China’s leading artificial intelligence chipmakers are sharply raising prices for current and next-generation processors as a global shortage of high-bandwidth memory squeezes efforts to build domestic alternatives to Nvidia, according to people familiar with the matter.
Huawei Technologies has raised the indicated price of its Ascend 950DT accelerator card to more than 250,000 yuan ($37,255), two people said. The increase represents a rise of between 20% and 50% from prices quoted to customers just two months ago, depending on contract terms.
The Shenzhen-based company has said publicly that the 950DT, its most advanced AI chip, will become available in the fourth quarter of 2026.
Beijing-based Cambricon has also raised the indicative price of its next-generation processor, tentatively known as the 690, by between 20% and 30% compared with levels indicated two months ago, according to two sources. Smaller Chinese rivals MetaX and Iluvatar CoreX have made similar increases, the sources said.
The people declined to be identified because the pricing discussions are not public.
The increases illustrate a major challenge for China’s AI industry as Beijing pushes domestic companies to reduce their reliance on Nvidia and develop home-grown alternatives. Even as Chinese chipmakers seek to expand production and capture a larger share of the country’s AI computing market, they remain exposed to a critical component whose supply is heavily concentrated outside China.
High-bandwidth memory, or HBM, consists of vertically stacked memory chips that allow AI processors to access large amounts of data at high speed. It is a crucial component in AI computing, particularly as increasingly demanding applications require processors to handle vast quantities of information rapidly.
The advanced HBM market is dominated by South Korea’s SK Hynix and Samsung Electronics, as well as US-based Micron Technology.
Since Washington tightened export controls on certain advanced HBM products to China in December 2024, Chinese chipmakers have increasingly relied on grey-market channels to obtain supplies, according to the sources.
Those supplies typically cost several times more than the prices paid by buyers outside China, they said. Because memory represents a substantial portion of an AI accelerator’s production cost, the increase is flowing directly into the price of finished accelerator cards.
Huawei has said that its Ascend 950 series will use two proprietary HBM technologies, HiBL 1.0 for the 950PR and HiZQ 2.0 for the 950DT. The company has not detailed where or how the memory is manufactured.
The two processors are designed for different stages of AI computing. The 950DT is intended mainly for developing AI models and generating responses, while the 950PR processes user requests before they are passed to the model.
Huawei, Cambricon, MetaX and Iluvatar CoreX did not respond to requests for comment.
The pressure is not limited to China’s newest processors. Prices of Huawei’s earlier-generation products have also risen sharply this year.
The Ascend 950PR, which sold for about 60,000 yuan per card at the beginning of the year, now commands more than 80,000 yuan, representing an increase of roughly 30%, according to two sources.
The older Ascend 910C board has also increased in price, rising to more than 110,000 yuan from about 90,000 yuan at the beginning of the year, the sources said.
The increases are adding to the cost of expanding AI computing capacity in China at a time when domestic chipmakers have benefited from a gap created by US export restrictions on Nvidia’s most advanced processors.
Those restrictions have helped create an opening in China’s $50 billion AI chip market for domestic manufacturers. But the HBM shortage demonstrates that replacing Nvidia is not simply a question of designing alternative processors.
The supply constraints are forcing Chinese technology companies to make difficult choices over how scarce computing resources are allocated.
Iluvatar CoreX has doubled its shipments of graphics processing units, or GPUs, to ByteDance, the technology giant behind TikTok, to 100,000 units this year, according to one source. The increase comes as computing constraints have intensified.
The company has also diverted GPUs originally intended for its own internal use to help meet ByteDance’s requirements, the source said.
The development illustrates the growing competition among Chinese technology companies for domestic AI computing resources as demand expands while access to key components remains constrained.
Cambricon has yet to formally launch its 690 processor, with the company’s current flagship remaining the 590. Huawei is ByteDance’s largest domestic chip supplier, followed by Cambricon and Iluvatar CoreX, according to the three sources.
ByteDance did not respond to a request for comment.
For Beijing, the price increases expose a difficult contradiction at the centre of China’s drive for technological self-reliance. US export controls have created powerful incentives for Chinese companies to develop domestic AI processors, while simultaneously restricting access to advanced components needed to compete at the highest level.
The result is an industry attempting to expand its domestic chip ecosystem while paying a premium for critical memory through alternative supply channels.
For Chinese AI companies, that means the race to build a domestic alternative to Nvidia is becoming not only a contest of chip design, but also a struggle over access to the components needed to make those chips commercially viable.
The surge in HBM costs therefore threatens to make China’s push for AI chip independence more expensive just as demand for domestic processors is increasing. With suppliers already reshuffling allocations and some companies diverting their own computing resources to meet major customers’ needs, the memory shortage is emerging as a significant bottleneck in China’s attempt to build a self-sufficient AI hardware industry.
