ByteDance Pursues Custom AI Chips to Cut US Reliance

ByteDance, the parent company of TikTok, is reportedly developing its own custom AI CPUs to reduce reliance on US chipmakers [2]. This internal effort coincides with a reported agreement with Qualcomm to manufacture custom Application-Specific Integrated Circuits (ASICs) for ByteDance's data centers [1].

Key takeaways

• ByteDance is designing custom AI CPUs inspired by Groq's inference chips to reduce dependence on US suppliers [2].
• Qualcomm reportedly secured a multi-million unit deal to build custom ASICs for ByteDance's AI agents and Doubao chatbot [1].
• The custom chips are optimized for inference rather than training, allowing them to potentially navigate strict US export controls [1].
• ByteDance lacks internal chip design teams and is partnering with external firms like InnoStar Semiconductor for memory technology [2].

Qualcomm's ASIC Partnership and ByteDance's In-House Push

Qualcomm has reportedly landed a major deal to supply custom artificial intelligence chips for ByteDance's data centers, representing a shift for the chipmaker beyond its traditional mobile and PC markets [1]. The agreement involves Application-Specific Integrated Circuits (ASICs) rather than general-purpose processors, which are hard-wired to execute specific tasks with maximum efficiency [1]. These chips will be dedicated to running ByteDance's AI agent software and its Doubao chatbot, which a report claims recently became China's most-downloaded AI application [1]. The partnership allows ByteDance to turn an already completed, in-house chip architecture into a production-ready semiconductor built at scale [1].

Separately, ByteDance is pursuing its own custom AI CPU project, which is reportedly in the concept and design stage [2]. These chips are said to be inspired by Groq's "language processing units" and are optimized for inference tasks—running AI models instead of training them—with the company evaluating both Arm and RISC-V designs [2]. This initiative complements ByteDance's existing SeedChip AI accelerator project, which started with TSMC in 2024 and is expected to tape out and be mass-produced this year [2].

Navigating Geopolitics and Supply Chain Constraints

The drive for custom silicon is occurring against a backdrop of geopolitical tension and trade restrictions. Washington has implemented strict computing thresholds on advanced processors exported to Chinese firms to slow the development of frontier AI models [1]. However, because the custom ASICs are optimized specifically for model inference, they can be designed to sit precisely under current legal performance caps, allowing Qualcomm to sell the chips without running afoul of US restrictions [1]. This follows the Chinese government banning the purchase of Nvidia H200 Blackwell chips after changes in US export controls [2].