On October 15, Bloomberg reported that ByteDance is in advanced talks with MediaTek over data-center AI chips and is "close to an agreement". Lianhe Zaobao described the chips as custom-designed for ByteDance's own data centers; Phoenix Technology and Guandian added that if the deal closes, ByteDance could become the second-largest customer of MediaTek's ASIC business.
The shape of the deal, cross-checked
Deal points (all "in talks" as of publication)
· Target: AI ASICs for data centers
· Status: close to a procurement agreement (Bloomberg)
· Form: custom silicon for ByteDance data centers (Lianhe Zaobao)
· Standing: potentially MediaTek ASIC's No.2 customer (Phoenix/Guandian)
· Corroboration: relayed by EETimes China,ijiwei, TechNode and others
Why MediaTek
Within the broader trend of "inference workloads moving to ASICs", the majors have already set several templates: Google builds TPU with Broadcom's help, Microsoft develops Maia in-house, Meta co-designs with Broadcom, and Amazon pairs in-house Trainium with Marvell. ByteDance's "deep co-design plus volume procurement" balances risk and reward: no chip team built from scratch, yet an architecture tailored to its own inference workloads — plus supply certainty. MediaTek's mass-production and cost-control credentials from smartphone SoCs are the industrial foundation; as one analysis put it, with Nvidia supply tight and in-house tape-outs slow, "MediaTek became the optimal answer."
Once buyers start routing around commodity GPUs, the compute market shifts from fighting for cards to defining them.
What it signals on the demand side
ByteDance is widely seen as one of the fastest-growing buyers of inference capacity, with its Doubao app family and agent products pushing token consumption higher. If the deal lands, put it on the same coordinate axis as Microsoft's Maia 300 entering volume production and IDC's finding that non-GPU accelerator shipments have surpassed GPUs: the supply structure of inference compute is diversifying, and a second supply system beyond commodity GPUs is taking shape.
A dose of caution: talks are not contracts, and ASICs take years from architecture definition through tape-out to volume production, with advanced packaging still a bottleneck. The real lesson for enterprises is that compute procurement is shifting from "buying cards" to "defining cards" — multi-cloud, multi-chip scheduling and TCO modeling beat betting on a single supply chain. NineZenith keeps the compute layer decoupled precisely so clients keep capturing the cost dividends of supply-side competition.
(Compiled from Bloomberg reporting as relayed by EETimes China, Lianhe Zaobao, Phoenix Technology, Guandian, ijiwei and other public sources)