HBM shortage drives up Huawei Ascend 950DT price to over $37,000
Reuters reports that the anticipated Huawei Ascend 950DT AI accelerator will cost over 250,000 Yuan due to global HBM scarcity and US export restrictions. Prices for other Chinese chips like Cambricon and MetaX have also risen significantly.
HBM shortage is driving up prices for China's AI chips, with Huawei's upcoming Ascend 950DT reportedly costing over 250,000 yuan according to Reuters.
According to a September 10, 2026 report by Reuters, Chinese manufacturers have significantly raised prices for their current and upcoming AI accelerators. Huawei is expected to charge more than 250,000 yuan (approximately $37,000 USD) for the Ascend 950DT, which is anticipated in Q4 2026. This represents a price increase of 20% to 50% compared to offers made just two months prior.
While Huawei has not officially confirmed these specific figures, Reuters cites multiple sources familiar with the negotiations. Other Chinese chipmakers, including Cambricon, MetaX, and Iluvatar CoreX, have also reportedly increased their prices by similar margins. For context, the Ascend 950PR model has risen from around 60,000 yuan to over 80,000 yuan, while the Ascend 910C board now costs more than 110,000 yuan.
The core driver behind this surge is a global shortage of High Bandwidth Memory (HBM). Unlike standard DDR memory, HBM uses vertical stacking of DRAM dies and requires extremely wide interconnects to the accelerator. This architecture delivers multi-terabyte per second bandwidth but drastically increases manufacturing complexity, wafer usage, and packaging costs.
Compounding the issue are US export restrictions tightened in December 2024, which limit China's access to modern HBM products. Consequently, Chinese chipmakers are increasingly relying on alternative or gray-market supply chains. The HBM available through these channels can cost several times more than standard market rates, and this premium is directly passed on to the final product price.
This creates a strategic bottleneck for China's tech sector. Domestic cloud and internet giants are forced to use local accelerators due to US restrictions, yet they still require massive amounts of high-speed memory to build competitive AI systems. The memory shortage persists regardless of whether the accelerator is made by NVIDIA, AMD, or a domestic alternative like Huawei.
Reuters estimates the Chinese market for AI chips at around $50 billion. Huawei remains the primary local supplier for major clients like ByteDance, followed by Cambricon and Iluvatar CoreX. The latter reportedly doubled its deliveries to ByteDance this year, even diverting units originally intended for its own use.
Ultimately, the HBM shortage is shifting from a mere availability issue to a pricing crisis. While China can develop its own processors to reduce reliance on NVIDIA, it cannot bypass the fundamental need for high-bandwidth memory. Until the supply chain stabilizes, the cost of building new AI data centers will remain prohibitively high.
Source: Reuters (cited by TrendForce analysis confirming Huawei's Q4 2026 timeline).
Гика
