In October 2022, the United States imposed the most sweeping semiconductor export controls in history, explicitly designed to deny China access to advanced chip technology and slow its AI development. Three and a half years later, the controls have produced the opposite of their intended effect. SMIC has achieved high-volume production of a 5nm-class process using DUV multipatterning — a workaround that circumvents the EUV lithography controls that were the centrepiece of the restriction strategy. Huawei plans to produce 600,000 Ascend 910C AI chips in 2026, doubling output, manufactured by SMIC on an enhanced 7nm process. The Ascend 910C delivers roughly one-third the per-chip performance of Nvidia’s B200 — but Huawei compensates with scale, building 384-processor SuperPod systems positioned as direct alternatives to Nvidia’s infrastructure. Nvidia’s China AI chip market share has fallen from over 90% to approximately 50%. Chinese firms purchased $38 billion in semiconductor equipment from the five largest allied suppliers in 2024 — up 66% from 2022 when controls began. ASML continues to service DUV lithography machines in China, potentially extending their operational lifespans to 30 years. China has announced a 50% domestic equipment mandate for its fabs. Alibaba, Tencent, ByteDance, and DeepSeek are migrating to domestic alternatives. An EUV prototype reportedly exists but will not reach commercial viability until 2030 at the earliest. The export controls did not prevent China from building a parallel semiconductor ecosystem. They accelerated it. The at-risk thesis is not about China catching up. It is about a permanent bifurcation of the global semiconductor supply chain into two incompatible stacks — with consequences for every company, country, and AI model that depends on chips.
High-volume production using DUV multipatterning. Circumvents EUV controls. Yields ~30%, below commercial viability threshold of ~70%, offset by state subsidy.[1]
Huawei’s 2026 production target. Double 2025 output. Total Ascend line: up to 1.6M dies. SMIC 7nm process. Dual-chiplet design. ~1/3 B200 per-chip performance.[2]
Chinese firms purchased from top 5 allied equipment suppliers in 2024. Up 66% from 2022 when controls began. DUV machines still flowing freely.[3]
Down from >90% pre-controls. 25% tariff + 30-day Congressional review (AI Overwatch Act) pushing Chinese firms to domestic alternatives.[1]
China requires all domestic fabs to source at least half their equipment from local vendors. ASML warned of significant 2026 China revenue decline.[1]
950PR (Q1 2026), 950DT (Q4 2026), 960 (Q4 2027), 970 (Q4 2028). Each generation doubles compute. Full sovereign AI stack under construction.[4]
The chokepoint strategy rested on a specific theory: deny China access to EUV lithography machines, and it cannot manufacture chips below 7nm. The theory was correct about EUV. It was wrong about everything else. SMIC demonstrated that DUV multipatterning — using older lithography tools with multiple exposures — can produce chips at 5nm-class geometry. The process is slower, more expensive, and lower yield than EUV-based manufacturing. But China’s strategy is volume over performance: produce enough chips at good-enough quality to fill the domestic market, subsidise the cost difference with state funding, and optimise software for domestic architectures. The strategy is working.[1][3]
The equipment loophole is the structural flaw. Current export controls ban only the most advanced DUV tools to all of China. Less advanced DUV tools are banned only for specific entities, leaving most Chinese fabs free to purchase them. Even banned entities can access equipment purchased before controls or obtained through intermediary companies. Chinese firms established shell companies specifically to ensure unhindered equipment flow. ASML continues to service existing DUV machines in China — maintenance that extends machine lifespans to 30 years, far beyond original projections. The controls created a wall with holes large enough to drive a semiconductor supply chain through.[3][5]
Huawei’s Ascend 910C is the headline, but the story is the full stack being assembled underneath it. SMIC provides the foundry. HiSilicon (Huawei’s design subsidiary) provides the architecture. CXMT (ChangXin Memory Technologies) is building domestic HBM production — estimated at 2 million stacks in 2026, sufficient for 250,000–300,000 Ascend 910C packages. Huawei open-sourced its CANN (Compute Architecture for Neural Networks) software framework in 2025, triggering a wave of domestic developer adoption. The MindSpore AI framework provides the training and inference layer. The CloudMatrix 384 SuperPod — integrating 384 Ascend processors — is positioned as a direct alternative to Nvidia’s GB200 platform.[4][6]
