Exclusive-China’s AI chipmakers raise prices as high-bandwidth memory shortage bites

Huawei raised the indicated price of its Ascend 950DT AI accelerator to more than 250,000 yuan ($37,255), up 20%-50% from two months earlier, while Cambricon increased indicative pricing for its next-generation 690 chip by 20%-30%. A global high-bandwidth-memory shortage, compounded by U.S. export controls on advanced HBM to China, is pushing up the cost of domestic Nvidia alternatives and constraining Chinese AI-computing expansion. Huawei's 950PR has risen about 30% to more than 80,000 yuan, while its older 910C board increased to above 110,000 yuan from roughly 90,000 yuan; supply shortages have also prompted Iluvatar CoreX to double GPU shipments to ByteDance to 100,000 units this year.

Analysis

The relevant signal is not domestic accelerator pricing power; it is a constrained-memory tax on China AI buildouts. Higher system cost raises the capital intensity and extends the payback period for Chinese cloud and model-training deployments, likely shifting scarce capacity toward hyperscale customers and away from smaller enterprise AI projects over the next 1-3 quarters. That should slow the volume ramp assumed in local-substitution valuations, even if reported accelerator ASPs rise.

For global memory, this reinforces that HBM remains the marginal bottleneck rather than GPU logic. MU is a cleaner listed beneficiary of sustained HBM scarcity than Chinese accelerator vendors, but its upside is principally ex-China: export restrictions mean China-specific shortages do not translate one-for-one into Micron unit sales. SK Hynix has the strongest HBM product position, yet a prolonged China supply workaround market adds policy and customer-concentration risk rather than incremental investable demand.

NVDA has a nuanced read-through. More expensive local alternatives widen the performance-per-dollar and software-stack advantage of legally sellable Nvidia products, making any future licensing flexibility a high-convexity upside catalyst; absent that, the near-term effect is a smaller and more capital-constrained China AI TAM. The contrarian point is that domestic chip makers may lose margin despite higher ASPs if memory procurement costs and allocation uncertainty are rising faster than realized selling prices.

Watch Chinese hyperscaler capex guidance, domestic accelerator delivery lead times, and HBM contract pricing through the next earnings cycle. The bearish China-buildout thesis is falsified if local vendors demonstrate stable gross margins while materially expanding shipments, or if domestic HBM yields improve enough to compress system costs within 6-12 months.

Key Decisions for Investors

• Maintain/accumulate MU on 3-6 month pullbacks rather than chase a China-specific headline: the actionable thesis is global HBM mix and pricing durability. Use a 10-15% downside risk budget; reduce if next earnings guidance implies HBM supply is loosening or DRAM pricing turns negative.
• Avoid adding to Cambricon (688256.SS) solely on higher quoted ASPs over the next 1-3 months. Treat gross-margin disclosure, receivables growth, and confirmed shipment volume as the gating data; price increases without margin expansion would indicate cost pass-through, not improved economics.
• Use a tactical long MU / short KWEB pair over 1-3 months if Chinese cloud capex commentary weakens: memory scarcity supports MU's global HBM economics while higher domestic compute costs pressure the return profile of China internet AI investment. Exit if Chinese hyperscalers reaffirm materially higher capex or policy opens broader access to imported accelerators.
• Keep NVDA as a watch-list upside optionality trade, not a direct response position. Add only on evidence that approved China-specific products are gaining volume or export licensing is easing; otherwise, China substitution friction is more likely to reduce addressable demand than create near-term Nvidia revenue.