ByteDance is in talks with MediaTek to design custom AI chips, according to a report by Korea's Digital Today. The negotiations come as the Chinese company faces tightening access to Nvidia accelerators.

The talks signal a shift in how China's largest internet firms source AI compute. Custom silicon, designed with a Taiwan-based partner, offers a path around export controls that now limit Nvidia's most advanced data center GPUs.

Why is ByteDance seeking its own chips?

ByteDance operates some of China's largest AI workloads. Its recommendation systems and large language model efforts require substantial accelerator capacity. Nvidia hardware has been the default supplier for that demand.

United States export rules have narrowed what Nvidia can legally sell into China. Each successive rule round has cut the performance ceiling for permitted parts. Reports of further tightening have pushed Chinese buyers toward alternatives.

A custom chip designed with MediaTek gives ByteDance direct control over its accelerator roadmap. It also reduces exposure to future policy shifts in Washington.

What role would MediaTek play?

MediaTek has built a business designing chips for others. The Taiwanese company supplies smartphone SoCs at scale and has expanded into custom silicon programs for large cloud customers.

Under such arrangements, the customer defines the architecture and the design partner delivers a manufacturable product. Production would typically run at TSMC, where MediaTek is a major customer. Neither company has confirmed the scope or timing of the proposed project.

The arrangement would resemble other custom-accelerator programs in the industry. Google, Amazon and Microsoft all design their own AI chips and buy Nvidia GPUs in parallel. ByteDance appears to be moving toward the same two-track model.

How does this fit the broader supply picture?

Chinese technology firms have pursued domestic AI silicon for several years. Huawei's Ascend line is the most prominent local alternative. Baidu developed its Kunlun chips. Alibaba designs its own accelerators through TGI Cloud affiliate work.

Custom design through MediaTek differs from those efforts. The chips would likely be fabricated in Taiwan rather than mainland China. That preserves access to leading-edge foundry capacity but keeps the design outside United States export licensing.

The talks also illustrate pressure on Nvidia's China business. The company has already created modified products to comply with export rules. Continued tightening raises the risk that even those parts face restriction.

What remains unconfirmed?

Digital Today reported the talks but cited no figures on volume, process node or delivery timing. Neither ByteDance nor MediaTek has commented publicly. No agreement has been announced.

Custom chip programs typically take two to three years from design start to volume shipment. Any MediaTek-designed accelerator for ByteDance would not arrive quickly.

If the talks produce a deal, expect ByteDance to keep buying Nvidia hardware where permitted while it builds an alternative. The competitive question is whether custom silicon plus TSMC manufacturing can match Nvidia's software ecosystem and annual product cadence.