Chinese AI chip prices climb as HBM memory runs short

Chinese producers of AI accelerators are increasing list and indicative prices because they cannot secure enough high-bandwidth memory (HBM) at normal market rates. HBM is a stacked, high-throughput memory family essential for modern AI workloads. A small number of suppliers—Micron, Samsung Electronics and SK Hynix—supply advanced HBM at scale.

Since the United States tightened export controls on advanced HBM in December 2024, some Chinese buyers have turned to unofficial procurement channels to get the chips they need. Those channels cost more than regular market routes. Chipmakers are passing the extra cost to customers and raising quotes for both current products and upcoming releases.

Who is affected and how much prices rose

• Cambricon raised its planned pricing for the upcoming 690 chip by about 20%–30%.
• Smaller domestic players such as Iluvatar CoreX and MetaX have made similar price adjustments.

Higher memory costs are redirecting supply and changing delivery patterns. Iluvatar CoreX reportedly doubled GPU deliveries to ByteDance this year to 100,000 units and diverted chips originally intended for internal use to meet customer demand. Huawei remains a primary supplier to ByteDance, followed by Cambricon and Iluvatar CoreX, but delivery bottlenecks and elevated prices limit expansion of AI compute capacity in the region.

China aims to reduce reliance on foreign AI infrastructure and fill a market gap estimated at $50 billion created by restrictions on Nvidia-related imports. The current HBM shortage imposes financial pressure on that plan by raising costs for domestic accelerators and slowing rollouts. Some domestic memory manufacturers are moving to produce HBM variants, but commercial timelines, yields and volume remain uncertain.

• Commercial ramp and qualification of domestic HBM sources and their ability to scale to accelerator makers' needs.

The shortage of HBM and export controls are directly raising the cost of Chinese AI accelerators. That increases prices for customers and complicates China's pursuit of a self-sufficient AI hardware stack, at least until more affordable, reliable HBM supply becomes available.