Nvidia invests $3.5B in MediaTek to scale custom AI chip networking

This investment signals a shift toward deeper hardware integration between GPU leaders and custom silicon designers to solve data bottlenecks in high-scale AI systems. By embedding NVLink Fusion and NVHBM memory designs into MediaTek’s architecture, Nvidia ensures its proprietary networking standards become the backbone for third-party AI accelerators and automotive chips. For the streaming ecosystem, this infrastructure evolution supports the massive compute requirements of real-time video AI and software-defined vehicle interfaces. Watch for MediaTek’s next revenue update to see if this partnership accelerates their goal of growing the data center unit severalfold beyond the current $2 billion target.

MediaTek has been building its custom AI silicon business aggressively over the past year, positioning itself as a key design partner for hyperscalers seeking alternatives to merchant GPUs. In May 2025, MediaTek announced its Dimensity Auto platform targeting software-defined vehicles with integrated AI processing, signaling the company's intent to extend its mobile chip expertise into automotive and edge AI workloads. The Nvidia investment validates that strategy by providing both capital and access to NVLink Fusion, which MediaTek can embed in custom accelerators designed for cloud providers. Marvell Technologies, a direct competitor in the custom AI silicon space, has pursued a similar model of integrating proprietary interconnects into customer-designed chips, making the Nvidia-MediaTek deal a direct competitive response in the custom accelerator market.

The business case for Nvidia's convertible bond structure reflects a broader trend of GPU vendors locking in design partnerships through equity stakes rather than simple licensing agreements. In March 2025, Nvidia reported data center revenue of $35.6 billion for its fiscal fourth quarter, up 93% year over year, underscoring the scale of demand that motivates the company to secure custom silicon partners capable of extending its networking standards into non-Nvidia accelerators. Jensen Huang has framed NVLink Fusion as an open ecosystem play, allowing third-party chip designers to connect their silicon to Nvidia GPUs at the same bandwidth as Nvidia's own products. For MediaTek, the partnership provides a path to grow its data center business beyond the current $2 billion run rate by offering hyperscalers a turnkey solution that combines MediaTek's design services with Nvidia's interconnect and memory interface technology.

On the technical side, NVLink Fusion represents Nvidia's answer to the bandwidth bottleneck that limits multi-chip AI systems. The interconnect delivers 1.8 terabytes per second of bidirectional bandwidth per link, a specification that Nvidia detailed at GTC 2025 alongside the NVL72 rack-scale system. For streaming and video AI workloads, the relevance is indirect but significant: as hyperscalers deploy custom accelerators built on MediaTek silicon with NVLink Fusion connectivity, the resulting compute density supports real-time video transcoding, generative AI inference, and recommendation engines at scale. MediaTek's existing relationships with cloud providers, combined with Nvidia's networking stack, could accelerate the deployment of heterogeneous AI clusters where custom chips handle inference while Nvidia GPUs manage training, a configuration that for cost efficiency.