China’s attempt to replace Nvidia hardware is running into a costly constraint that lies beside, rather than inside, the accelerator: high-bandwidth memory. Reuters reports that people familiar with nonpublic pricing discussions said Huawei and Cambricon had lifted indicated prices for current and planned AI processors as HBM became more expensive and harder to obtain. The report matters because China’s domestic-computing strategy is often described in terms of processors, software, and server clusters. Memory availability can determine whether those processors are affordable enough to deploy at scale.
The clearest reported change concerns Huawei’s Ascend 950DT accelerator card. Reuters said the indicated price had risen above 250,000 yuan, or about US$37,255, putting it 20% to 50% above quotations given roughly two months earlier, depending on contract terms. The price was described as indicative rather than a public list price, and Huawei has not confirmed it. Huawei has said publicly that the 950 series is due to become available in the fourth quarter of 2026.
Cambricon’s proposed next-generation 690 processor was also cited as 20% to 30% above the level indicated two months earlier. Reuters noted that Cambricon has not formally launched the 690 and that its 590 remains its current flagship. A separate report carried by Sina said Cambricon rejected claims that it had issued price-increase notices, calling such reports market rumors. The distinction is important: the Reuters account concerns prices discussed with customers, while the company response contests the idea of a public, formal notice.
HBM Has Become the Cost Center for China AI Chips
HBM stacks memory chips vertically so an accelerator can move large quantities of data quickly. That makes it central to training and serving large models, but it also makes memory a substantial share of an AI card’s cost. Reuters said the worldwide HBM market remains dominated by SK Hynix, Samsung Electronics, and Micron Technology. Washington tightened controls on exports of some advanced HBM products to China in December 2024, leaving Chinese buyers more dependent on expensive gray-market channels, according to the report’s sources.
The result is an unusually direct transmission from supply-chain friction to computing prices. A chip designer can improve an accelerator’s architecture, but it cannot simply substitute low-bandwidth memory without changing the performance profile of the system. Huawei has disclosed that its Ascend 950 series will use proprietary memory technologies called HiBL 1.0 for the 950PR and HiZQ 2.0 for the 950DT. It has not publicly detailed where that memory is made or how it is sourced. That disclosure leaves a significant gap between an announced technical architecture and a fully transparent supply chain.
China’s domestic memory efforts are therefore part of the AI-computing story, not a side issue. EastFrontier recently examined CXMT’s move into HBM3E for domestic AI-chip designers, while another report tracked Huawei’s effort to deepen its software position through AscendNPU IR for China’s AI compiler ecosystem. The new pricing evidence shows why both memory and software matter. A domestic accelerator must be available, usable, and economically viable.
Huawei and Cambricon Prices Signal a Wider Infrastructure Problem
Reuters also reported increases for older Huawei products. The Ascend 950PR was said to have moved from roughly 60,000 yuan at the start of the year to more than 80,000 yuan, while the Ascend 910C board reportedly rose from about 90,000 yuan to more than 110,000 yuan. These figures were attributed to people familiar with the prices, not to a Huawei price list, and should be treated accordingly. Still, the pattern suggests the pressure is not confined to one unreleased flagship.
The timing is uncomfortable for Chinese cloud providers and model developers that are trying to add domestic capacity after restrictions limited access to Nvidia’s most advanced products. Higher card prices do not stop deployments, but they change how many cards a fixed budget can purchase and can raise the cost of training, inference, or both. The impact matters most for organizations planning dense clusters rather than isolated servers.
JD Cloud and Moore Threads, for example, said they planned a 100,000-GPU domestic cluster. That proposal does not depend on Huawei hardware alone, and it is a plan rather than an operating facility. Yet it illustrates the scale at which a higher per-card memory burden can multiply. When a system needs tens of thousands of accelerators, memory costs are no longer a procurement detail. They become a strategic input into the economics of Chinese AI infrastructure.
Domestic Alternatives Must Compete on More Than Chip Design
The Reuters report named MetaX and Iluvatar CoreX as smaller rivals that had also made similar indicated increases. It further said Iluvatar CoreX had doubled GPU shipments to ByteDance to 100,000 units this year, citing one source, and had redirected hardware previously intended for internal use to meet ByteDance demand. Neither company responded to Reuters’ requests for comment. The reported allocation change, if sustained, would show how buyers and suppliers are reallocating scarce domestic compute as demand intensifies.
This pressure does not mean China’s domestic accelerator effort has failed. It instead exposes the layers required to make it durable. Chip designers need processors; system builders need boards, interconnects, and software; data centers need power and racks; and all of them need memory in sufficient quantity and at a cost that permits customers to build. Export restrictions have helped create demand for local accelerators, but they have also made the surrounding memory supply chain more consequential.
For investors and enterprise buyers, the immediate question is not whether every reported quote becomes a standard price. The question is whether HBM scarcity remains a temporary premium or becomes a persistent tax on China’s domestic AI buildout. The answer will shape the cost of expanding clusters, the competitiveness of locally designed accelerators, and the urgency of China’s effort to make advanced AI memory at home.