Chinese AI chipmakers raise prices amid growing memory shortages and export restrictions.

Huawei’s Ascend 950DT pricing was reportedly set in line with Nvidia’s B200, and both companies sell their AI processors in batches of at least eight, with final prices varying according to order size and purchases of auxiliary networking equipment.
Huawei’s Ascend 950DT is designed for both AI-model training and the decode stage of inference, and is expected to use the company’s HiZQ 2.0 memory system, with 144GB of capacity and 4TB-per-second memory bandwidth.
The Ascend 950DT is expected to provide 2TB-per-second interconnect bandwidth and support several low-precision data formats intended to improve AI-computing efficiency.
Following tighter U.S. export controls imposed in December 2024, Chinese chipmakers have increasingly relied on grey-market channels to obtain advanced HBM, which sources said can cost several times more in China than for buyers elsewhere.
Huawei has not disclosed where or how the proprietary memory for its upcoming accelerator will be manufactured, leaving an additional uncertainty around the chip’s supply chain.
Chinese AI chipmakers are raising prices by 20% to 60% as a shortage of high-bandwidth memory (HBM) drives up production costs and squeezes their ability to compete with Nvidia. Inshorts reports that Huawei has lifted the price of its Ascend 950DT accelerator to more than 250,000 yuan, or about $37,000, while smaller competitors Cambricon, MetaX, and Iluvatar CoreX have made similar adjustments. The shortage is particularly acute in China following U.S. export restrictions imposed in December 2024, forcing chipmakers to rely on costly grey-market channels that charge several times more than global prices.
HBM is essential memory for training and running advanced AI models, but China has limited access because the technology is dominated by South Korean makers SK Hynix and Samsung, plus U.S.-based Micron. Modern Diplomacy notes that Huawei's new Ascend 950DT won't arrive until the fourth quarter of 2026, adding more uncertainty to Beijing's push for a self-sufficient AI computing industry.
Huawei set its Ascend 950DT price to match Nvidia's B200 processor, both sold in batches of at least eight units. Crypto Briefing reports that Huawei's price increase reflects the chip's advanced specs: 144GB of proprietary memory with 4TB-per-second bandwidth and 2TB-per-second interconnect bandwidth. The exact final price varies based on order size and additional networking equipment purchases, making direct comparisons difficult.
The pricing pressure is not limited to Huawei. ITP.net reports that Cambricon has raised preliminary pricing for its next-generation 690 processor by 20% to 30%, while smaller suppliers MetaX and Iluvatar CoreX have also increased their prices. These simultaneous hikes across competitors suggest the HBM shortage is industry-wide, forcing all Chinese chipmakers to pass costs onto customers.
U.S. export controls tightened in December 2024, severely limiting China's access to advanced HBM from global suppliers. Global Sources notes that Chinese chipmakers now rely on grey-market channels where advanced memory costs several times more than prices elsewhere in the world. This supply chain disruption adds another layer of cost pressure, making it nearly impossible for Chinese AI chips to compete on price.
Huawei has not disclosed where its proprietary HiZQ 2.0 memory will be manufactured, creating additional uncertainty about whether the Ascend 950DT can actually reach production at scale by late 2026. If supply remains constrained, Chinese chipmakers may struggle to meet demand even at higher prices.
The Ascend 950DT won't ship until the fourth quarter of 2026 — roughly 18 months away — giving Nvidia time to release newer generations and maintain its AI chip lead. Modern Diplomacy highlights that this delay compounds China's challenge: higher prices, uncertain supply, and a moving target from competitors. Beijing's goal of building a self-sufficient AI computing industry faces mounting headwinds as costs spiral and timelines slip.
Publishers
13
Articles
28
Reach
41