- China’s AI chip prices rise as HBM supply tightens.
- US controls limit China’s HBM access as CXMT expands output.
Chinese AI chipmakers are raising prices for current and upcoming processors as a shortage of high-bandwidth memory (HBM) pushes up production costs.
Huawei has lifted the indicated price of its upcoming Ascend 950DT accelerator card to more than 250,000 yuan (US$37,255), between 20% and 50% above quotes given to customers two months earlier, according to Reuters. Cambricon has raised the indicated price of its next-generation 690 processor by between 20% and 30%, while MetaX and Iluvatar CoreX have also increased prices.
Memory becomes the constraint
TrendForce said DRAM supply is expected to remain tight through 2027, while uncertainty around production and validation of next-generation HBM4E is already affecting planned AI accelerator configurations. Nvidia began evaluating lower-HBM alternatives for its Rubin Ultra platform during the third quarter of 2026, including configurations using fewer memory layers and HBM4 rather than HBM4E.
Several cloud service providers are also evaluating lower HBM capacities for their next-generation in-house AI accelerators, according to TrendForce. The research firm said memory shortages had already led cloud providers and server manufacturers to reduce RDIMM capacities in some server configurations during the first half of the year.
TrendForce expects HBM bit shipments to increase by 50% to 60% year on year in 2027, but said that growth would still be insufficient to keep pace with demand. The firm expects HBM suppliers to retain pricing power under those conditions.
TrendForce reported on September 7 that inventories among DRAM suppliers remained at historically low levels in the second quarter of 2026. Additional supply was mainly being allocated to servers, while overall DRAM bit shipments increased only modestly during the quarter.
DRAM industry revenue reached US$154.73 billion during the quarter, up 59.5% from the previous quarter. TrendForce also reported higher shipments of HBM3E, LPDDR5X, and high-capacity RDIMMs.
The supply pressure extends beyond HBM. Server DRAM revenue reached US$75.58 billion in the second quarter, up 53% quarter on quarter, while average contract prices increased by between 53% and 58%, according to TrendForce.
Memory manufacturers are allocating more production capacity to HBM, restricting growth in the supply of RDIMMs used in servers. TrendForce expects server DRAM shortages and higher prices to continue through 2027 as manufacturers prioritise HBM and higher-capacity server products.
SK hynix said in July that it had finalised long-term agreements with around 10 customers and was continuing discussions with major customers on multi-year contracts. The company began mass shipments of HBM4 during the second quarter and plans to increase production in the second half of 2026.
Samsung said it began mass production of HBM4 and shipped commercial products to customers in February. Its HBM4 uses sixth-generation 10-nanometre-class DRAM and a 4nm logic base die.
TrendForce estimates that HBM and RDIMM will account for 51% of global DRAM bit supply in 2026. It projects DRAM and NAND flash will rise from 47% of capital expenditure among major cloud service providers in 2026 to 68% in 2027.
China faces tighter HBM access
The US introduced additional controls on exports of advanced HBM to China in December 2024. The Bureau of Industry and Security said the measures cover US-origin HBM as well as certain foreign-produced HBM subject to US export rules.
BIS added HBM-specific controls alongside new restrictions on semiconductor manufacturing equipment and software. It described HBM as critical to AI training and inference at scale and as a key component of advanced computing chips.
Chinese AI chip companies have increasingly sourced restricted HBM through grey-market channels following the controls, Reuters reported, with some supplies costing several times the prices available to buyers outside China. The higher procurement costs have contributed to the latest accelerator price increases.
Huawei is also increasing memory capacity and bandwidth across its Ascend processor roadmap. The company has developed two proprietary HBM products for its Ascend 950 series, with different memory configurations for separate AI workloads.
The Ascend 950PR uses Huawei’s HiBL 1.0 HBM and is designed for the prefill stage of inference and recommendation workloads. Huawei said these workloads require more parallel computing but less memory-access bandwidth than model training and the decode stage of inference.
Huawei said HiBL 1.0 is more cost-effective than HBM3E and HBM4E. The Ascend 950PR and 950DT use different HBM configurations for workloads with different memory-bandwidth requirements.
The Ascend 950DT, scheduled for availability in the fourth quarter of 2026, is designed for model training and the decode stage of inference. Its HiZQ 2.0 memory provides 144GB of capacity and 4TB per second of memory-access bandwidth, while the chip provides 2TB per second of interconnect bandwidth.
Huawei plans to increase those memory specifications alongside processor performance in subsequent generations. The company said the planned Ascend 960 will double the Ascend 950 series’ computing power, memory capacity, memory-access bandwidth, and number of interconnect ports.
China expands domestic HBM production
China is developing domestic HBM capacity alongside its efforts to produce more AI processors locally. CXMT, China’s leading DRAM chipmaker, held about 7.7% of the global DRAM market in 2025, according to Reuters.
CXMT’s HBM programme remains at an earlier stage than its established DRAM business. TrendForce said in January that CXMT was focusing on HBM3-generation products and advanced packaging but continued to face technical barriers and was still in early verification stages.
Reuters reported on August 31 that CXMT had begun small-scale production of HBM3E. Alibaba’s T-Head and Cambricon are testing the memory with their processors, while CXMT plans to expand production in 2027.
CXMT has progressed further in other categories of memory. The company said on September 7 that it had entered mass production of LPDDR6, with the memory first being used in Xiaomi’s 18 Fold smartphone. The product follows CXMT’s previous work on DDR5 and LPDDR5X.
TrendForce said US export controls announced in January 2025 affect CXMT’s process technology, wafer input, yield rates, and market deployment. The restrictions cover all of the company’s process nodes and have affected its development plans for 2025 and 2026.
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
The post China’s AI chip push runs into a costly problem: memory appeared first on TechWire Asia.