Huawei’s Ascend Production Ramp: TSMC Die Banks Help, HBM Remains the Bottleneck

Compute capacity has become a central resource in artificial intelligence. The United States currently leads global deployed compute, with more than 70% of deployed FLOPs, while export controls are intended to limit China’s access to advanced hardware.

Those restrictions have encouraged Chinese companies to adapt. Techniques such as high batch sizes and disaggregated serving have helped offset hardware constraints, although much of DeepSeek’s inference still runs on Western hardware. The use of Huawei chips also delayed training of DeepSeek’s next-generation model.

China’s long-term objective is greater control over its domestic compute infrastructure. Huawei is central to that strategy because its semiconductor ecosystem spans chip design, manufacturing, packaging, networking, and related tools.

Huawei’s vertically integrated chip strategy

Huawei aims to control more of the complete manufacturing chain for its Ascend accelerators. That includes the logic die, memory, and advanced packaging. The company has also established SiCarrier, a tool company intended to replicate foreign semiconductor manufacturing equipment, and has purchased more than $9 billion of equipment for its fabs and reverse-engineering efforts.

SiCarrier recently raised $2.8 billion. The funding is being used to build fabs dedicated to Huawei and staffed by Huawei employees. Reporting on these facilities suggests that their combined production could exceed SMIC’s output by next year. As Huawei brings more capacity online, SMIC could allocate additional capacity to other Chinese chip designers, including Cambricon, whose accelerators are used by Chinese companies such as ByteDance.

Huawei’s own fabs would increase China’s production capacity while giving Huawei greater control over process development, yields, and manufacturing iteration. Huawei and SMIC could work together on yield improvement, research for future process nodes, and broader semiconductor manufacturing capabilities.

At present, high-volume production of Huawei chips is outsourced to SMIC, China’s leading pure-play foundry. This includes the Ascend accelerator family and Kirin mobile processors. SMIC’s 7 nm-class processes have experienced low yields because of process immaturity, export controls, and the difficulty of manufacturing large dies such as Ascend chips. As a result, only a limited portion of SMIC’s capacity has been assigned to Ascend, while smaller mobile processors remain easier to manufacture profitably.

Huawei production and SMIC’s expansion

The available estimates indicate that Huawei shipped 507,000 Ascend units in 2024, most of them Ascend 910B chips. Shipments are projected to reach 805,000 units this year, including 653,000 Ascend 910C units. The 910C is the more advanced version, and the estimate includes dies produced by both TSMC and SMIC.

SMIC initially struggled to ramp production because of export controls. During that period, Huawei obtained Ascend dies fabricated at TSMC. The source analysis estimates that Huawei received more than 2.9 million Ascend dies, which can be used in both 910B and 910C products.

This inventory, described as a foreign “die bank,” supports Huawei’s Ascend production through 2024 and 2025. Without it, Huawei’s shipment totals would be considerably lower.

The TSMC die bank is expected to run out within nine months. SMIC, however, is expected to have enough capacity to produce meaningful Ascend volumes by the end of the year. Under a baseline scenario, producing millions of Ascend dies per month would require no more than 20,000 wafers per month (wspm) of SMIC capacity.

For comparison, a conservative estimate places SMIC’s total advanced-node capacity, covering 7 nm and below, at:

45,000 wspm by the end of 2025
60,000 wspm in 2026
80,000 wspm in 2027

Huawei is also building fabs of its own, some of which are not subject to the same export controls. Its collaboration with SMIC on process technology could allow advanced-node production to increase further.

If all advanced-node capacity were assigned to Ascend dies, annual production could reach tens of millions of units. The estimates use conservative assumptions for yield and yield improvement. Higher yields would allow SMIC to produce several million Ascend dies while using a smaller share of its total capacity.

SMIC may need only a single-digit percentage of its capacity to produce more than one million Ascend dies as soon as early next year. Producing several million would require additional allocation, but not necessarily a complete conversion of the foundry’s output.

