China’s AI chipmakers hike prices up to 50% as high-bandwidth memory shortage bites

China’s leading AI chipmakers are pushing through aggressive price increases on their accelerator products, driven by a worsening shortage of high-bandwidth memory, the specialized component that makes modern AI chips actually useful. The increases, ranging from 20% to 50% depending on the product, reflect the compounding pressure of US export restrictions and a global AI infrastructure buildout that has stretched HBM supply chains to their limits.

The price tag keeps climbing

Huawei’s Ascend 950DT accelerator card now costs more than 250,000 yuan, roughly $37,255. That’s a 20-50% jump compared to quotes from just two months ago.

The company’s older models haven’t been spared either. The Ascend 950PR has climbed approximately 30% to exceed 80,000 yuan since the start of 2026. The Ascend 910C has risen about 22%, now priced above 110,000 yuan.

Expected availability for some models has been pushed to Q4 2026 at the earliest, meaning companies placing orders today are essentially locking in premium prices for hardware they won’t receive for months.

Huawei isn’t alone. Cambricon, one of China’s other prominent AI chip designers, has raised the price of its upcoming 690 chip by 20-30%. Smaller players like MetaX and Iluvatar CoreX have implemented similar increases in the same range.

Why HBM is the bottleneck

High-bandwidth memory stacks memory chips vertically and connects them with through-silicon vias, enabling the massive data throughput that large language models and other AI workloads demand.

US export restrictions on advanced HBM took effect in December 2024, specifically targeting the kinds of memory modules that power cutting-edge AI accelerators. For Chinese chipmakers, this meant their primary channels for sourcing top-tier HBM from suppliers like SK Hynix and Samsung effectively closed.

The result has been a migration toward grey-market sources, where HBM is available but at significantly higher prices and with far less supply reliability. Major suppliers have allocated most of their output for 2026 already, leaving latecomers and restricted buyers fighting over scraps.

The scramble for supply

Iluvatar CoreX doubled its shipments to ByteDance to 100,000 GPU units by reallocating inventory from other customers.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.