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

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

ByteDance has been steadily raising investment in computing power in recent years and has split its training and inference chip supply chains. If a deal is sealed, Iluvatar will become the Beijing-based firm'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 Shanghai-based Iluvatar CoreX's Zhikai series provides massive online traffic-oriented inference capabilities.

China’s internet giants are all building data center and AI computing capacity at scale. Baidu is rolling out AI clusters nationwide with tens of thousands of GPUs, and Alibaba Group Holding's capital expenditure on cloud and intelligent computing hardware exceeded CNY38 billion (USD5.6 billion) in the first quarter of fiscal year 2026, and it expects to invest more than CNY380 billion (USD56.2 billion) over the next three years.

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

Major internet firms are building multi-vendor computing infrastructures to ensure supply security, strengthen procurement leverage, and optimize costs, Li Yuxuan, head of AI infrastructure technology at ModelBest, told Yicai. He noted that inference workloads account for a far larger share of computing demand than model training, while placing less stringent requirements on chip interconnects, memory bandwidth, and software ecosystems.

Chinese chips have already reached a usable level for inference workloads, making a broader range of suppliers viable options for large-scale commercial deployments, Li said.

China’s LLM sector is entering a period of rapid application growth this year, according to Wang Zhan, co-chief executive 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 do well in terms of cost-effectiveness and performance per watt in specific use cases.

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 released by China Insights Consultancy.

Demand for AI inference chips is growing rapidly, with the market forecast to increase to CNY3 trillion (USD453.9 billion) by 2030, with the Chinese market alone nearing CNY1.2 trillion, the report said.

Editor: Futura Costaglione