China's AI Chipmakers Raised Prices Because Memory Ran Out

Reuters reported on September 10 that China's domestic AI chipmakers have raised prices by 20% to 50% from what they were quoting customers two months ago. The reason given is specific and unglamorous: a worldwide shortage of high-bandwidth memory.

The numbers, where they are quotable:

• Huawei Ascend 950DT — now indicated above 250,000 yuan (about $37,255) per accelerator card, up 20% to 50% from two-month-old quotes. Huawei has said publicly the 950DT arrives in Q4 2026.
• Ascend 950PR — roughly 60,000 yuan per card at the start of the year, now above 80,000 yuan, up about 30%.
• Ascend 910C boards — from about 90,000 yuan at the start of the year to above 110,000 yuan.
• Cambricon repriced its next-generation part, tentatively the 690, at 20% to 30% above two-month-old levels.
• MetaX and Iluvatar CoreX made similar increases.

This is a supply story wearing a pricing story's clothes, and it says three things worth separating.

One: the constraint is memory, not lithography

The entire Western framing of Chinese AI silicon has been about process nodes — SMIC's 7nm yields, the absence of EUV, how many generations behind the logic is. That framing has always been half the picture, and this pricing round is the half nobody discusses becoming visible.

AI accelerators are memory-bandwidth-bound in most real workloads. A 7nm die with enough HBM outperforms a 5nm die that is starved. Which means the binding constraint on Chinese AI compute is not the transistor — it is the stack of DRAM sitting next to it, and that stack comes from a global market with three real suppliers, all of whom are sold out and all of whom are reallocating capacity toward HBM4 for Nvidia and AMD.

Export controls restricted advanced HBM to China specifically because it is the component that turns a functional domestic accelerator into a competitive one. The controls worked in the narrow sense. What they did not anticipate is that the same shortage would price Western buyers out too.

Two: this is the other end of the CXMT story

Two weeks ago the memory news out of China was CXMT sampling HBM3E to Alibaba's T-Head and to Cambricon roughly a year ahead of forecast, with risk production started and volume targeted for as early as 2027.

At the time that read as ambition. Read alongside these price quotes, it reads as necessity. Cambricon is not qualifying domestic HBM3E because it prefers local supply on principle. It is qualifying domestic HBM3E because the alternative is repricing its flagship part 30% upward and telling customers to wait.

The two stories are one story: China's accelerator industry has a memory problem severe enough to reprice the market, and the domestic fix is in qualification but not in volume. 2027 is when those two curves are supposed to cross. Everything before that is paid for in yuan.

Three: Iluvatar's volume is the number that matters

Buried in the same reporting is a detail with more forward signal than any price: Iluvatar CoreX doubled its GPU shipments to ByteDance to 100,000 units this year.

That is not a shortage story. That is a substitution story. ByteDance is one of the largest compute buyers on earth and has spent years routing around Nvidia restrictions through gray channels, offshore capacity, and cloud arbitrage. A hundred thousand units from a second-tier domestic vendor means the qualification work is done and the software stack is tolerable enough for production inference.

Prices go up when buyers cannot walk away. Chinese hyperscalers have largely stopped being able to walk away from domestic silicon, because the Nvidia parts they are permitted to buy have been narrowed to the point where the comparison is no longer close. Scarcity pricing on Ascend is what a captive market looks like.

The uncomfortable symmetry

Beijing mandated domestic AI chips for state data centers earlier this year. The stated goal was independence from Nvidia. The unstated cost is that independence from Nvidia does not mean independence from Samsung, SK hynix, and Micron — because the memory is upstream of everybody.

A Chinese AI lab in September 2026 is paying 30% to 50% more for domestic compute than it budgeted two months ago, on a schedule where its best accelerator does not ship until Q4, using memory it cannot make at volume until 2027 at the earliest. That is not a story about export controls failing. It is a story about a supply chain with one genuine chokepoint, and everyone — American labs included — standing in the same line for it.

What to watch

Whether the increases stick. Quoted prices in a shortage are not transacted prices. If HBM supply loosens in Q1 2027 as new capacity lands, these numbers retreat and this becomes a footnote about a tight quarter.

CXMT's HBM3E yields. This is the real variable. Credible volume yields in 2027 decouple Chinese accelerator pricing from the global memory market for the first time. Stalled yields mean another year of paying scarcity rent.

Whether Ascend 950DT actually ships in Q4. A part quoted at 250,000 yuan that slips to 2027 is a very different market from one that arrives on time. Huawei has hit its stated dates more often than its critics expect, and missed them more often than its customers would like.