Huawei is also expanding its own manufacturing network. Its 2024 expenditure on wafer fabrication equipment reached $7.3 billion, up 27% year over year. The company moved from effectively zero spending in 2022 to becoming the world’s fourth-largest wafer-fabrication-equipment customer within two years. Since then, its fab construction and network of semiconductor-related companies have continued to expand.

TSMC access and export-control timing

Huawei produces many mobile chips, but other Chinese companies, including Oppo and Xiaomi, continue to fabricate mobile system-on-chips at TSMC. Their access to TSMC can reduce pressure on SMIC to manufacture mobile processors, leaving more SMIC capacity available for AI accelerators.

SMIC’s expansion has also benefited from differences in export-control timing. The United States often coordinates equipment restrictions with Japan and the Netherlands, but matching controls may not take effect immediately. In some cases, the delay is six months or longer. Chinese companies can use that period to place large equipment orders and build inventories.

Some Japanese and Dutch equipment vendors have continued selling into China during these gaps. Certain semiconductor suppliers now receive more than 40% of their revenue from China. Re-export routes can also allow restricted equipment to reach China through third countries when matching controls do not cover re-exporting.

Equipment restrictions based on “advanced ICs” also leave gaps. ASML’s NXT:1980 scanners are capable of producing 7 nm-class logic and may support more advanced production when throughput and yield requirements are reduced or subsidized. These systems remain permitted in China and at some SMIC facilities.

The analysis argues that export controls would be more effective with faster coordination among allies, stronger re-export rules, and coverage of semiconductor subsystems. Restrictions would also need to address not only CXMT, but the OSATs and subsidiaries that can perform critical manufacturing steps outside the main company.

Memory is particularly important because it is the main constraint on Huawei’s accelerator output.

HBM is the main bottleneck

The source analysis identifies high-bandwidth memory, or HBM, as the primary bottleneck for Huawei’s Ascend production. China has reportedly asked US officials to relax HBM controls during trade discussions, without making comparable requests for greater TSMC access or lithography equipment.

Huawei was able to build an HBM inventory in addition to its stockpile of TSMC logic dies. Samsung, which had struggled to establish a strong position in the Western accelerator supply chain, sold memory to Chinese customers that ultimately directed the inventory to Huawei.

The estimates attribute 11.4 million HBM stacks directly to Samsung, including 7 million stacks shipped during the one-month period between the announcement and enforcement of controls. Including other suppliers and shipment routes, China is estimated to have obtained 13 million HBM stacks.

On December 2, 2024, the Bureau of Industry and Security announced controls covering HBM more advanced than HBM2E. Full compliance was required by December 31, 2024. Samsung exported a large amount of HBM to China during that quarter, and the shipment is estimated to account for most of China’s HBM inventory.

HBM continued to reach China after the rules took effect. Companies including CoAsia Electronics and Faraday were reported to have supplied non-functional chips containing HBM. That particular flow is believed to have declined, although other sources of HBM smuggling may remain.

The estimated 13 million HBM stacks are sufficient for approximately 1.6 million Ascend 910C packages. Even so, the analysis expects China to become HBM-constrained by the end of the year as foreign inventory declines.

China may have enough TSMC and SMIC capacity to produce more than 805,000 Huawei Ascends this year, but the number of completed accelerators will be limited by the available HBM. SMIC’s projected output includes approximately 1 million 910Cs and nearly 500,000 910Bs, although not all of those dies can be assembled into finished accelerators without sufficient memory.

Without additional foreign HBM, Huawei may be unable to manufacture even one million Ascend chips next year. Slower GDDR and LPDDR memory could be used instead, but that would not be suitable for leading language models using modern reinforcement learning or for large-scale inference deployments.

