Now various capitals are competing for market share for AI infrastructure, AI model training, and AI database construction, and ByteDance is also developing its own AI GPU to reduce its dependence on Nvidia and cope with US export controls. According to reports, ByteDance has designed two GPUs, one for AI training and the other for AI reasoning, which are expected to be manufactured by TSMC using the N4/N5 process and are scheduled to be deployed around 2026. These GPUs will cost less than Nvidia, but their performance may not exceed Nvidia's existing products due to export restrictions.
ByteDance invested more than $2 billion this year and purchased more than 200,000 Nvidia H20 GPUs, each priced at about $10,000. However, the shortage and high price of Nvidia GPUs have pushed ByteDance to develop its own hardware. Despite this, ByteDance still faces the challenge of building a compatible software ecosystem for its self-developed GPUs. Currently, its AI tasks rely on Nvidia's CUDA and supporting tools.
Nvidia expects to deliver more than 1 million H20 GPUs to China this year, with revenue exceeding $12 billion, far exceeding the sales and revenue of other AI hardware manufacturers in China.
