China AI Chipmakers Raised Prices Up to 50%. The Memory Is the Bottleneck, Not the Logic.

Huawei's most advanced AI accelerator is not shipping yet, and it has already become up to 50 percent more expensive. Reuters reported on Sept. 10 that the indicated price of the Ascend 950DT — the card Beijing is counting on to displace Nvidia inside Chinese data centers — now sits above 250,000 yuan, roughly $37,255, a jump of 20 to 50 percent from what customers were quoted two months ago, depending on contract terms. Huawei says the part arrives in the fourth quarter of 2026. The repricing has almost nothing to do with the logic die. It is about the memory stacked beside it.

Huawei is not alone. Cambricon has repriced its next-generation chip, tentatively called the 690, at 20 to 30 percent above the levels it indicated two months earlier, according to three people familiar with the pricing discussions who spoke to Reuters. Smaller rivals MetaX and Iluvatar CoreX have pushed through similar increases. Older silicon is moving too: the Ascend 950PR, roughly 60,000 yuan per card in January, now fetches more than 80,000 yuan, up about 30 percent, while the previous-generation Ascend 910C board has climbed past 110,000 yuan from about 90,000. None of the four chipmakers responded to requests for comment.

The common input is high-bandwidth memory, the vertically stacked DRAM that lets an AI processor reach enormous volumes of data at speed. HBM is a large share of an accelerator's bill of materials, and the advanced end of that market belongs to three companies: South Korea's SK Hynix and Samsung Electronics, and US-based Micron Technology. Since Washington tightened controls on exports of certain advanced HBM products to China in December 2024, Chinese chipmakers have leaned increasingly on grey-market channels, where the sources said the same parts cost several times what buyers outside China pay. That premium now flows directly into finished cards.

It is landing on top of a genuine global squeeze rather than instead of one. SK Hynix has told investors its HBM, DRAM and NAND capacity is essentially sold out for 2026, and Micron exited consumer memory entirely to concentrate on AI and enterprise buyers. Wonjin Lee, president and head of global marketing at Samsung, put the year plainly in January: “In 2026, there's going to be issues around semiconductor supplies, and it's going to affect everyone, not just Samsung,” he told Bloomberg. “I think it's an industry-wide reality that we're going to see some supply issues.”

Analysts describe a market that has lost its usual dampeners. “Memory manufacturers once functioned as shock absorbers for the tech ecosystem, using scale, inventory discipline, and long-term contracts to provide pricing and supply predictability,” said Manish Rawat, semiconductor analyst at TechInsights. “Samsung's inability to cushion volatility despite its unmatched capacity indicates a market in disequilibrium.” Gartner has forecast a 47 percent rise in DRAM prices in 2026; Meritz Securities analyst Kim Sunwoo estimates suppliers are meeting only 75 to 80 percent of DRAM demand in the second half of this year, possibly falling toward 60 percent in 2027. HBM consumes roughly three times the wafer capacity of standard DRAM per gigabyte, by Micron's own accounting, which is why every additional AI accelerator cannibalises the commodity market behind it.

Memory, not logic, is the binding constraint

For three years the consensus framing of US export controls was about logic: lithography, transistor density, how far behind SMIC's 7-nanometer process ran Nvidia's TSMC-fabricated parts. Chinese designers have partially routed around that. They cannot route around bandwidth the same way. A shortfall in raw compute can be papered over with more dies and rack-scale architecture — Huawei's SuperPoD systems are an explicit bet on that substitution. Memory bandwidth and capacity per package sit much closer to hard physical ceilings, and modern inference economics scale with them rather than with peak flops. Serving a long-context model is a memory problem wearing a compute costume.

Huawei has said the Ascend 950 series will use two proprietary memory technologies — HiBL 1.0 on the 950PR, specified at 128GB and 1.6 TB/s, and HiZQ 2.0 on the 950DT, at 144GB and 4 TB/s — without detailing where or how that memory is made. That omission is the story. Designing an HBM interface is tractable; yielding tall stacks with through-silicon vias, fine interconnect, warpage control and thermal management is the part the rest of the industry took a decade to learn.

What the December 2024 controls actually achieved, then, is narrower and stranger than halting Chinese AI compute. Domestic vendors captured the roughly $50 billion Chinese AI chip market that opened when Nvidia's top parts were restricted, and volumes are still climbing: Iluvatar CoreX has doubled GPU shipments to ByteDance to 100,000 units this year, diverting parts earmarked for internal use to do it. The controls did not cap output. They moved the chokepoint downstream and imposed a tax — collected largely by grey-market intermediaries rather than by Samsung, SK Hynix or Micron, and paid by the domestic champions Beijing is subsidising into existence. China's AI buildout is not so much throttled as made structurally more expensive than everyone else's.

The escape hatch is domestic. ChangXin Memory Technologies began risk production of HBM3E in late August, according to The Information — a generation behind the HBM4 now in mass production at the big three, and by most estimates three to five years off the frontier. Alibaba's T-Head and Cambricon are already evaluating the parts, with commercial use possible as early as 2027, while Naura, Maxwell and U-Preseason build the domestic stacking and bonding tools to go with them.

Watch three things. Whether the 950DT ships in volume in the fourth quarter at the quoted 250,000 yuan, or slips. Whether CXMT's HBM3E passes qualification at Cambricon and T-Head, converting a grey-market cost into a domestic supply line inside 12 to 18 months. And whether Chinese buyers keep paying: above 250,000 yuan a card, domestic silicon now competes with the landed cost of smuggled Nvidia parts, and that arbitrage cuts both ways.