Nvidia Bets $3.5 Billion on MediaTek for AI Data Center Chips
The $3.5 billion deal helps hyperscalers integrate custom AI chips into Nvidia rack infrastructure.
Nvidia has invested $3.5 billion in Taiwanese chipmaker MediaTek to enable custom AI accelerators to integrate directly into Nvidia data center rack architecture. As cloud providers and AI developers build proprietary ASICs to curb rising compute costs, the deal ensures custom silicon coexists with Nvidia infrastructure. The partnership highlights how hardware manufacturers are adapting to hyperscaler demands for tailor-made AI compute solutions.
Key details
Under the terms of the deal, MediaTek will adopt Nvidia technology to design custom application-specific integrated circuits (ASICs) for cloud operators and AI companies. MediaTek expects its custom data center ASIC business to generate $2 billion in revenue in 2026, targeting a growing portion of hyperscaler silicon spending. Major cloud and AI providers—including Amazon, Google, Microsoft, OpenAI, and Anthropic—have increasingly invested in custom chip designs to reduce reliance on off-the-shelf GPUs and lower infrastructure operating expenses. By opening up its rack-scale interconnect technology, Nvidia allows client-customized accelerators to operate alongside Nvidia GPUs using standardized data center scaffolding.
Why this matters
Hyperscale AI training and inference costs have soared as model parameters and compute workloads expand across global data centers. Cloud providers are actively seeking custom ASIC designs to achieve higher performance per watt and cut the total cost of ownership for AI workloads. Nvidia's $3.5 billion backing of MediaTek demonstrates that even dominant chip suppliers must accommodate custom hardware to maintain control over data center rack architecture and interconnect standards.
Context
This deal follows a broader shift across the AI hardware ecosystem as tech giants seek alternatives to standard GPU deployments. Last week, AWS announced the deployment of an additional 2 million Nvidia GPUs alongside custom Trainium accelerators within its infrastructure. Meanwhile, hyperscaler capital expenditures are projected to reach $725 billion in 2026, driven by intense demand for high-density compute, power delivery, and cooling infrastructure.
What happens next
MediaTek plans to leverage Nvidia's technology stack to scale its custom chip design pipeline for cloud customers throughout 2026 and beyond. Industry observers will watch whether other custom ASIC designers adopt similar rack-scale integration standards or build independent server architectures. As custom AI chips enter volume production, data center operators will monitor whether tailor-made silicon yields measurable energy efficiency gains and compute cost reductions.
Source: TechCrunch Published on AI Usage Global, author: AUG Bot
