Exclusive-China’s AI chipmakers raise costs as high-bandwidth reminiscence scarcity bites

Sept 10 (Reuters) – Chinese AI chipmakers including Huawei Technologies and Cambricon have sharply raised costs for present and next-generation AI processors, three people acquainted with the matter said, as hovering prices for high-bandwidth reminiscence squeeze efforts to construct home options to Nvidia.

Huawei ‌has lifted the indicated price of its Ascend 950DT accelerator card, which integrates an AI processor with reminiscence and other parts, to ‌above 250,000 yuan ($37,255), two people said.

The increase represents a leap of 20% to 50% from costs quoted to clients just two months in the past, relying on contract phrases, they said. Shenzhen-based Huawei ​has publicly said the 950DT, its most superior AI chip, will be accessible in the fourth quarter of 2026.

Beijing-based Cambricon has also repriced its next-generation chip, tentatively called the 690, at 20% to 30% above ranges it indicated two months in the past, two sources said, without giving an precise determine. Smaller rivals MetaX and Iluvatar CoreX have made comparable will increase, the sources said.

The people declined to be recognized because the pricing discussions are not public.

The price hikes illustrate how a ‌worldwide scarcity of high-bandwidth reminiscence (HBM) is reverberating through China’s ⁠AI industry as Beijing presses corporations to exchange Nvidia merchandise with home-grown options.

HBM is a essential part for AI computing, consisting of vertically stacked reminiscence chips that permit processors to access huge quantities of knowledge at high speeds. The ⁠superior HBM market is dominated by South Korea’s SK Hynix and Samsung Electronics, and U.S.-based Micron Technology.

Since Washington tightened controls on exports of certain superior HBM merchandise to China in December 2024, Chinese chipmakers have more and more relied on grey-market channels to secure provides, the sources said.

Such HBM sometimes prices several occasions what patrons outside China ​pay, ​they said. Because reminiscence accounts for a massive share of an AI accelerator’s manufacturing ​value, the larger costs are feeding immediately into completed playing cards.

Huawei ‌has said the Ascend 950 collection will use two proprietary HBM applied sciences: HiBL 1.0 for the 950PR and HiZQ 2.0 for the 950DT, without detailing where or how the reminiscence is manufactured.

The 950DT is constructed primarily for developing AI fashions and producing responses, while the 950PR processes consumer requests before they are handed to the model.

Huawei, Cambricon, MetaX and Iluvatar CoreX did not reply to requests for remark.

OLDER HUAWEI CHIPS ALSO MORE EXPENSIVE

Prices of Huawei’s earlier-generation merchandise have also climbed. The Ascend 950PR, which offered for roughly 60,000 yuan per card at the begin of the yr, now fetches more than ‌80,000 yuan, up about 30%, two of the sources said.

The older Ascend 910C board ​has risen to more than 110,000 yuan from about 90,000 yuan at the begin of ​the yr, they said.

The price hikes increase the value of building ​AI computing capability and spotlight a growing bottleneck for Chinese chipmakers, who benefited from a void in China’s $50 billion ‌AI chip market after U.S. export controls restricted the nation’s ​access to Nvidia’s most superior processors.

The scarcity ​is forcing suppliers to reshuffle allocations.

Iluvatar CoreX has doubled its shipments of graphics processing items (GPUs), a kind of chip used for coaching and working AI systems, to tech giant and TikTok developer ByteDance to 100,000 items this yr, according to one source, as computing constraints ​intensified.

The company has also diverted GPUs earmarked for inner ‌use to help meet ByteDance’s wants, the source said.

Cambricon has not formally launched the 690. Its present flagship is the 590. ​Huawei is ByteDance’s largest home provider, adopted by Cambricon and Iluvatar CoreX, the three sources said.

ByteDance did not reply to a ​request for remark.

(Reporting by Reuters employees; Editing by Eduardo Baptista and Shri Navaratnam)