China’s HBM Shortage Is Raising AI Chip Prices—and Slowing the Nvidia Replacement Push

High-bandwidth memory has become a central choke point in China’s AI-accelerator market. A global shortage is increasing component costs for domestic chipmakers, while U.S. controls introduced in December 2024 restrict access to covered advanced HBM for China. The result is higher quoted prices, tighter availability and a more difficult path to replacing Nvidia systems with locally made alternatives. 11819

The reported price increases

Reuters reported that Huawei lifted the indicated price of its Ascend 950DT accelerator card to more than RMB 250,000 (about $37,255). That was 20% to 50% higher than quotes given to customers two months earlier, depending on contract terms. 119

Cambricon reportedly increased the indicated price of its next-generation accelerator, tentatively called the 690, by 20% to 30% over the level indicated two months earlier. The available reporting does not establish a reliable absolute RMB price for that product. 12

Reuters also reported price increases at MetaX and Iluvatar CoreX, alongside Huawei and Cambricon, as HBM costs rose and supply tightened. Public reporting supplied here does not provide a dependable product-by-product price schedule for MetaX or Iluvatar. 119

Some secondary reports also put Huawei’s Ascend 950PR at more than RMB 80,000, versus roughly RMB 60,000 earlier in 2026, and its Ascend 910C at more than RMB 110,000, versus roughly RMB 90,000 at the start of the year. Those figures imply increases of about 30% and 22%, respectively, but they are less well substantiated in the supplied primary-source excerpts than the 950DT and Cambricon 690 figures. They should be treated as reported market indications, not confirmed list prices. 91213

Why export controls make China’s shortage more acute

On December 2, 2024, the U.S. Commerce Department announced new controls on HBM, describing it as critical to AI training and inference at scale. The controls cover U.S.-origin HBM and certain foreign-produced HBM subject to the advanced-computing Foreign Direct Product rule; specified provisions had a delayed compliance date of December 31, 2024. 18

Congressional Research Service material describes the package as adding China-wide controls on HBM, DRAM and advanced-packaging semiconductor-manufacturing equipment. 17

Reuters reported that Chinese chipmakers increasingly relied on grey-market channels after the restrictions tightened. That may secure some supply, but it also adds cost and uncertainty. The supplied evidence supports the existence of this reliance; it does not establish a precise average grey-market premium or total volumes purchased. 19

HBM changes both price and allocation

HBM is not a peripheral input: it is a key component of advanced AI accelerators. When supplies are scarce, chipmakers must either obtain memory at higher cost, restrict shipments, redesign around available memory, or prioritize selected products and customers. 18

That helps explain why a domestic accelerator can still be difficult to deploy at scale even when the processor itself is available. Higher memory costs raise the delivered system price, while constrained HBM availability can reduce the number of boards that can actually be built and shipped. Reuters described the price hikes as evidence that the worldwide HBM shortage is reverberating through China’s effort to deploy home-grown alternatives to Nvidia. 119

The available sources do not substantiate a specific allocation share, shipment volume or priority arrangement in which Iluvatar CoreX reserved a defined amount of supply for ByteDance. That claim should not be treated as established on the evidence available here.

Huawei’s Ascend roadmap: what is known

Huawei planned to ship roughly 750,000 Ascend 950PR units in 2026, according to Reuters sources. Samples were sent to customers in January, mass production was expected to begin the following month, and full shipments were expected in the second half of 2026. Reuters reported a roughly RMB 50,000 version using traditional DDR memory and a faster-HBM version at roughly RMB 70,000. 33

For the higher-end Ascend 950DT, reports place commercial availability in the fourth quarter of 2026. Huawei has said the 950DT is designed to use its HiZQ 2.0 memory technology, with 144 GB of memory and 4 TB/s of bandwidth. 3435

Even with that roadmap, scaling supply is the harder task. Reuters reported that constraints in high-end domestic AI chips were expected to persist during production ramp-up, with broader availability improving only once Ascend 950 supernodes began shipping at scale. 32

What the shortage means for China’s Nvidia-replacement effort

China’s domestic AI strategy is therefore a systems-supply challenge, not just a logic-chip challenge. Accelerators need memory, packaging, manufacturing capacity and qualified supply chains. HBM scarcity increases the bill of materials and can cap the volume of deployable hardware, making domestic alternatives less economical or less available for cloud providers and AI developers. 11819

The pressure also creates a powerful incentive to develop domestic memory capability. That could strengthen Chinese suppliers over the longer term, but it does not immediately remove the manufacturing and qualification constraints that make advanced HBM difficult to scale.

For Korean memory leaders, scarcity reinforces the strategic importance of HBM capacity. SK Hynix is Nvidia’s main HBM supplier and, according to Reuters citing Counterpoint, held a 57% share of the HBM market. 40 But sustained high prices and China’s need for alternatives also create incentives for new competitors and localized supply chains.

How long could the squeeze last?

There is no precise, source-backed expiry date for the HBM shortage in the materials available. Capacity additions take time, and the outlook remains contingent on AI demand, production yields and how quickly new facilities and packaging capacity come online.

The strongest supported indication of duration is broader rather than HBM-specific: SK Group Chairman Chey Tae-won said in March 2026 that the global chip-wafer shortage could persist until 2030, because AI-driven demand continued to outpace supply. He specifically linked HBM production to substantial wafer consumption. 40

That should not be read as a firm forecast that HBM will remain equally constrained until 2030. It does, however, support the conclusion that the current supply pressure is structural rather than a short-lived disruption—and that memory access will remain a decisive variable in the competitiveness of Chinese AI accelerators for years to come.