Home โ€บ Blog โ€บ China's Chip Stack
AI Compute Stack ยท Part 15 ยท Geopolitics

China's Chip Stack: Export Controls, Huawei and the Race for Self-Reliance

Since 2022 the United States has tried to cut China off from the most advanced AI chips, the machines that make them and the memory they need. China's answer has been to build a second, parallel AI stack: Huawei's Ascend accelerators made at SMIC without EUV lithography, home-grown HBM, a new software ecosystem, and labs such as DeepSeek that squeeze more out of less. This guide maps that stack layer by layer, measures where it is strong and where it is stuck, and asks whether the controls are slowing China down or speeding up its independence.

FL
FrontierAGI Team
โš ๏ธ How this was researched This draws on web searches run up to 23 October 2026: the Federal Register and BIS for US rules, Nvidia's 10-K, Huawei's own roadmap announcement, Epoch AI's analysis of Huawei, and Reuters, Caixin, DigiTimes, ChinaTalk and trade-press reporting. Chinese production volumes, yields and HBM progress are estimates or unconfirmed reports; chip specifications are vendor-announced peaks; how the 25% charge on H200 sales works is described differently by different sources. The calculators are illustrative. See Reading the numbers. The article describes policy; it does not take a side on it.

๐Ÿงฑ Two AI stacks

Every earlier part of this series described one global supply chain: Nvidia designs, TSMC fabricates with ASML machines, SK hynix supplies memory. Since October 2022, US export controls have tried to keep China out of the top of that chain. China cannot legally buy Nvidia's Blackwell or Rubin chips, ASML's EUV machines or the latest HBM, and most of its access to TSMC's leading nodes is closed.

Yet Chinese labs keep shipping competitive models, often open-weight, and Huawei has published a roadmap of a new Ascend chip every year to 2028. The result is a second AI stack, less efficient and more expensive at almost every layer, but increasingly complete.

~1.5MEpoch AI's median estimate of Huawei Ascend output in 2026
8,192Ascend chips in one Atlas 950 SuperPoD (Huawei, Q4 2026)
<1%Licensed H200 sales as a share of Nvidia's data-centre revenue last quarter
2028โ€“30When China's EUV prototype is expected to make working chips (reported)

๐Ÿ•ฐ๏ธ Controls and countermoves

โ†”๏ธ Two-lane timeline

US actions on the left, China's responses on the right. Filter by lane.

๐Ÿ’ก How the rules are drawn US controls do not name chips; they set performance thresholds. The January 2026 rule, for example, allows case-by-case licences for chips below a "total processing performance" of 21,000 and memory bandwidth below 6,500 GB/s, "such as the Nvidia H200 or AMD MI325X", with conditions on US supply, security and third-party testing. Anything above, including Blackwell, remains under a presumption of denial. Chipmakers then design "just under the line" products, and regulators move the line.

๐Ÿ—บ๏ธ The Chinese stack, layer by layer

๐Ÿงฉ How self-sufficient is each layer?

Click a layer. Ratings are our qualitative summary of the reporting, as of October 2026.

๐Ÿญ SMIC: advanced chips without EUV

SMIC, China's largest foundry, cannot buy EUV machines. To print features small enough for 7 nm and 5 nm-class chips, it uses older deep-ultraviolet (DUV) immersion scanners and multi-patterning: exposing each critical layer two, three or four times with shifted masks. It works, as Huawei's phone and Ascend chips show, but every extra exposure adds steps, time and chances for defects.

2025 reports put SMIC's 5 nm-class yields at roughly a third of TSMC's on comparable chips, with wafer costs perhaps 50% higher. Epoch AI says the Ascend 950 and 960 compute dies use SMIC's N+3 process (a 5โ€“7 nm-class DUV node), with N+4 expected for the 970, though other reports describe N+3 as still maturing because of a shortage of advanced DUV scanners. Reliable figures for SMIC's advanced-node capacity are not public.

๐Ÿงฎ What multi-patterning costs

๐ŸŽ›๏ธ Cost per working AI die: EUV vs DUV multi-patterning

Standard yield model (good dies = eโˆ’area ร— defect density). All inputs illustrative; defaults loosely follow the 2025 reports.

