Chinese AI chipmakers, including Huawei and Cambricon, have sharply raised prices for current and next-generation processors as a worldwide shortage of high-bandwidth memory pushes up their costs. Three people familiar with the matter told Reuters that the increases range from 20% to 50% depending on the product and contract terms.
Huawei lifted the indicated price of its Ascend 950DT accelerator card to more than 250,000 yuan, about $37,255, a jump of 20% to 50% from quotes given to customers just two months ago. The company has said the 950DT, its most advanced AI chip, will be available in the fourth quarter of 2026. The card pairs an AI processor with memory and other components on one board.
Cambricon repriced its planned next-generation chip, tentatively called the 690, at 20% to 30% above the levels it indicated two months earlier. Smaller rivals MetaX and Iluvatar CoreX made similar moves. None of the four companies responded to requests for comment, and the people who described the pricing asked not to be named because the discussions are not public.
Why memory is the bottleneck
High-bandwidth memory, HBM, consists of vertically stacked memory chips that let processors read and write huge amounts of data at speed. It is a core component of every AI accelerator, and the advanced HBM market is dominated by three suppliers: South Korea’s SK Hynix and Samsung Electronics, and US-based Micron Technology.
Chinese designers sit at the end of that queue. US export controls restrict China’s access to Nvidia’s most advanced processors, which left a roughly $50 billion domestic market for local alternatives, but the same controls and supply tightness make HBM hard to secure. Huawei has said its Ascend 950 series will use two proprietary HBM technologies, HiBL 1.0 for the 950PR and HiZQ 2.0 for the 950DT, without detailing where or how the memory is manufactured.
Older chips are more expensive too
The repricing reaches back through the product line. Huawei’s Ascend 950PR, which sold for roughly 60,000 yuan per card at the start of the year, now fetches more than 80,000 yuan, up about 30%. The older Ascend 910C board has climbed to more than 110,000 yuan from about 90,000 yuan.
For Chinese companies building AI computing capacity, the price hikes land as a direct cost increase, and they expose a growing bottleneck for the domestic chip sector. Beijing has been pressing firms to replace Nvidia products with home-grown alternatives, and demand is concentrated among a handful of buyers. Huawei remains ByteDance’s largest domestic chip supplier, followed by Cambricon and Iluvatar CoreX, according to the sources. Iluvatar CoreX has doubled its GPU shipments to ByteDance this year to 100,000 units and has reportedly reassigned GPUs originally meant for its own use to meet the TikTok owner’s requirements. ByteDance did not respond to a request for comment.
| Chip | Price earlier in 2026 | Price now |
|---|---|---|
| Huawei Ascend 950DT | Implied 167,000-208,000 yuan | Above 250,000 yuan |
| Huawei Ascend 950PR | About 60,000 yuan | Above 80,000 yuan |
| Huawei Ascend 910C board | About 90,000 yuan | Above 110,000 yuan |
| Cambricon 690 (planned) | Indicated price | Up 20-30% |
Relief may be coming
Domestic HBM supply is starting to move. According to Seoul Economic Daily, citing sources from The Information, China’s CXMT has begun small-scale production of HBM3E and plans a significant output expansion in 2027. Chinese chip designers, including Cambricon and Alibaba’s T-Head, are testing that memory with their processors. If CXMT scales as planned, the dependence on Korean and American memory suppliers would shrink, though not before several quarters of tight supply.
Demand on the other side is also growing. Bloomberg reported, via TechNews, that DeepSeek plans to deploy at least 160,000 Ascend 950DT chips at a large data center under construction in Inner Mongolia, mainly for AI inference. Orders of that size from a single buyer show how quickly Chinese AI firms are absorbing whatever domestic supply exists. More buyers chasing constrained memory suggests the pricing pressure will not ease quickly, and Chinese cloud builders will be paying Nvidia-market prices for home-grown silicon for some time.