China’s domestic HBM effort: CXMT

CXMT is China’s main DRAM manufacturer. It has advanced rapidly through domestic engineering, recruitment of engineers from Samsung, SK Hynix, and Micron, and technical support from equipment vendors including Applied Materials, Lam Research, and Tokyo Electron.

CXMT can ship DDR5 memory, reportedly only a few years behind SK Hynix, Micron, and Samsung. The company is reducing profitable DDR4 production for PCs and mobile devices while pursuing a more aggressive HBM roadmap.

CXMT’s capacity is expected to approach Micron’s by next year, although the ramp is not expected to arrive soon enough to prevent a near-term reduction in Huawei Ascend production. The estimates project CXMT at 257,000 wafers per month in 2026, equivalent to just under 15% of global DRAM production, rising to 490,000 wafers per month by 2030.

China’s Big Fund III, established in May 2024, invested $2 billion in CXMT. The company is also expanding in Shanghai and Beijing, with HBM packaging subsidiaries being established in Shanghai.

CXMT stockpiled equipment during 2024 in anticipation of possible US entity-listing action and may still be adding HBM-specific tools. Some advanced tools, such as Hanmi’s thermal-compression-bonding systems for HBM3, are restricted, but older systems can be operated more slowly without imposing prohibitive costs.

CXMT can also obtain leading-edge equipment for through-silicon-via formation from Japanese suppliers. TSV processing is essential for HBM production. Chinese OSAT companies such as JCET and Tong Fu are developing capacity for TSV and stacking, allowing them to package CXMT’s front-end HBM wafers.

Because these processes can be distributed among multiple companies, controls focused only on CXMT may leave significant gaps. The broader manufacturing ecosystem would also need to be covered.

CXMT production scenarios

CXMT currently produces slightly more than 250,000 wafers per month and is expected to reach 300,000 wafers per month by the end of the year. It has not yet installed all the equipment required to convert standard DRAM lines to HBM production, but such a conversion is expected.

The amount of HBM CXMT can produce will depend on how much wafer capacity is assigned to it. China could potentially produce tens of millions of HBM stacks using less than half of CXMT’s capacity, although sanctions could slow the conversion of existing DRAM lines.

The analysis estimates that CXMT will produce approximately 2 million HBM stacks next year, enough for only 250,000 to 300,000 Ascend 910C accelerators. Yield improvement and capacity conversion are expected to take time.

The outcome could change if CXMT continues to stockpile equipment or improves yields more quickly. The estimates are described as conservative, and CXMT may be able to produce HBM3e in 2026.

If all available leading-edge logic dies could be paired with HBM, Huawei’s Ascend output would rise from 805,000 units this year to 1.175 million in 2025. Under a scenario with substantially more domestic HBM, annual output could rise from approximately 300,000 to more than 5 million Ascend 910C units in the following year.

The analysis concludes that export controls have constrained China’s chip production. Without additional HBM smuggling, China is expected to produce fewer Ascend accelerators next year rather than more, because CXMT’s HBM output will not yet match available logic-die capacity.

Networking and datacenter CPUs

AI accelerators operate as parts of larger systems. Clusters can contain tens of thousands of interconnected chips, making networking hardware and system design as important as the accelerator microarchitecture.

Huawei’s CloudMatrix 384, or CM384, is one example of this system-level approach. The source analysis estimates that its scale-up switches are being manufactured at TSMC rather than SMIC through shell companies, with some equipment being stockpiled.

Huawei is also believed to have manufactured datacenter CPUs at TSMC. The analysis identifies this as evidence that existing controls on Huawei’s TSMC access leave gaps. Although the Ascend accelerator uses a 7 nm-class process, Huawei may have obtained 5 nm technology through insufficient screening.

Producing networking and CPU chips at TSMC reduces pressure on SMIC and leaves more capacity for Ascend and Cambricon accelerators, subject to HBM availability. China has an incentive to move as much production as possible to TSMC, and existing compliance frameworks may not be sufficient to prevent the use of intermediary companies.

