NextFin News - Huawei Technologies raised the price of its most advanced artificial-intelligence chip by as much as 60% this summer, a striking reversal for the company at the center of Beijing's drive to build a domestic alternative to Nvidia. The increase, which also lifted prices across Huawei's older Ascend boards and forced rival Chinese designers to follow suit, is being driven less by chip design than by a global shortage of high-bandwidth memory that has pushed the cost of a critical component several times higher for Chinese buyers. The episode lays bare a contradiction at the heart of China's semiconductor self-reliance campaign: U.S. export controls have handed domestic chipmakers a protected market, but the same controls have made the components needed to compete at the frontier far more expensive to obtain.
The price increases, chip by chip
Huawei has raised the indicated price of its Ascend 950DT accelerator card to more than 250,000 yuan, about $37,255, according to people familiar with the matter. That represents an increase of between 20% and 50% from prices quoted to customers just two months ago, depending on contract terms. The 950DT is Huawei's most advanced AI chip and is scheduled to become available in the fourth quarter of 2026.
The pricing pressure is not confined to the flagship. Huawei's Ascend 950PR, which entered mass production in the first quarter of this year, sold for about 60,000 yuan per card at the start of 2026 and now fetches more than 80,000 yuan — an increase of roughly 30%. The older Ascend 910C board has risen to more than 110,000 yuan from about 90,000 yuan at the beginning of the year. Beijing-based Cambricon Technologies has raised the indicative price of its next-generation processor, tentatively known as the 690, by between 20% and 30% compared with levels indicated two months ago, and smaller rivals MetaX and Iluvatar CoreX have made similar increases, the sources said.
The increases illustrate a major challenge for China's AI industry as Beijing pushes domestic companies to reduce their reliance on Nvidia and develop home-grown alternatives. Even as Chinese chipmakers seek to expand production and capture a larger share of the country's roughly $50 billion AI chip market, they remain exposed to a critical component whose supply is heavily concentrated outside China.
High-bandwidth memory, or HBM, consists of vertically stacked memory chips that allow AI processors to access large amounts of data at high speed. It is a crucial component in AI computing, particularly as increasingly demanding applications require processors to handle vast quantities of information rapidly. The advanced HBM market is dominated by South Korea's SK Hynix and Samsung Electronics, as well as U.S.-based Micron Technology. Since Washington tightened export controls on certain advanced HBM products to China in December 2024, Chinese chipmakers have increasingly relied on grey-market channels to obtain supplies, according to the sources. Those supplies typically cost several times more than the prices paid by buyers outside China. Because memory represents a substantial portion of an AI accelerator's production cost, the increase is flowing directly into the price of finished accelerator cards.
Huawei, Cambricon, MetaX and Iluvatar CoreX did not respond to requests for comment.
The bottleneck moved from lithography to memory
For years, the story of China's AI chip ambitions was told in terms of lithography: without access to extreme-ultraviolet machines from ASML, Chinese foundries could not produce the most advanced logic, and domestic accelerators would remain a generation behind. That narrative is being rewritten. Chinese designers have learned to work around the logic constraint — Huawei's Ascend 910C achieves performance comparable to Nvidia's H100 by combining two 910B processors into a single package through advanced integration techniques, according to people familiar with its design, and the 950 series uses proprietary low-precision formats and custom memory interfaces to compensate for a less advanced process node.
The binding constraint has shifted one layer up the stack, to the memory that sits beside the logic die. HBM is not merely more expensive than conventional DRAM; it is structurally scarce. Samsung, SK Hynix and Micron have reallocated wafer capacity toward HBM because it carries substantially higher margins, and the transition from HBM3E to HBM4 will increase the number of DRAM dies required per AI accelerator from 12 to 16 per stack — a 33% rise in die consumption per chip even as overall production capacity grows only incrementally. For Chinese buyers, the problem is compounded: since December 2024, access to the most advanced HBM products has been restricted, forcing purchases through alternative channels at a large premium.
