Chinese AI Chipmakers Raise Prices Amid High-Bandwidth Memory Shortage

Major Chinese artificial intelligence chipmakers, including Huawei Technologies and Cambricon, have initiated substantial price increases across their current and next-generation processor lineups. The adjustments come as severe supply bottlenecks and rising international acquisition costs for high-bandwidth memory begin to impact China's domestic semiconductor industry. According to industry sources, memory cost pressures are filtering directly into customer quotes as Chinese technology companies attempt to construct indigenous alternatives to restricted hardware from Nvidia.

The market shifts underscore how global memory constraints are compounding operational hurdles for Chinese hardware vendors. Following export controls enacted by the United States government in December 2024 that constrained access to advanced high-bandwidth memory products, Chinese chip designers have increasingly sourced crucial components through grey-market channels. This reliance has elevated production expenditures, driving up total bill-of-materials costs for AI accelerator boards across the domestic sector.

The price adjustments are led by Shenzhen-based Huawei Technologies, which has increased the indicated price of its flagship Ascend 950DT accelerator card to more than 250,000 yuan, equivalent to roughly $37,255. Sources familiar with contract discussions noted that the current figure represents an increase of 20% to 50% compared to customer quotes issued just two months prior, depending on specific contractual terms. Huawei has previously slated the Ascend 950DT, its most capable artificial intelligence processor, for commercial availability in the fourth quarter of 2026.

Beijing-based designer Cambricon has enacted parallel price increases on its upcoming enterprise silicon. The firm has repriced its next-generation processor, tentatively designated as the 690, at levels 20% to 30% higher than estimates offered two months ago. Smaller domestic market participants, including MetaX and Iluvatar CoreX, have similarly adjusted their pricing expectations upward by comparable percentages. None of the four chip designers responded to formal requests for comment regarding the pricing revisions.

The widespread revisions reflect systemic cost increases rather than isolated commercial decisions. As memory components represent a substantial percentage of total manufacturing expense for high-density computing boards, finished accelerator prices have risen in direct response to component-level inflation across domestic supply chains.

The cost escalations extend beyond upcoming hardware releases into Huawei's currently available silicon product lines. The Ascend 950PR accelerator card, which traded at approximately 60,000 yuan per unit at the start of the year, is now quoted at over 80,000 yuan, marking a price jump of approximately 30%. Similarly, Huawei's older Ascend 910C board has increased from roughly 90,000 yuan earlier in the year to over 110,000 yuan.

The two variants within the Ascend 950 generation fulfill distinct operational roles in enterprise data centers. The higher-priced 950DT is designed primarily for building and training large language models as well as generating AI responses. In contrast, the 950PR focuses on handling pre-processing tasks for inbound user requests before passing data to underlying models.

To power these systems, Huawei disclosed that the Ascend 950 series utilizes two proprietary high-bandwidth memory architectures, incorporating HiBL 1.0 for the 950PR and HiZQ 2.0 for the 950DT. However, the company has not publicly clarified the specific manufacturing origin or fabrication partners responsible for producing these specialized memory stacks.

The underlying driver of the price increases is an acute global shortage of high-bandwidth memory, a vital technology consisting of vertically stacked DRAM chips that enable high-speed data transfer to processing cores. The global market for advanced high-bandwidth memory is heavily concentrated among three primary producers: South Korea's SK Hynix and Samsung Electronics, alongside U.S.-based Micron Technology.

Following the tightening of U.S. export controls in December 2024 aimed at limiting advanced semiconductor deliveries to China, domestic chipmakers were largely cut off from direct supply arrangements for cutting-edge memory stacks. In response, Chinese vendors turned to indirect grey-market acquisition routes to secure necessary component volume.

Industry sources report that high-bandwidth memory procured through these grey-market channels carries premium pricing that is several times higher than baseline market rates paid by international competitors outside China. Because high-bandwidth memory accounts for a dominant share of total component expense in AI accelerators, these inflated procurement costs pass directly to enterprise buyers.

The escalating hardware costs and restricted memory allocations are forcing Chinese chip developers to reorganize customer delivery priorities. Iluvatar CoreX has doubled its graphics processing unit shipments to ByteDance, the parent company of TikTok, reaching 100,000 units this year. To meet ByteDance's computing needs, Iluvatar CoreX diverted GPU inventory originally allocated for its internal operational requirements.

ByteDance represents one of the largest corporate consumers of artificial intelligence hardware within China's domestic ecosystem. Industry data indicates that Huawei currently serves as ByteDance's primary domestic chip supplier, with Cambricon ranking second and Iluvatar CoreX holding the third position.

These supply reallocations highlight the operational pressures facing Chinese technological enterprises as they navigate a local artificial intelligence chip market valued at $50 billion. While domestic vendors gained significant market opportunity following U.S. restrictions on Nvidia's flagship chips, memory procurement bottlenecks continue to complicate sustained infrastructure expansion across China.

Looking ahead to the final quarter of 2026, the anticipated rollout of Huawei's Ascend 950DT will provide a key test for domestic high-density AI cluster deployments. While Cambricon continues development toward the formal commercial release of its 690 processor to succeed its current flagship 590 chip, enterprise buyers face elevated hardware costs across all domestic tiers.

The sustained cost premium required to procure grey-market memory stacks indicates that Chinese AI infrastructure builders will face elevated capital expenditure demands in the near term. As domestic chipmakers balance internal inventory with contract commitments to major technological clients like ByteDance, component availability and pricing volatility remain central factors shaping the domestic computing sector.