Nvidia, Huawei, and China-specific GPUs

The US government previously prohibited Nvidia’s H20 from being shipped to China, but later indicated that Nvidia could receive an export license. The H20 has more memory than the banned H100, although it provides fewer FLOPs. Its larger memory capacity is useful for inference.

Nvidia is positioned to sell existing H20 and H20E inventory, with total shipments potentially reaching several hundred thousand chips and billions of dollars in revenue.

Reports have also described a more advanced Blackwell-based chip for China. The proposed B30A could provide more than 10 times the FLOPs of the H20, while offering roughly half the price and half the performance of the B300. Purchasing two B30A chips could therefore provide comparable performance to one B300, depending on workload scaling.

Reports have linked the B30A proposal to weaker interest in the H20 among Chinese customers. However, the analysis attributes that reported decline partly to government pressure rather than to a lack of Chinese demand for foreign compute.

The H20 and H20E have substantially better software support than the Ascend 910C, although their overall hardware capabilities may be closer. The B30A, by contrast, would occupy a higher performance class. Its availability would give Chinese companies more capacity for inference and model-serving workloads.

US policy therefore faces a balance between allowing Chinese companies to remain on the US AI stack and limiting the development of China’s domestic alternatives. The analysis argues that the performance and volume of chips shipped to China should be monitored alongside the growth of Huawei, Cambricon, and CXMT.

Implications for AI development

Additional H20 or H20E shipments would provide more compute for companies such as DeepSeek and Alibaba to use in reinforcement learning. Much of the compute used for reinforcement learning is inference, an area where the H20 and H20E are suitable. A Blackwell-based chip would provide greater performance.

DeepSeek has continued improving its models, but limited compute has slowed work on a planned multimodal V4 model. Alibaba’s Qwen and Moonshot’s Kimi K2 have also concentrated mainly on text models while facing compute limitations. Additional Blackwell shipments could accelerate their development.

More compute would also improve the speed and availability of model serving in China. DeepSeek has deliberately served R1 at low speeds to conserve compute, demonstrating how hardware shortages can affect user experience and adoption.

Models do not need to be trained and served on the same type of accelerator. Claude 4, for example, used GPUs and TPUs during different stages of research and training, while inference is available on Nvidia GPUs, Google TPUs, and Amazon Trainiums. Chinese companies are expected to continue training primarily on Nvidia hardware even if Huawei’s domestic accelerators become more widely available.

China’s compute position

Approval of the H20 would materially increase China’s access to FLOPs and memory. The Ascend 910C is expected to become China’s first significant domestically produced accelerator in terms of realized FLOPs and memory. Unauthorized re-exporting could also provide China with meaningful volumes of H100 chips and smaller quantities of B200 chips.

A Blackwell-based China-specific chip would increase China’s available FLOPs further because of its substantially higher compute capability.

Chinese companies, including ByteDance, can also rent advanced compute outside China. Providers such as Oracle and Google offer access to Blackwell GPUs in countries including Malaysia. ByteDance’s Seed models were trained in the United States on a US cloud platform.

Cloud rental allows Chinese companies to use Nvidia hardware without placing the chips inside China. Physical inspection is also easier in locations such as Malaysia than in China. However, rented capacity still enables Chinese companies to train advanced models and potentially serve them to users. The difference is that access to rented infrastructure can be withdrawn.

China is unlikely to send large volumes of its data outside the country, so domestic AI infrastructure remains necessary. Foreign chip shipments may provide a temporary buffer, but they do not remove China’s long-term objective of silicon self-sufficiency.

China’s pursuit of semiconductor sovereignty predates US export controls. As Huawei’s foreign HBM inventory declines, domestic production will become increasingly important. Future decisions about Blackwell shipments therefore depend partly on the pace at which Huawei, Cambricon, and CXMT expand their own capabilities.

Huawei’s Ascend Production Ramp: TSMC Die Banks Help, HBM Remains the Bottleneck