The mechanism is straightforward and unforgiving. Memory represents a substantial portion of an AI accelerator's production cost. When the price of that memory rises severalfold for one set of buyers, the cost increase flows directly into the price of the finished card. Huawei's own product strategy acknowledges the pressure: the Ascend 950PR was designed around HiBL 1.0, Huawei's proprietary "low-cost" HBM, which rotating chairman Eric Xu described as more cost-effective than HBM3E and HBM4E. Even so, the 950PR's price has risen roughly 30% this year. The company has not detailed where or how the memory for its proprietary HBM technologies is manufactured.
Demand is not the problem — supply is
The price increases are arriving at a moment of unusually strong demand for domestic silicon. Huawei expects revenue from its AI chips to surge at least 60% this year, to about $12 billion from $7.5 billion in 2025, driven by strong demand from Chinese companies looking for local alternatives to Nvidia, whose sales are constrained by regulatory hurdles in China. Major Chinese technology companies, including Alibaba, ByteDance, Tencent and DeepSeek, are driving the surge in demand for Huawei's Ascend chips.
The order books show how concentrated that demand has become. DeepSeek plans to deploy at least 160,000 of Huawei's top accelerators at a massive data center it is building in Inner Mongolia, which could create one of the largest known clusters of Huawei AI chips. At a street price near 111,000 yuan, about $16,000, per chip, the order carries a face value of roughly $2.6 billion. Huawei is ByteDance's largest domestic chip supplier, followed by Cambricon and Iluvatar CoreX, according to people familiar with the supply chain. Iluvatar CoreX has doubled its shipments of GPUs to ByteDance to 100,000 units this year and has diverted GPUs originally intended for its own internal use to help meet ByteDance's requirements — a telling sign that computing constraints have intensified enough to force suppliers to cannibalize their own capacity.
On the supply side, Huawei plans to make about 600,000 of its Ascend 910C chips in 2026, roughly double this year's output, people familiar with the matter have said. The company will increase its Ascend product line in 2026 to as many as 1.6 million dies. But those projections include dies that Huawei has in inventory as well as internal estimates of yields, and they depend on access to HBM. Domestic production offers only partial relief: CXMT, China's top DRAM supplier, is estimated to be able to produce approximately 7 million HBM3 dies in 2026 — sufficient for about 600,000 AI chips of comparable performance to Nvidia's H100 accelerator, assuming each chip uses eight HBM stacks and a 70% yield. CXMT has likely acquired enough equipment for HBM production through 2026 or 2027, after which it will encounter obstacles both in ramping up volumes for current HBM generations and in developing more advanced ones.
Cyclical shortage, structural squeeze
The critical analytical question is whether this is a cyclical price spike that will fade as memory capacity comes online, or a structural change in the cost base of China's AI industry. The answer is both, and the distinction matters for how the story resolves.
The HBM shortage itself has cyclical characteristics. Memory markets have historically oscillated between shortage and glut, and SK Hynix, Samsung and Micron are making massive capital investments to expand HBM production. Micron raised its 2026 capital expenditure to more than $25 billion, and SK Hynix has committed over $30 billion to new advanced packaging and fabrication plants. If those investments land on schedule, the global HBM balance should ease, and the grey-market premium that Chinese buyers pay should compress.
But the structural component will not self-correct. The premium that Chinese chipmakers pay for HBM is a direct function of export controls, and there is no evidence those controls will be relaxed. Even if global HBM supply normalizes, Chinese buyers will remain a separate, higher-cost market so long as they are barred from purchasing advanced memory through official channels. That means the cost base of China's domestic AI accelerator industry is likely to remain structurally elevated relative to Western competitors — a permanent tax on the very self-reliance campaign the controls were meant to slow.
