Chinese AI chip prices surge 50% year-on-year amid HBM memory shortage, constraining domestic accelerator production and limiting deployment scaling in mainland data centers.

China's homegrown AI chipmakers are raising prices across the board, driven by a shortage of high-bandwidth memory (HBM), the specialized stacked memory that powers modern AI accelerators. In September 2026, Huawei, Cambricon, MetaX, and Iluvatar CoreX all repriced their product lines within weeks. The increases range from 20% to more than 50%, hitting at the exact moment Beijing is pushing its chip industry to cut its dependence on Nvidia.

Huawei's Ascend 950DT accelerator card is now quoted at more than 250,000 yuan, or roughly $37,255, according to Reuters—20% to 50% higher than quotes from just two months earlier. Cambricon raised the indicated price of its next-generation 690 processor by 20% to 30%. MetaX and Iluvatar CoreX have also moved prices up, though exact percentages remain undisclosed.

The repricing did not happen all at once. Reuters first reported Huawei's higher Ascend 950DT quotes on September 10, 2026. A follow-up report on September 11 filled in price increases from Cambricon, MetaX, and Iluvatar CoreX, making clear this was an industry-wide move. Intel CEO Lip-Bu Tan added his own memory-constraint warning five days later, on September 16, tying the China-specific story to a global pattern already building for months. Three separate disclosures inside a single week is notable—pricing shifts of this size typically roll out over a full quarter, not days.

HBM sits at the center of nearly every AI accelerator built since 2023. It stacks memory dies vertically and connects them to the processor through a silicon interposer, giving GPUs and NPUs the bandwidth needed to keep pace with massive AI models. Only a handful of companies—chiefly SK Hynix, Samsung, and Micron—can manufacture it at scale, and all three have redirected capacity toward the highest bidders in the Nvidia and AMD supply chains. Chinese chipmakers now compete for whatever HBM output remains, much of it sourced through gray-market channels or domestic substitutes that lag the leading edge.

On September 16, Tan warned that memory constraints could worsen before improving. His comments echoed Intel's earlier observations that DRAM and NAND prices were climbing five to sevenfold across parts of the memory market. The China-specific squeeze on HBM is a subset of that broader crunch, but it hits AI chipmakers harder because HBM has no substitute in accelerator design. A GPU or NPU without enough memory bandwidth simply cannot run large models at competitive speed, so vendors either pay higher prices or ship slower products.

The knock-on effect is showing up in ordinary consumer memory too. Reports place a single-week jump in Chinese DRAM prices at 14%, as AI data center buildouts and HBM production absorb wafer capacity that once went toward standard memory chips. AI's memory appetite is reshaping pricing for products with nothing to do with AI at all.

Four companies make up the core of this story, each responding to the same shortage in slightly different ways. Huawei's Ascend line carries the highest public number. Cambricon's 690 processor increase is smaller in percentage terms but still material for a chip aimed at large-scale training and inference. MetaX and Iluvatar CoreX have both confirmed price increases without disclosing exact figures, signaling how fluid the pricing environment has become.

| Company | Product | Price Change | New Price / Detail |

|---------|---------|---------------|--------------------|

| Huawei | Ascend 950DT accelerator | +20% to +50% | More than 250,000 yuan (~$37,255) |

| Cambricon | 690 processor (next-gen) | +20% to +30% | Not publicly disclosed |

| MetaX | AI processor line | Confirmed increase, % undisclosed | Not publicly disclosed |

| Iluvatar CoreX | AI processor line | Confirmed increase, % undisclosed | Doubling ByteDance shipments to 100,000 units in 2026 |

Iluvatar CoreX's shipment trajectory deserves attention. The company is reportedly on pace to ship 100,000 GPUs to ByteDance in 2026, double its prior volume, even as it raises prices. That combination—higher unit prices alongside rising shipment volume—signals that demand inside China for domestic AI silicon is strong enough to absorb cost increases rather than force customers toward alternatives. ByteDance, the parent company of TikTok, has been one of the most aggressive Chinese buyers of AI compute. Its willingness to keep scaling purchases from a domestic vendor despite higher prices suggests how constrained alternative supply—meaning smuggled or gray-market Nvidia chips—has become.

China's push toward domestic AI silicon accelerated sharply after US export controls tightened access to Nvidia's top-tier accelerators. Huawei's Ascend series, Cambricon's processor line, and newer entrants like MetaX and Iluvatar CoreX all emerged or scaled up as direct responses to that squeeze. The goal was straightforward: build enough local capacity that Chinese cloud providers, social platforms, and state-backed AI labs would not depend on chips that Washington could cut off at any moment.

That strategy worked well enough to create real demand, as evidenced by ByteDance's shipment numbers, but it ran straight into a second bottleneck unrelated to export policy: the physical scarcity of HBM. Unlike logic chips, where China has made real progress domestically through SMIC and other foundries, advanced memory manufacturing remains dominated by SK Hynix, Samsung, and Micron. All three are currently prioritizing shipments to Nvidia, AMD, and hyperscale cloud buyers in the US and elsewhere. China's chip self-sufficiency drive solved the logic problem before it solved the memory problem, and September 2026's price hikes are the clearest evidence yet of that gap.

HBM scarcity has rewritten pricing across the industry before. Nvidia faced its own version of this squeeze, and yield challenges in HBM4 manufacturing have kept supplies tight industry-wide. What's different about the Chinese chipmakers' situation is that they sit lower in the queue than Nvidia when memory suppliers allocate limited HBM output, explaining why their percentage increases are, in several cases, sharper than what Nvidia itself has passed on to customers.

Even after the price increases, Huawei's Ascend 950DT and Cambricon's 690 remain positioned as lower-cost alternatives to Nvidia's export-restricted data center GPUs inside China, though the gap has narrowed. A $37,255 price tag for the Ascend 950DT puts it in a similar range to some configurations of Nvidia's China-specific compliant chips, eroding what used to be a clearer cost advantage for domestic silicon. The performance gap has not closed at the same