Chinese AI chips jump up to 50% on HBM shortage

Huawei and Cambricon have raised AI accelerator prices by 20% to 50% in two months, and the cause is memory. Reuters reported on September 10 that a worldwide high-bandwidth memory shortage is pushing up the cost of Chinese AI silicon faster than anything happening on the logic side. It is the clearest public price signal yet that the constraint in AI hardware moved off the GPU die and onto the memory stacked next to it.

What actually happened

Huawei repriced the Ascend 950DT accelerator card above 250,000 yuan — about $37,255 — a jump of 20% to 50% over quotes given to customers two months earlier, depending on contract terms. The Ascend 950PR went from roughly 60,000 yuan to more than 80,000. The older 910C climbed from about 90,000 yuan to more than 110,000.

Cambricon repriced its next-generation part, tentatively the 690, at 20% to 30% above its earlier indications. MetaX and Iluvatar CoreX are in the same market with the same input costs.

The mechanism is straightforward. HBM is supplied almost entirely by SK Hynix, Samsung, and Micron. Washington tightened export controls on advanced HBM to China in December 2024, so Chinese buyers work through grey-market channels at several times the price the rest of the world pays. Memory is a large share of an accelerator's bill of materials, so that cost lands directly on the finished card.

Why an AI hardware shortage matters for your business

Read past the China framing. Export controls explain why the increase shows up first and worst in Chinese pricing. They do not explain the shortage, which is global. Everyone building inference capacity is bidding for the same HBM out of the same three suppliers, and that is upstream of every token you buy.

We covered Positron's $875 million round for inference silicon that routes around HBM two days ago. That is a $5 billion bet on exactly this constraint. When a component shortage is severe enough to fund an architectural detour, it is not a quarter-long blip.

So do the unglamorous work now. Instrument tokens per completed task — not per request, per task — so you know what a 20% input-cost increase does to your margin before your vendor tells you. Turn on prompt caching if your provider offers it; most teams leave a large discount on the table because nobody checked. Look hard at whether your agent needs a frontier model for every hop, because a lot of orchestration burns premium tokens on "yes" and "confirm the order."

And keep an open-weights path warm. Not because you want to run inference yourself — you almost certainly don't — but because a 7B model on commodity hardware handling your classification and routing steps is a real hedge, and it takes a weekend to prove out. The teams that get squeezed by hardware pricing are the ones with a single provider and no measurement.

Sources: Reuters exclusive via Investing.com: China's AI chipmakers raise prices as high-bandwidth memory shortage bites.