HBM Shortage Drives Up Costs as Huawei, Cambricon and Other Chinese AI Chip Makers Raise Prices Sharply
nashnova research
Huawei, Cambricon and other Chinese AI chip makers have hiked prices 20%–50% in two months as soaring high-bandwidth memory costs — inflated by grey-market procurement — erode the price advantage of domestic chips positioned as Nvidia alternatives.
How much have prices risen — and which chips are affected?
Huawei's flagship Ascend 950DT accelerator now carries an indicative price above ¥250,000, up 20%–50% from two months ago depending on contract terms; the card is slated for Q4 2026.
Cambricon's next-generation 690 chip (working name) rose 20%–30% over the same period; smaller rivals MetaX and Iluvatar CoreX made comparable increases.
This means → the hikes are not one company's pricing decision but a supply-chain-wide pressure hitting every domestic AI chip vendor at once.
Why are older chips also getting more expensive?
Huawei's Ascend 950PR climbed from roughly ¥60,000 per card at the start of the year to over ¥80,000 — about a 30% jump; the earlier 910C rose from around ¥90,000 to above ¥110,000.
In plain terms = older cards also contain HBM; when memory costs rise, finished-product prices follow — regardless of generation.
This reflects a root cause that sits upstream in memory supply, not in chip design.
Why is HBM so expensive — and what is the grey market?
HBM — high-bandwidth memory, layers of storage chips stacked vertically so a processor can access massive data at extreme speed — is a critical component of AI accelerators. Global advanced capacity is concentrated at SK Hynix, Samsung and Micron.
After the US tightened advanced-HBM export controls on China in December 2024, Chinese vendors increasingly rely on grey-market channels where costs run several times the normal price.
This means → memory accounts for a large share of an AI card's total cost; grey-market premiums feed straight through to end-product pricing — vendors are essentially passing procurement costs on to customers.
Can Huawei's in-house HBM solve the problem?
Huawei says its 950 series will use two proprietary HBM technologies: HiBL 1.0 for the 950PR and HiZQ 2.0 for the 950DT.
The company has not disclosed where or how the memory is manufactured.
In plain terms = the self-developed path exists on paper, but whether it can reach volume production at competitive cost remains an open question.
How is the supply crunch reshaping the competitive landscape?
Iluvatar CoreX doubled its GPU shipments to ByteDance this year to 100,000 units and redirected GPUs originally reserved for internal use to meet ByteDance's demand.
Among ByteDance's domestic AI chip suppliers, Huawei ranks first, followed by Cambricon and Iluvatar CoreX.
This reflects how big buyers' bargaining power grows during a price-hike cycle — suppliers will sacrifice their own inventory to hold on to top-tier orders.
What does this mean for China's chip-substitution strategy?
The price hikes come at a critical juncture in China's roughly $50 billion AI chip market: domestic vendors were already benefiting after US controls blocked Nvidia's most advanced processors from entering China.
But persistently rising HBM costs are testing whether the domestic substitution path can sustain enough price-performance appeal.
This means → the real competitor for Chinese chips is not just Nvidia's performance — it is their own supply chain's cost ceiling. If grey-market premiums persist, the economics of "substitution" become harder to justify.
市场有风险，内容仅供研究参考，不构成投资建议。