Chip prices usually fall as production scales. In China's domestic AI accelerator market in 2026, they are doing the opposite. According to a Reuters report on September 10, 2026, Huawei, Cambricon, MetaX and Iluvatar CoreX have all raised prices on current and next-generation AI processors, with some increases running from 20% to as much as 50% over quotes given just two months earlier. The cause is not weak manufacturing or thin demand. It is a shortage of the one component that has become the true bottleneck of the AI era: high-bandwidth memory.
The numbers behind the increase
The specifics show how broad the move is. Huawei's forthcoming Ascend 950DT, expected in the fourth quarter of 2026, is now quoted above 250,000 yuan, roughly 37,255 dollars, a jump of 20% to 50% depending on contract terms. The Ascend 950PR, which sold for about 60,000 yuan at the start of the year, now fetches more than 80,000, a rise of around 30%. Even the older Ascend 910C has climbed from about 90,000 yuan to over 110,000. Cambricon has repriced its next-generation 690 part 20% to 30% above where it stood two months ago.
How far China's AI chip prices moved in 2026
Reported increases on domestic accelerators as the high-bandwidth-memory squeeze feeds into sticker prices.
Now above 250,000 yuan (about $37,255); due Q4 2026
About 60,000 yuan at the start of the year, now over 80,000
About 90,000 yuan at the start of the year, now over 110,000
Next-generation part; exact price not disclosed
Ranges shown at their upper bound for the bar length. Figures as reported by Reuters, September 2026.
Rising prices on last year's silicon, not just on unreleased flagships, is the tell. When a chip that has been shipping for months gets more expensive, the pressure is not coming from a new design or a fab upgrade. It is coming from an input cost that every one of these products shares.
Why high-bandwidth memory is the choke point
Modern AI accelerators are memory machines as much as compute machines. Moving the enormous parameter sets of large models in and out of the processor fast enough is what high-bandwidth memory, or HBM, exists to do, and it accounts for a large share of an accelerator's bill of materials. If HBM gets scarce or expensive, the whole chip does too, no matter how efficient the logic die is.
For Chinese chipmakers, HBM scarcity is compounded by policy. Washington tightened export controls on advanced HBM products bound for China in December 2024, cutting off the straightforward supply route. Since then, Chinese buyers have leaned on grey-market channels, where the same memory reportedly costs several times what buyers outside China pay. So domestic accelerators carry a double penalty: they need a lot of HBM, and the HBM they can get is far more expensive than it is for competitors abroad. The price increases are that penalty passed through to customers.
Export controls were meant to slow China's access to compute. One second-order effect is that China's own compute is now more expensive to build.
On the economics of the HBM squeeze
The policy paradox
This is where the story becomes genuinely interesting rather than just a set of price tags. Export controls on advanced memory were designed to constrain the amount of frontier AI compute China can field. In the near term, they are working in a specific and perhaps unintended way: not by making the chips unavailable, but by making them costlier to produce, which raises the price of every training and inference run built on them.
The honest reading is that this cuts both ways, and it is too early to call the net effect. Higher input costs do slow deployment at the margin and squeeze the economics of Chinese AI labs, which is aligned with the controls' intent. But scarcity and a price umbrella are also exactly the conditions that pull domestic investment into the constrained component. Demand is not going away: Iluvatar CoreX reportedly doubled its planned GPU shipments to ByteDance to 100,000 units for 2026. When a bottleneck is this lucrative and this strategic, the strong incentive is to build a way around it at home, and China has moved before to localize supply chains it was cut off from.
What it means beyond China
For the rest of the market, the episode is a reminder that HBM, not logic, is the pacing item for AI hardware everywhere. The same memory crunch driving Chinese prices up is the reason memory makers globally are enjoying a supercycle and why accelerator supply stays tight far beyond China's borders. Compute cost is increasingly a memory story, and memory is concentrated among a handful of suppliers.
That concentration has a practical implication for anyone whose plans depend on AI compute. When the cost and availability of the underlying hardware can swing on a policy change or a supply shock in a single component, the ability to move workloads across providers, regions and model options is not a nicety, it is insulation. The same model-agnostic discipline that guards against vendor lock-in at the software layer, the approach platforms like Metir take by keeping work portable across providers rather than tied to one, is ultimately a hedge against volatility that starts all the way down at the memory die. You cannot control the price of HBM. You can control how exposed a single supplier's constraints leave you.
China's rising AI chip prices are not a sign of failure. They are a sign of a market straining against a real physical and geopolitical bottleneck, and paying up to keep building anyway. The number worth watching next is not the sticker price. It is how quickly domestic HBM capacity comes online to relieve it.
Sources:
- Exclusive: China's AI chipmakers raise prices as high-bandwidth memory shortage bites (Reuters via Yahoo Finance, Sept 10, 2026)
- China's AI chipmakers raise prices as high-bandwidth memory shortage bites (Reuters via Investing.com)
- Huawei Lifted Prices for Its Best AI Chip This Summer (Bloomberg, Sept 10, 2026)
- China's AI Chipmakers Raise Prices as HBM Shortage Drives Up Costs (ITP.net)
Image credits
- Hero: 12-inch silicon wafer with integrated-circuit dies. Wikimedia Commons, licensed CC BY-SA 3.0.
- In-body: Technician holding a silicon wafer in a semiconductor cleanroom. Wikimedia Commons, public domain.