MMarket PulseScope依頼 / お問い合わせ
All market news
AI HARDWARE

HBM Squeeze Pushes Chinese AI Accelerator Quotes Up by as Much as 50%

Huawei has raised the indicative price of its Ascend 950DT accelerator card above RMB 250,000, 20-50% more than quotes two months earlier, and Cambricon repriced its next-generation 690 by 20-30%, as scarce high-bandwidth memory and grey-market sourcing costs feed into AI hardware bills of materials.

Huawei has lifted the indicative price of its Ascend 950DT accelerator card to more than RMB 250,000, according to people familiar with the quotes, which is 20-50% above the level quoted to customers roughly two months earlier depending on contract terms. Bloomberg separately reported suggested pricing up about 60% over three months to around the same level, with one source saying the quote was set in line with Nvidia's B200 for comparable multi-card deals. Huawei has publicly said the 950DT, which uses the company's proprietary HiZQ 2.0 memory technology, remains scheduled for the fourth quarter of 2026. Older boards are not exempt: the Ascend 950PR has moved above RMB 80,000 from about RMB 60,000 at the start of 2026, roughly 30%, and the Ascend 910C board has climbed from about RMB 90,000 to more than RMB 110,000.

The repricing extends across China's domestic AI chip vendors. Cambricon has repriced its next-generation part, tentatively called the 690, by 20-30% above the reference quotes it gave two months ago, although it has not formally launched the product and told investors it has issued no price announcement. Smaller rivals MetaX and Iluvatar CoreX have made similar increases. Bloomberg and Reuters both note that list economics still vary substantially with order size and how much networking, boards and other auxiliaries a customer buys alongside the accelerators.

Memory is the binding constraint. Since Washington tightened export controls on certain advanced HBM products to China in December 2024, Chinese chipmakers have increasingly turned to grey-market channels where HBM typically costs several times what buyers outside China pay, and because memory accounts for a large share of an accelerator's production cost, those premiums flow straight into finished card prices. Supply is also being reallocated: Iluvatar CoreX has roughly doubled its GPU shipments to ByteDance this year to about 100,000 units, diverting chips originally intended for its own use, and Huawei remains ByteDance's largest domestic supplier ahead of Cambricon and Iluvatar. TrendForce, citing Seoul Economic Daily, reported that CXMT has begun small-scale production of HBM3E and plans to expand output significantly in 2027, with Cambricon and Alibaba's T-Head said to be testing it.

Demand keeps running ahead of what infrastructure can absorb. TrendForce's 7 September report warned that AI server shipments are lifting data centre power demand faster than the US grid can deliver, with the gap between shipment-implied demand and grid-deliverable capacity widening from 2026 and far wider by 2028, and deferral risk concentrated in PJM, ERCOT and MISO. On the demand side, DeepSeek is reported to plan to deploy at least 160,000 Ascend 950DT chips at a data centre under construction in Inner Mongolia, primarily for inference — roughly a full year's initial output for the part, which is understood to be constrained to the low hundreds of thousands of units this year.

Sources