Chinese AI chipmakers including Huawei and Cambricon have reportedly raised customer price indications for current and upcoming accelerators as a global shortage of high-bandwidth memory, or HBM, increases production costs. The increases show that China's attempt to replace restricted Nvidia hardware faces a bottleneck beyond the processor itself.
Reuters reported that Huawei now indicates a price above 250,000 yuan, about $37,255, for its Ascend 950DT accelerator card. Depending on contract terms, that is 20% to 50% higher than quotes given two months earlier, according to people familiar with private pricing discussions.
Cambricon's planned 690 chip has reportedly been repriced 20% to 30% higher, while smaller competitors MetaX and Iluvatar CoreX have made similar changes. The companies did not respond to Reuters' requests for comment, and final prices may vary across customers and volumes.
Reported Chinese AI accelerator price changes
| Product | Earlier indication | Latest reported level | |---|---:|---:| | Huawei Ascend 950DT | Contract-dependent | Above 250,000 yuan; 20%–50% rise | | Cambricon 690 | Not disclosed | 20%–30% rise | | Huawei Ascend 950PR | About 60,000 yuan early in 2026 | Above 80,000 yuan | | Huawei Ascend 910C | About 90,000 yuan early in 2026 | Above 110,000 yuan |
These are reported market indications, not a universal public retail list. Enterprise hardware is often negotiated as part of larger systems including memory, networking, software and support.
What is high-bandwidth memory?
HBM is a specialised type of memory built by stacking multiple layers vertically and connecting them through extremely fast pathways. It sits close to an AI processor and supplies data at far higher bandwidth than conventional server memory.
Modern AI workloads perform enormous numbers of calculations in parallel. A powerful processor can remain underused if it must wait for model weights and intermediate data to arrive. HBM reduces that bottleneck, which is why memory capacity and bandwidth are central specifications for training and inference accelerators.
The advanced HBM market is dominated by South Korea's SK Hynix and Samsung Electronics and US-based Micron. Expanding supply requires sophisticated fabrication, packaging and testing, so production cannot increase instantly when AI demand rises.
Why the shortage affects China more severely
The United States tightened restrictions on exports of certain advanced HBM products to China in December 2024. Chinese chipmakers have consequently relied more heavily on indirect or grey-market supply channels, according to Reuters' sources.
Memory obtained through those routes can cost several times the price paid by buyers outside China. Because HBM represents a large share of an accelerator card's total manufacturing cost, the premium flows directly into the finished product.
Export controls therefore affect more than access to Nvidia processors. A domestically designed accelerator still needs memory, advanced packaging, interconnects and manufacturing equipment. Restricting one essential component can limit the performance or volume of the entire system.
Huawei's Ascend 950 strategy
Huawei has said the Ascend 950DT will become available in the fourth quarter of 2026. The card integrates an AI processor, memory and supporting components and is intended mainly for model development and generating responses.
The Ascend 950PR is designed to process user requests before they move deeper into a model-serving system. Huawei says the 950 family will use two proprietary HBM technologies, HiBL 1.0 for the 950PR and HiZQ 2.0 for the 950DT, but has not publicly detailed the manufacturing source.
Prices for older Huawei accelerators are also rising. That suggests customers are competing for available computing capacity across generations rather than simply waiting for the newest product.
Impact on Chinese cloud and AI companies
More expensive accelerators increase the cost of building data centres and training models. Large firms can absorb some of that pressure, but startups may need to rent computing time, reduce experiment size or raise more capital.
ByteDance is receiving increased shipments from Iluvatar CoreX, Reuters reported. The supplier has doubled GPU shipments to the TikTok developer to around 100,000 units in 2026 and redirected hardware originally reserved for internal use.
That allocation illustrates a shortage economy: major customers with scale and strategic importance can secure supply, while smaller buyers face delays or higher prices.
Does this help Nvidia?
Nvidia remains the global AI accelerator leader, but US controls restrict access to its most advanced products in China. Higher domestic prices could make permitted Nvidia models relatively more attractive where available, although compliance rules and performance requirements limit substitution.
Chinese firms also have strategic reasons to keep adopting local hardware. Domestic systems reduce long-term exposure to changing export policy and allow software teams to build expertise around Chinese platforms. A temporary cost disadvantage may be accepted if supply independence is the larger objective.
What happens next
The decisive variable is HBM production. More output from established suppliers, successful Chinese memory alternatives or improved accelerator designs could ease the pressure. Continued global AI demand and tighter export enforcement would push in the opposite direction.
Buyers should watch delivered system cost rather than headline chip price alone. Electricity, networking, cooling, software compatibility and utilisation can matter more over the life of a cluster than the purchase price of one accelerator.
The shortage shows how the AI race depends on a tightly connected semiconductor supply chain. Processor architecture receives most of the attention, but memory may determine how many usable systems can actually be built. Follow related developments in MatchUpWorld's Technology section.