AI accelerator dies are among the largest chips made
Extra exposures, steps and time

Dies per 300 mm wafer approximated as (ฯ€ยท150ยฒ รท area) โˆ’ (ฯ€ยท300 รท โˆš(2ยทarea)). Real yields also depend on redundancy, binning and packaging losses.

๐Ÿ”ด Huawei Ascend: the national champion

At its September 2025 Connect conference, Huawei published its first multi-year chip roadmap: the Ascend 950PR (Q1 2026, tuned for inference prefill and recommendation), 950DT (Q4 2026, decode and training), 960 (2027) and 970 (2028), each roughly doubling compute, and from the 950 onwards using Huawei's own HBM. Earlier Ascend 910B and 910C chips were made on SMIC's 7 nm process; teardowns of some 910-series parts found TSMC-made dies obtained through intermediaries before TSMC cut off supply.

Huawei's strategy is to win at the system level. If each chip is weaker, connect more of them, with a powerful network. The Atlas 950 SuperPoD joins 8,192 Ascend chips into one machine, versus 72 GPUs in an Nvidia NVL72 rack. Epoch AI estimates median Ascend output of about 1.5 million units in 2026.

โš–๏ธ Chip vs system: closing the gap with scale

Vendor-announced peak FP8 compute (dense). Toggle between per-chip and per-system. Different precision definitions make exact comparisons uncertain.

Ascend 950: ~1 PFLOPS FP8 per chip; Atlas 950 SuperPoD: ~8 EFLOPS FP8 (Huawei, Sep 2025). Nvidia B200: ~4.5 PFLOPS dense FP8; GB200 NVL72: ~360 PFLOPS dense FP8 (Nvidia specifications). Power and space are not shown: the SuperPoD needs far more of both.

๐Ÿ’พ Memory: the binding limit

Analysts increasingly argue that HBM, not logic chips, sets how many accelerators China can build. US controls since December 2024 restrict HBM sales to China. Chinese firms are reported to have stockpiled about 11.4 million HBM stacks from Samsung before the rules tightened, but stockpiles run down.

China's answer is CXMT, its leading DRAM maker. Reports say it aims for HBM3 mass production by the end of 2026, with analysts warning of slippage; an April 2026 report said no mass-production orders had yet been placed. In August 2026 The Information reported that CXMT had begun small-scale HBM3E production, which the company has not confirmed. Huawei says its own HBM will appear in the Ascend 950. Even on the optimistic reading, Chinese HBM is a generation or two behind SK hynix's HBM4 (see The Memory Wall).

๐Ÿ The challengers and the money

๐Ÿ“‹ China's AI chip companies

CompanyMain productsLatest numbersNotes
Huawei (HiSilicon)Ascend 910C, 950PR/DT; Atlas SuperPoDs; CANN software~1.5M Ascend units in 2026 (Epoch median estimate)Vertically integrated: chips, HBM, systems, cloud, models
CambriconSiyuan (MLU) acceleratorsRevenue ยฅ9.53B in Janโ€“Sep 2026, +107%; Q3 ยฅ3.53B, net profit ยฅ1.40BProfitable; growth slowing sequentially; supply capacity a constraint
Moore ThreadsMTT GPUsIPO Dec 2025 on Shanghai's STAR Market: raised ยฅ8B; shares +468% on day oneLoss-making
MetaXGPUs for AI and HPCIPO Dec 2025: ~$600M raised; shares +~700% on day one2024 revenue ยฅ743M, net loss ยฅ1.4B
Hyperscaler chipsAlibaba (T-Head), Baidu (Kunlun), othersLimited public dataMostly for their own clouds
CXMTDRAM, HBM3 (planned)H1 2026 revenue ยฅ150B (+874%) amid the memory boomHBM3 mass production targeted late 2026

Cambricon Q3 report (23 Oct 2026, via TechFlow); Caixin; DigiTimes; CNBC; Taipei Times; TrendForce; roic.ai/The Information. Some figures from single reports.

Software is the quieter battle. Huawei's CANN stack and the MindSpore framework are Huawei's answer to CUDA, and Chinese labs increasingly port models to Ascend. Developers report the tooling is less mature, which adds engineering cost to every model.

