ByteDance Aims for 50,000 Local AI Chips, China's Computing Needs Soar

A row of servers in an AI data center. ByteDance is reportedly exploring the purchase of 50,000 Chinese-made AI chips to expand its computing capacity. (Photo Yica Global)

JAKARTA - ByteDance is reportedly exploring the purchase of 50,000 domestically produced AI chips to expand its computing capacity. This move comes as China's major internet companies compete to strengthen their artificial intelligence infrastructure.

According to a report by Yicai Global quoted on Thursday, June 18, the parent company of TikTok is negotiating with Shanghai Iluvatar CoreX Semiconductor to buy tens of thousands of AI chips.

An industry source quoted by Yicai said the chips discussed consisted of the Zhikai series for AI inference needs and the Tiangai series for AI model training.

Inference is the stage when the AI model that has been trained is used to answer questions, create content, or execute user commands.

If the deal is reached, Iluvatar CoreX will become ByteDance's third local AI chip supplier after Huawei Technologies and Cambricon Technologies.

Each supplier has a different role. Huawei's Ascend chips are used for training large-scale AI models. Cambricon targets medium to high-end inference needs. Meanwhile, Zhikai from Iluvatar CoreX is designed to handle large amounts of user traffic online.

According to Yicai Global, ByteDance in recent years has continued to increase its computing investment and has begun to separate the chip supply chain for training and inference needs.

The race is also being waged by other Chinese internet giants.

Baidu is building an AI computing cluster with tens of thousands of processing cards in various regions of China.

Alibaba poured more than 38 billion yuan or about 5.6 billion US dollars into cloud computing and intelligent computing hardware in the first quarter of fiscal year 2026. In the next three years, Alibaba's investment in the sector is expected to exceed 380 billion yuan or about 56.2 billion US dollars.

Meanwhile, Tencent is building high-performance computing centers in various regions and plans to expand the use of domestic computing power in the second half of this year.

Tencent uses the infrastructure for a variety of services, ranging from Hunyuan's large language model, AI-based content creation for games, to video inference services for external customers.

Head of AI ModelBest Infrastructure Technology Li Yuxuan said large internet companies are now building computing systems with multiple suppliers. Considerations include supply certainty, price bargaining power, and cost structures.

According to Li, the current inference needs are much greater than the needs for training AI models. However, the chip requirements for inference are much lower than those for training.

"Domestic chips have reached a level that is suitable for use for inference. Therefore, more and more suppliers can be a realistic technical choice," Li told Yicai Global.

Sunrise Co-CEO Wang Zhan believes this year is a period of explosive use of large language models in China. Models such as DeepSeek V4 trigger a surge in the AI token market, while AI agents are growing rapidly.

An AI agent is a system that can carry out tasks automatically on behalf of a user.

According to Wang, industry competition is now shifting to the ability of companies to offer cheaper token costs. In certain scenarios, China's artificial inference chips are considered to show good cost and power consumption efficiency.

The China Insights Consultancy report said the global AI chip industry is undergoing a major change, from previously centered on model training to being more focused on inference.

Demand for AI inference chips is expected to continue to grow rapidly. Its global market value is projected to reach 3 trillion yuan or about 453.9 billion US dollars by 2030. China's market alone is expected to approach 1.2 trillion yuan.