(Yicai) June 18 -- Chinese internet giant ByteDance is in talks with Shanghai Iluvatar CoreX Semiconductor to purchase 50,000 artificial intelligence chips, according to industry sources.

The AI chips involved in the negotiations are the Zhikai series cloud inference graphics processing units, which will be used for large language model inference work, and the Tiangai series for training scenarios, Yicai learned from the insiders.

In recent years, ByteDance has continuously invested in computing power and separated its chip supply for training and inference. If the deal is sealed, Iluvatar CoreX will become ByteDance's third Chinese AI chip supplier after Huawei Technologies and Cambricon Technologies.

Huawei's Ascend chip series focuses on cluster training and serves as the foundation for pre-training AI models, Cambricon's chips cater to some mid-to-high-end inference needs, and Iluvatar CoreX's Zhikai series provides massive online traffic-oriented inference capabilities.

Other Chinese internet giants are also intensively expanding their data center operations. For example, Baidu is establishing AI computing power clusters with a scale of tens of thousands of cards nationwide. Alibaba Group Holding's expenditure on cloud and intelligent computing hardware exceeded CNY38 billion (USD5.6 billion) in the first quarter of fiscal year 2026, with projections to invest over CNY380 billion (USD56.2 billion) over the next three years.

Meanwhile, Tencent Holdings is building high-performance computing clusters in various locations across the country, planning to introduce domestic computing power on a large scale in the second half of the year. The firm provides services, such as model-as-a-service for the Hunyuan LLM, AI-generated content for gaming, and video generation inference to external clients.

Major internet firms are building multi-vendor computing power systems primarily based on supply certainty, bargaining power, and cost structure, Li Yuxuan, head of AI infrastructure technology at ModelBest, told Yicai. For them, inference volume demand is much greater than that for training, but the requirements for inference chips are significantly lower than those for training.

Domestic chips have reached a usable level on the inference side, which allows more suppliers to become viable engineering options, Li said.

This year marks the explosive application period for LLMs in China, said Wang Zhan, co-chief executive officer of Sunrise. Models such as DeepSeek V4 have led to a surge in the token market, and agents are rapidly becoming popular.

The key to competition in the industry has shifted to who can offer the lower token cost, Wang noted, adding that domestic inference chips have demonstrated good performance in terms of cost-effectiveness and performance per watt in specific scenarios.

The AI chip industry is undergoing a fundamental paradigm shift from being training-centered to inference-centered, according to the global AI inference chip industry report by China Insights Consultancy.

Demand for AI inference chips is growing rapidly, and it is estimated that the industry will reach a market size of CNY3 trillion (USD453.9 billion) by 2030, with the Chinese market alone nearing CNY1.2 trillion, the report predicted.

Editor: Futura Costaglione