Chinese AI chipmakers have started raising prices. The cause is a shortage of high-bandwidth memory (HBM), the stacked memory technology that modern AI accelerators depend on, according to a Reuters exclusive.

The price increases mark a shift for China's domestic chip industry. For two years, local AI chip designers competed mainly on price against restricted access to Nvidia hardware. Now the supply of a critical component, not demand alone, sets the ceiling on what they can ship and what they can charge.

Why does HBM matter for AI chips?

HBM sits next to the compute engine on an AI accelerator. It feeds the processor with training data at speeds conventional memory cannot reach. Large language models live or die by this bandwidth. Without enough HBM, a chip cannot serve today's models, no matter how strong its compute units are.

The memory is hard to make. Only a small group of producers — led by SK Hynix, Samsung and Micron — can manufacture it at scale. Capacity is allocated months in advance, and the largest buyers of AI hardware absorb most of it. That leaves smaller buyers, including Chinese chipmakers, competing for what remains.

What does this change for Chinese AI chip vendors? It converts a demand-side boom into a supply-side squeeze. They can sell more chips than they can build.

How does the shortage reach the customer?

The mechanism is simple. Chipmakers buy HBM on the open market or through contracted suppliers. When supply tightens, memory prices climb. The chipmaker passes that cost into the final accelerator price. The customer — a cloud provider, a data center operator, an AI lab — pays more for the same product.

Reuters reports the price rises as an exclusive finding, without yet naming specific vendors or percentage increases. The direction, however, is clear and it moves against the broader trend of falling compute costs per unit of performance.

Who feels the impact first?

The pressure lands unevenly across the market:

• Domestic chip designers face higher input costs and longer lead times for memory.
• Chinese cloud and AI customers see accelerator prices rise at the moment they scale up model training.
• Global memory makers gain pricing power, because every AI chip shipped needs their product.
• Overseas buyers of Chinese chips may see quotations revised upward on existing procurement talks.

The shortage also interacts with export controls. Washington has restricted China's access to advanced AI chips and to some memory technologies. A tight global HBM market compounds those policy limits with a plain commercial one: even unrestricted memory costs more when everyone wants it.

What comes next?

Memory producers have announced capacity expansions, but new HBM lines take quarters, not weeks, to ramp. Until supply catches up, chipmakers that secured memory contracts early hold an advantage. Those buying on the spot market carry the heaviest cost burden.

For buyers, the practical question is timing. Contracts signed now lock in higher prices. Waiting risks paying more later if the shortage deepens. For China's chip industry, the episode shows a structural weakness: the country can design competitive AI silicon, but it still depends on a short list of foreign suppliers for the memory that makes those chips useful.

The price increases reported by Reuters are the first visible symptom. If HBM stays scarce, they will not be the last.