According to a September 16 report by The Information, ByteDance is working with TSMC to independently develop two AI chips, aiming to reduce its dependence on NVIDIA's expensive chips and enhance its market competitiveness.

ByteDance plans to cooperate with TSMC and expects to achieve mass production of these two self-developed chips by 2026.

These two chips are designed specifically for model training and inference. One is used to accelerate the massive data calculations required for building AI models, known as the training chip; the other is used to accelerate model-based prediction and decision-making processes, known as the inference chip.

ByteDance has pre-ordered hundreds of thousands of chips from TSMC, and is expected to save billions of dollars in costs by purchasing self-developed chips.

Last year, ByteDance launched its first AI chatbot named "Doubao" and introduced a series of low-cost large language models this year. Meanwhile, ByteDance's expenses in developing generative AI models have continued to rise, with orders of over 200,000 NVIDIA H20 chips this year, totaling over $2 billion. ByteDance's chip development efforts follow a similar trend as companies like Google, Amazon, and Microsoft in recent years, which is to reduce reliance on NVIDIA chips.