ByteDance is reportedly developing custom AI CPUs to reduce reliance on US chips

ByteDance is reportedly working on custom AI CPUs as it tries to reduce its dependence on US chipmakers. The TikTok owner is said to be designing chips focused on inference, which means running AI models rather than training them from scratch.

The project is still believed to be in the early concept and design stage. According to the report, ByteDance is evaluating both Arm and RISC V designs for the chip. The goal appears to be a processor built for the growing demand around agentic AI, where systems run AI models repeatedly to answer questions, make decisions, and complete tasks.

This would be different from a traditional general purpose server CPU. ByteDance’s reported chip is said to be inspired by Groq style language processing units, which are designed to handle AI inference workloads efficiently. That could make sense for ByteDance because the company runs large AI services in China, including the Doubao chatbot app, along with several AI models.

The company is also reportedly working with Chinese startup InnoStar Semiconductor on memory technology for the project. That part matters because advanced AI chips often depend on expensive and limited high bandwidth memory. If ByteDance can use another memory approach, it may reduce its exposure to global HBM supply problems and export restrictions.

ByteDance does not appear to be building everything alone. The report says the company may rely on several external partners for chip design and manufacturing. That is not unusual in the chip industry, where companies often design parts of a product while using partners for silicon production, memory, packaging, and other specialized work.

The timing is important. China’s AI companies are facing tighter access to advanced US hardware, especially Nvidia accelerators. The Chinese government has also pushed local companies to use more domestic chips. That has made custom silicon more attractive for large firms that can afford long term chip development.

ByteDance has already been moving in this direction. The company started work on its SeedChip AI accelerator with TSMC in 2024, and that chip is expected to tape out and move toward mass production. A custom AI CPU could sit alongside that accelerator in future server designs, giving ByteDance more control over its AI infrastructure.

Still, ByteDance is unlikely to replace Nvidia quickly. The report suggests the company may use hybrid server architectures for now because Nvidia hardware remains important for many AI workloads. The bigger shift may happen over time, as ByteDance builds more internal hardware and local suppliers improve their chips.

The broader trend is clear. Large AI companies no longer want to rely fully on outside chip vendors. Google, Amazon, Meta, and other major firms have already invested in custom silicon. ByteDance now appears to be moving in the same direction, but with the added pressure of US export controls and China’s push for semiconductor independence.

If the project succeeds, ByteDance could lower costs, improve control over its AI systems, and reduce its exposure to political restrictions. The risk is that custom chips are expensive, difficult to build, and slow to mature. For now, this looks like an early step, but it shows how seriously ByteDance is preparing for an AI market where access to chips may matter as much as the models themselves.