Nvidia did not make a direct investment in AWS.
The MediaTek transaction is more financially circular. Nvidia is funding a company that will use Nvidia technology to build processors for customers seeking alternatives to Nvidia’s core GPUs. Yet the apparent contradiction is also the logic of the deal. Nvidia can concede part of the accelerator market while retaining influence over how custom chips are assembled into data centers.
For hyperscalers, the offer is a compromise. They can pursue workload-specific economics and design control without independently rebuilding every layer of rack-scale infrastructure. For Nvidia, each custom processor connected through its technology reinforces a platform that is harder to replace as a whole.
An open question over customer independence
Nvidia CEO Jensen Huang said AI is transforming factories, personal computers and cars, and argued that the partnership will create differentiated systems at scale. That is a vendor projection rather than a disclosed customer result. No buyer, deployment schedule or performance figure for the new custom chips has been announced.
The practical test will be how much independence customers gain. A chip tailored to one workload may lower costs or improve efficiency, but a design tied tightly to Nvidia’s interconnect and rack architecture can preserve dependence at another layer. The deal changes where the lock-in may sit rather than proving that it disappears.
MediaTek brings experience in custom silicon and a stated revenue target; Nvidia contributes capital and the system around the chip. Together they are positioning for a market in which the processor is no longer always an off-the-shelf GPU.
The $3.5 billion wager shows Nvidia preparing for that transition before custom silicon becomes a larger threat. If major AI companies insist on owning more of their compute designs, Nvidia intends to remain indispensable to the infrastructure those designs use.
