HBM supply is tight, and Chinese AI chip manufacturers have collectively raised product prices

Overseas export controls have tightened and HBM supply is tight. Domestic AI chip manufacturers have adjusted their prices. The latest price of Huawei's Ascend 950DT accelerator card has exceeded 250,000 yuan ($37,255). Compared to the price two months ago, the increase range is between 20% and 50%. This product is planned to be launched in the fourth quarter of 2026. Besides the 950DT, the prices of older Ascend models are also rising.

At the beginning of the year, the Ascend 950PR single card was about 60,000 yuan. Now, the quoted price has exceeded 80,000 yuan, with an increase of nearly 30%. The earlier generation Ascend 910C, which was priced at 90,000 yuan at the beginning of the year, has now surpassed 110,000 yuan. Cambricon has not officially released the 690 chip yet, but has already raised the provisional price of this new product by 20% to 30%. Other smaller chip manufacturers have also raised the prices of their products.

HBM is the core storage component of AI acceleration cards, and provides high speed data reading and writing capabilities for AI processors. The high-end HBM market is mainly controlled by three companies: Samsung, SK Hynix and Micron. Starting from December 2024, the United States tightened its export rules for advanced HBM to China. Chinese domestic chip enterprises find it difficult to purchase high-end HBM in large quantities through regular channels and can only turn to the gray market to acquire it.

The HBM in the gray market is priced much higher than the normal quotations in overseas markets. In the production cost of AI acceleration cards, the memory component accounts for a high proportion. The increase in HBM procurement costs will be passed on to the finished products. Huawei uses its own storage solution in the Ascend 950 series. The 950PR is equipped with HiBL 1.0, while the 950DT uses HiZQ 2.0. The two chips have different positioning, with 950DT used for large model training and generation, and 950PR responsible for the pre-processing of model inference.

The expansion of China's AI large model industry has led to high demand for computing power from leading internet enterprises. Under the condition of limited production capacity, chip manufacturers give priority to ensuring orders from major clients. Iluvatar CoreX has adjusted its supply volume to ByteDance's GPUs, increasing the shipment to ByteDance to 100,000 units this year, a direct doubling compared to before. The company even allocated the GPU production capacity that was originally reserved for its own use to meet ByteDance's computing power procurement needs.

ByteDance is a major purchaser of computing power in China. The domestic AI chip suppliers in the supply chain are divided into several tiers. Huawei is ByteDance's biggest local chip supplier in China. Cambricon ranks second, followed by Iluvatar CoreX. With the explosive demand for computing power, the market maintains an optimistic expectation for orders of domestic AI chips. However, the increase in chip prices will change the computing power procurement budget of AI enterprises.

Due to export control restrictions, the Chinese domestic market is unable to obtain Nvidia's high-end AI chips normally. The market space that originally belonged to Nvidia has been given to domestic AI chip manufacturers. The domestic AI chip market size is $50 billion. The supply chain of local manufacturers has obvious shortcomings. The shortage of HBM is currently the biggest bottleneck.

The increase in chip prices will have a dual impact. From the perspective of chip manufacturers, the price hike can offset the cost pressure brought by HBM and improve the revenue level of the enterprises. From the perspective of the purchasers, the rising procurement costs will compress the profits of AI enterprises. For some small and medium-sized AI startups, the threshold for building a computing power cluster will further increase.