ByteDance is close to a deal with MediaTek to secure AI chips for its data centres, according to a report by Asia Business Outlook. The arrangement, if completed, would give the TikTok parent a dedicated channel of AI accelerator silicon outside the tight supply of Nvidia-class GPUs that has defined the market since 2023.

The reported deal sits at the intersection of two pressures. ByteDance runs one of the largest recommendation and video workloads on the internet, and that workload now consumes inference capacity at data-centre scale. At the same time, United States export controls restrict which advanced accelerators Chinese customers can buy from American vendors, pushing Chinese hyperscalers toward alternative suppliers and custom silicon routes.

What would the deal cover?

According to Asia Business Outlook, the negotiation concerns AI chips for data centres. MediaTek, best known as a mobile SoC supplier, has spent several years expanding beyond handsets into ASIC design services and edge AI silicon. A data-centre AI chip engagement with ByteDance would extend that trajectory into the highest-value segment of the market.

The report does not specify the chip family, the process node, the volume, or the financial terms. It describes the deal as near completion rather than signed, so the commercial details remain open until the parties confirm them.

The mechanism matters more than the label. A fabless customer like ByteDance cannot build chips alone; it needs a design partner, foundry capacity, and packaging supply. MediaTek would supply the design-and-integration role, in a structure similar to how other cloud operators have contracted custom accelerators rather than buying merchant GPUs.

Why would ByteDance do this now?

Two forces shape the timing. First, demand for AI inference capacity at ByteDance continues to grow with its video and advertising products. Second, access to top-tier US accelerators remains constrained by export policy for Chinese buyers, which makes a non-US supply route strategically attractive.

A custom or semi-custom chip also changes the cost picture. Merchant AI accelerators carry premium pricing; purpose-built silicon, produced at scale, can lower the cost per unit of inference even if raw performance trails the leading GPUs. For a company whose workloads are recommendation and video rather than frontier model training, that trade can make commercial sense.

For MediaTek, a data-centre AI win would mark a step change. The company has told investors for several quarters that it expects AI-related revenue to grow across both edge and data-centre segments. Landing ByteDance as a customer would give that roadmap a named anchor.

What remains uncertain?

Neither ByteDance nor MediaTek has confirmed the deal publicly, and the report does not state a signing date, order value, or shipment schedule. Deals of this type typically require foundry allocation and multi-quarter design cycles before first silicon reaches a data centre, so any volume supply would follow well after a contract signature.

The competitive read is straightforward. If the deal closes, ByteDance joins the ranks of hyperscalers that build or commission their own AI silicon, and MediaTek gains a data-centre reference customer that could attract further custom-chip business from other operators facing the same supply constraints.