This is the second-order effect that the market has not fully priced. The conventional reading of the price increases is bullish for Huawei and its domestic rivals: pricing power in a protected market, rising revenue, and share gains against a constrained Nvidia. That reading is not wrong, but it is incomplete. Higher chip prices raise the cost of every data center that Chinese cloud providers, model developers and enterprises are trying to build. Beijing wants Chinese companies to replace Nvidia processors, but more expensive HBM raises the price of the very domestic accelerators intended to fill that gap. The substitution is real — but it is expensive, and the bill is being paid by the same industry the state is trying to accelerate.
The counter-thesis
The strongest argument against this reading is that China's domestic chip ecosystem is learning fast enough to outgrow the memory constraint. Huawei's Pangu Ultra MoE, a 718-billion-parameter model, was trained on more than 6,000 Ascend NPUs, demonstrating that the domestic stack can work at frontier scale. Huawei, Cambricon, Hygon and Moore Threads all adapted DeepSeek V4 on launch day in April 2026 — previously an achievement only Nvidia could claim. And Huawei's own fabs, staffed by Huawei employees and equipped with more than $9 billion of tools purchased for reverse-engineering and replication, could eventually exceed SMIC's production, freeing SMIC capacity for other partners.
Paul Triolo, a partner at consulting firm Albright Stonebridge Group, captured this view when he said the U.S. Commerce Department's export curbs on Nvidia's H20 "will mean that Huawei's Ascend 910C GPU will now become the hardware of choice for (Chinese) AI model developers and for deploying inference capacity."
In this telling, the memory premium is a temporary friction that a determined, well-capitalized national champion can absorb and eventually engineer around.
The falsifying signal is specific: if CXMT's domestic HBM output reaches more than roughly 14 million dies in 2027 — enough to supply about 1.2 million AI chips at eight stacks per chip — and the grey-market premium on advanced HBM for Chinese buyers compresses to below 50%, then the structural-cost thesis is wrong and the price increases should be read as a cyclical spike. Until then, the elevated cost base is the baseline, not the exception.
What to watch
The immediate beneficiaries of the price increases are clear: Huawei, Cambricon and the other domestic designers are capturing pricing power in a market where demand vastly exceeds supply. But the exposed parties are equally clear — Chinese cloud providers, AI startups and state-backed computing projects that must now pay more for every unit of domestic compute they deploy. The DeepSeek order alone, at a face value of roughly $2.6 billion, is a statement of commitment to domestic silicon; it is also a statement of how expensive that commitment has become.
In the short term, expect continued upward pressure on prices for both current and next-generation Chinese accelerators as the HBM shortage persists through 2026. Huawei's Ascend 950 series is due to become fully available in the fourth quarter, while Cambricon's 690 has not yet formally launched — and both will carry the higher memory cost into the market. In the medium term, the question is whether domestic HBM production can close the gap; CXMT began small-scale production of HBM3E in late August and is testing the memory with Alibaba's chip unit T-Head and Cambricon, but small-scale output is not yet proof of scale. In the long term, the structural question is whether China can build a memory ecosystem that is not subject to foreign controls — and that is a project measured in years, not quarters.
The scenarios are stark. In the base case, HBM supply remains tight through 2026, Chinese AI chip prices stay elevated, and domestic substitution continues at a higher cost — Huawei's AI chip revenue reaches its $12 billion target, but each yuan of revenue buys less compute than it would have a year ago. In the upside case for China, CXMT ramps faster than expected and the grey-market premium compresses, allowing domestic prices to stabilize even as volumes grow. In the downside case, the memory bottleneck tightens further, order fulfillment slips, and the cost of China's AI buildout rises faster than the capacity it purchases.
The takeaway is uncomfortable for both sides of the chip war. For Washington, the price increases show that export controls are raising the cost of China's AI advance — but they are also accelerating the very substitution they were meant to delay. For Beijing, the lesson is that self-reliance has a price tag, and that tag is being revised upward. The race for AI supremacy is no longer just about who can design the best chip. It is about who can secure the memory to put inside it.
Explore more exclusive insights at nextfin.ai.
