TikTok owner ByteDance in advanced talks to buy custom AI chips - The American Bazaar

ByteDance's push into custom AI silicon is a structural decoupling from NVIDIA, not a routine procurement. For three years, over 60% of China's hyperscaler AI capex flowed to NVIDIA data-center GPUs. Once ByteDance migrates inference workloads to purpose-built ASICs, NVIDIA's revenue moat in the Chinese market develops a structural crack. On the supply chain, custom silicon locks ByteDance into EDA toolchains, IP licensing, and advanced packaging capacity. Relying on US-based Synopsys/Cadence for design and TSMC (Taiwan, China) for CoWoS packaging creates dual exposure. The more probable play: ByteDance adopts a "design autonomy + foundry diversification" strategy, routing partial tape-outs to SMIC's N+2 node to hedge against escalating export controls. Competitively, Huawei's Ascend and Cambricon face direct margin pressure. A custom chip cutting inference costs by 30-40% forces domestic cloud providers to reprice compute procurement. NVIDIA's counter-move likely accelerates H200 "compliance-tier" shipments to China via price war. Within 18 months, Alibaba and Tencent will follow suit. China's AI chip market shifts from "buying cards" to "building cards." The true bottleneck is no longer the GPU die—it's advanced packaging capacity.