๐Ÿ”ฆ The last wall: lithography

The deepest gap is the one ASML fills in The Chokepoint. Reuters reported in December 2025 that a Chinese EUV prototype had been completed in early 2025 in Shenzhen, built with the help of former ASML engineers, and was being tested. By August 2026 reports said it generates 13.5 nm EUV light but has not yet produced a working chip; expectations for chip production range from 2028 to 2030. Domestic DUV scanner makers are also progressing, which matters because SMIC's multi-patterning depends on DUV tools that are themselves export-controlled.

โš ๏ธ Treat EUV claims carefully Several widely shared articles describe the Shenzhen prototype in dramatic terms ("Manhattan Project", "Silicon Curtain"), some from automated news sites. Generating EUV light is a long way from a production scanner with the throughput, overlay accuracy and uptime that ASML's machines achieve after decades of refinement.

๐Ÿ•ณ๏ธ Leaks and workarounds

Smuggling
In October 2026 US prosecutors arrested a California man accused of shipping servers with about $300 million of Nvidia AI chips to China through US freight companies and Malaysian transshipment firms. Taiwan opened an investigation in May 2026 into servers allegedly exported with forged documents, and in December 2025 the owner of Hao Global pleaded guilty in another smuggling case. Bloomberg reported in October 2026 that officials say Nvidia is not doing enough; Nvidia says it follows the rules.
Licensed sales
The H20, designed for China, was halted in April 2025 and later licensed; Nvidia's 10-K says US officials "expressed an expectation" of 15% of the revenue, never codified. The H200 was cleared in December 2025, and a BIS rule effective 15 January 2026 moved H200- and MI325X-class chips to case-by-case review. About ten Chinese firms were approved, but a Commerce official told Congress in July 2026 that shipments remained "very few"; Nvidia's first licensed H200 sales came in the quarter to 26 July 2026 and were under 1% of its $89 billion data-centre revenue. Sources describe the 25% charge differently: as a revenue share, or as a Section 232 tariff on chips imported into the US for re-export.
China's own brakes
Beijing has reportedly told domestic firms to stop buying Nvidia chips (from September 2025) and to favour Huawei, which helps explain why licensed sales are small.
Remote access
Chinese companies can rent GPUs in data centres outside China, a loophole regulators have debated closing.
Efficiency
DeepSeek trained V3 on Nvidia's H800, a cut-down China-market chip, and published unusually detailed papers on squeezing performance from limited hardware (see its hardware reflections paper). Constraint has pushed Chinese labs to excel at efficiency.

โš–๏ธ Are the controls working?

๐Ÿ‡บ๐Ÿ‡ธ The case that they work

China's total AI compute is far smaller than America's, and each unit costs more. SMIC's yields and HBM shortages cap volumes. Frontier training runs need hundreds of thousands of top chips, which China cannot assemble at the same scale. Every Chinese chip is a generation or more behind and uses more power for the same work.

๐Ÿ‡จ๐Ÿ‡ณ The case that they backfire

Nvidia's China share has collapsed and Huawei, Cambricon and others have gained a protected home market, funding rapid iteration. Chinese labs keep releasing strong open models. Controls created urgency and state funding that may produce a durable rival stack, including EUV, within a decade.

Both can be true: the controls slow China's total compute in the near term while accelerating its long-term independence. Which effect dominates depends on how fast China closes the HBM and lithography gaps.

๐Ÿ”ฎ Scenarios to 2030

๐ŸŽฒ Explore the futures

Pick assumptions; the panel summarises a plausible outcome. Qualitative and illustrative, not a forecast.

EUV
Domestic HBM
US controls

๐Ÿ”ญ What to watch

๐ŸŽฏ The bottom line US export controls have not stopped Chinese AI, but they have forced it onto a harder path: weaker chips made with multi-patterning at SMIC, joined into enormous Huawei SuperPoDs, limited by scarce HBM and paid for with more power and money. China is building every layer of a parallel stack, and its weakest links, HBM and EUV lithography, decide how fast it can scale. For the rest of the decade the world is likely to run two AI compute ecosystems, with different chips, software and supply chains, competing on efficiency as much as on raw scale.

๐Ÿ”ข Reading the numbers

๐Ÿ“š References