China’s domestic AI chip industry exists, in large part, to make compute cheaper and less dependent on foreign suppliers. As of September 10, Reuters reports those same chipmakers have jacked prices up by roughly 20 to 50 percent over quotes given just weeks earlier. Sovereignty, meet sticker shock.
According to that report, Huawei and Cambricon have both sharply raised prices on current and next-generation AI processors, per three people familiar with the matter. Huawei’s Ascend 950DT accelerator card, which bundles an AI processor with memory and other components, now carries an indicated price above 250,000 yuan, or about $37,255. The culprit, according to the reporting, is a shortage of high-bandwidth memory.
Why memory, not logic, is the chokepoint
If you spend your days evaluating AI tools rather than designing silicon, the instinct is to think of an AI chip as a processor and stop there. That instinct is wrong, and this story is the clearest recent illustration of why. HBM is the stacked memory sitting next to the compute die, and it’s what determines how fast you can shovel model weights into the thing that does the math. For inference in particular, memory bandwidth is often the binding constraint, not raw arithmetic throughput.
Which means a memory shortage doesn’t just make chips slightly more expensive. It hits the component that most directly governs how much useful work the chip can do. Huawei has said the Ascend 950 series would use two proprietary HBM technologies, HiBL 1.0 for the 950PR and HiZQ 2.0 for the 950DT, though details on those have not been disclosed. Even a company building its own memory stack apparently can’t route around the broader squeeze.
What this does to the people actually building things
The Reuters reporting notes the increase is affecting the cost of both AI model development and inference. Those are two different pains with two different timelines.
- Training costs are lumpy and visible. You budget a cluster, you buy the hardware, you eat the capex. A 20 to 50 percent jump on accelerator cards shows up as a number someone has to sign off on, and some of those projects simply don’t get signed.
- Inference costs are quiet and permanent. They ride along with every API call, every agent loop, every retry. If serving hardware gets more expensive, that pressure eventually reaches the per-token price, or it gets absorbed by rate limits and quotas that nobody announces in a blog post.
The second one is what I care about, because it’s the one that changes the tools you and I evaluate. Agent frameworks are among the most token-hungry things anyone has built. A single agent run can fan out into dozens of model calls, most of which produce output the user never sees. That architecture only made sense because inference kept getting cheaper. When the hardware underneath starts moving the other way, the economics of “just let the agent think about it for a while” get uglier fast.
The uncomfortable read for the sovereignty story
Domestic chip programs are usually pitched as a way to escape supply constraints. What the price hikes suggest is that swapping out the logic vendor doesn’t get you out from under the memory vendor. You can design your own accelerator and still be standing in the same queue as everyone else for the stacked DRAM that makes it useful.
I want to be careful here, because the facts on the table are limited. We have a Reuters exclusive built on unnamed sources, an indicated price on one Huawei card, and a percentage range for the increases. We do not have official confirmation, volume figures, or details on how long chipmakers expect the shortage to last. Anyone telling you exactly what this means for global AI pricing next quarter is filling in blanks that haven’t been filled in.
What I’d actually watch
If you build on top of models, the useful signals are downstream. Watch whether providers serving Chinese-market inference quietly adjust pricing or tighten limits. Watch whether teams start shipping smaller models tuned for lower memory bandwidth instead of chasing parameter counts. Watch whether “efficient by design” stops being a marketing line and becomes an actual procurement requirement.
The broader lesson is one I keep relearning while testing AI products: the impressive part of the stack is rarely the fragile part. Everyone benchmarks the model. Almost nobody asks what the model is running on, who makes that part, and how many other buyers want it. This week, the answer to that last question got expensive.
🕒 Published: