Huawei’s AI chip ambitions are often framed as a manufacturing race. According to SemiAnalysis, that picture misses the most important pressure point: high-bandwidth memory, or HBM.
The report describes a supply chain that can still produce large numbers of Ascend AI accelerators, supported by stockpiled dies, expanding domestic capacity and workarounds for export controls. But it also argues that memory supply, not logic chip production, is the constraint most likely to shape how far Huawei can scale.
Why the chip numbers need context
SemiAnalysis estimates that Huawei shipped about 507,000 Ascend AI accelerators in 2024. It projects 805,000 units for 2025. On their own, those figures suggest a fast-moving production push.
The report says the picture is more fragile because Huawei is relying on a large inventory of Ascend dies made at TSMC. SemiAnalysis puts that stockpile at over 2.9 million Ascend dies, saying it has helped Huawei get around export controls and sustain production through 2024 and 2025.
That stockpile is not permanent. Analysts cited in the source expect the die bank to run out within nine months. If that happens, Huawei’s ability to keep increasing output depends more heavily on domestic manufacturing and on access to the memory needed to complete finished AI accelerators.
The company is also trying to tighten control over its own supply chain. Huawei has launched a chipmaking tools outfit called SiCarrier and is expanding its own factories. SemiAnalysis says that by next year, those plants could rival or even surpass the capacity of SMIC, China’s largest contract chipmaker.
SMIC may not remain the main bottleneck
Huawei currently relies on SMIC for outsourced chip production, and that has been a limiting factor. SemiAnalysis expects the situation to shift by year’s end as SMIC expands advanced process capacity at 7nm and below.
The targets cited in the report are substantial: 45,000 wafers per month by the end of 2025 and 80,000 per month by 2027. SemiAnalysis says even a portion of that capacity would be enough for millions of Ascend chips each year.
Several factors help explain why SMIC can focus more attention on AI chips. The source says other Chinese firms still use TSMC for mobile chips, reducing pressure on SMIC. It also says Huawei continues to obtain advanced datacenter CPUs and networking components made at TSMC through shell companies.
Another factor is timing. The report points to slow adoption of US controls by Japan and the Netherlands, which allows Chinese companies to stock up on equipment. Together, those details support SemiAnalysis’s broader argument: logic chip manufacturing is important, but it is no longer the only or even the clearest constraint.
HBM is the harder problem
SemiAnalysis identifies HBM supply as China’s real weak spot in AI hardware. The report says Beijing’s trade priorities are revealing: it is pushing for relaxed HBM restrictions, not for lithography tools or TSMC access.
China has already tried to reduce the risk by building inventory. SemiAnalysis says China stockpiled 13 million HBM stacks, including 11.4 million from Samsung. Samsung alone reportedly shipped 7 million stacks in the month between the announcement and the start of US restrictions beyond HBM2E in December 2024.
That inventory matters because it can support finished chips. The report says the stockpile is enough for around 1.6 million Ascend 910C chips. But SemiAnalysis expects it to be gone by year-end.
After that, the numbers become much less favorable for Huawei. Without foreign HBM, the report says Huawei will not be able to make even a million Ascend chips in 2025 because domestic supply cannot fill the gap.
China’s top DRAM maker, CXMT, is moving quickly but remains far short of demand in the report’s forecast. SemiAnalysis projects CXMT will produce just 2 million HBM stacks next year, enough for only 250,000 to 300,000 Ascend 910C chips.
- Ascend shipments: about 507,000 in 2024, with 805,000 projected for 2025.
- Ascend die inventory: over 2.9 million dies made at TSMC.
- HBM stockpile: 13 million stacks, including 11.4 million from Samsung.
- CXMT forecast: 2 million HBM stacks next year, supporting 250,000 to 300,000 Ascend 910C chips.
Nvidia and the strategic tradeoff
The source also places Huawei’s position inside a wider US-China compute strategy. The US government’s move to let Nvidia export its H20 chip to China is described as part of a more complex approach than a simple ban.
SemiAnalysis suspects China’s public lack of interest in the H20 is a tactic aimed at winning approval for a more powerful Blackwell-series chip, the B30A. The report also points to weak performance from models like Deepseek on Huawei hardware as evidence that Chinese companies still depend on Western chips.
That dependence creates a dilemma for the US. Selling modified chips to China can keep Chinese companies tied to US technology and slow domestic alternatives. At the same time, those sales can accelerate China’s AI progress.
The report suggests a conditional approach: approve more powerful chips only once China can make an equivalent at scale. Until then, Chinese companies such as Bytedance must rely on cloud providers in places like Malaysia for top GPUs, access the US can cut off at any time.
For Nvidia, the stakes are also commercial. SemiAnalysis says sales of modified chips to China could bring in as much as $30 billion a year. That revenue sits alongside the broader strategic question of whether controlled access slows China’s independence or speeds up its AI development.
What the bottleneck reveals
The central lesson from the report is that AI chip capacity is not a single number. Huawei can expand factories, use SMIC, draw down TSMC-made dies and keep pushing Ascend production. But without enough HBM, those pieces do not translate cleanly into unlimited AI accelerator output.
That is why HBM restrictions appear to carry unusual force. SemiAnalysis argues they are effective and recommends adding CXMT to the sanctions list as well. The recommendation follows directly from the report’s view that domestic HBM supply is the gap China cannot yet close at the scale Huawei needs.
For Huawei, the near term may still look strong because existing inventories are doing real work. The harder test comes when the stockpiles run lower and domestic memory production has to carry more of the load.