The performance gap is real but narrowing through a different strategy. Each Ascend 910C delivers roughly one-third the BF16 throughput of Nvidia’s B200. But Huawei compensates horizontally: more silicon, more power, more systems. The SuperPod architecture enables customers to aggregate compute across multiple data centres. The per-chip efficiency penalty is absorbed at the system level. For China’s domestic market, the question is not whether Huawei matches Nvidia chip-for-chip. It is whether the sovereign stack is good enough for the workloads that matter — and whether the cost is subsidised enough to make the economics work.[7]
The HBM bottleneck is the remaining constraint. SMIC can produce the logic dies. CXMT is ramping HBM but yields are early-stage. SemiAnalysis estimates that once foreign HBM stockpiles are exhausted, domestic production becomes the binding constraint. If CXMT can scale HBM3e production in 2026, the Ascend production targets become achievable. If yields remain low, Huawei will be limited to 250,000–300,000 packaged units regardless of how many logic dies SMIC can produce.[6]
The revenue impact on Western firms is material. Nvidia estimated $5.5 billion in lost revenue from China restrictions. AMD lost approximately $800 million. ASML has warned of significant China revenue decline due to the 50% domestic equipment mandate. The controls are not just failing to prevent Chinese self-sufficiency — they are actively destroying revenue streams that US and allied firms had used to fund their own R&D. Every dollar of lost China revenue is a dollar less for Nvidia to invest in Rubin, for ASML to invest in high-NA EUV, for Applied Materials to invest in next-generation deposition tools. The controls are weakening the Western ecosystem they were designed to protect.[8]
| Dimension | Evidence |
|---|---|
| Regulatory / Geopolitical (D4)Origin · 85 At Risk | Export controls are the origin because the entire cascade flows from the policy decision to restrict China’s semiconductor access. October 2022 initial controls, tightened October 2023 and December 2024. 65+ entities added to Entity List in 2025. AI Overwatch Act (30-day Congressional review). H200 approved with 25% tariff and end-use certification. VEU revoked for Samsung/SK Hynix China operations. The controls correctly identified the strategic risk but the execution created a paradox: restrictions that accelerate the very self-sufficiency they aim to prevent. The regulatory dimension scores highest because it is the forcing function driving every other dimension.[5][9] |
| Operational (D6)L1 · 78 At Risk | SMIC 5nm-class (N+3) production via DUV multipatterning. Total advanced-node capacity estimated 45K wafers/month in 2025, growing. Huawei 600K Ascend 910C target for 2026. CXMT ramping HBM domestically (~2M stacks target). Shell companies circumventing Entity List restrictions. $38B in equipment purchased from allied suppliers in 2024 (+66% from 2022). ASML servicing extends DUV lifespan to 30 years. 50% domestic equipment mandate. China EUV prototype exists but not commercially viable until ~2030. The operational dimension captures the physical manufacturing reality: China is building capacity at a pace that makes the controls increasingly irrelevant for everything except the absolute leading edge.[1][3][6] |
| Revenue / Financial (D3)L1 · 75 | Nvidia estimated $5.5B lost China revenue. AMD ~$800M. ASML warning of significant China decline. Nvidia China AI market share: >90% → ~50%. Huawei Ascend revenue growing as domestic cloud and enterprise customers migrate. $240B rally in Chinese tech stocks driven by domestic champion thesis. China semiconductor investment: $70B proposed injection (largest government semiconductor investment of any nation). Alibaba, Tencent, ByteDance, DeepSeek migrating to domestic alternatives. The revenue dimension captures both the Western revenue destruction and the Chinese domestic market creation. The flow of funds is reversing.[1][8] |
| Customer (D1)L1 · 72 | Chinese AI companies — Alibaba, Tencent, ByteDance, Baidu, DeepSeek — are the world’s second-largest market for AI compute. They need millions of AI chips annually. Nvidia sold an estimated 1 million H20 chips in China in 2024. As Ascend production scales, these customers are migrating to Huawei’s ecosystem. Huawei open-sourced CANN software in 2025, lowering switching barriers. The customer dimension captures demand-side bifurcation: China’s AI companies are building on a stack that may become permanently incompatible with the Western ecosystem. Once software is optimised for Ascend, the switching cost to Nvidia becomes a barrier in both directions.[2][4] |
| Employee / Talent (D2)L1 · 68 | China’s semiconductor workforce is expanding rapidly. State-backed programmes are training fab engineers, process designers, and AI chip architects. Huawei employs 195,000+ globally. The domestic ecosystem is creating career paths that did not exist five years ago. On the Western side, the controls have disrupted talent flows: Chinese nationals in US semiconductor programmes face restrictions; reverse brain drain is pulling experienced engineers back to China. The employee dimension captures the human capital shift: talent is flowing toward the ecosystem that is growing, not the one that is restricting.[5] |
| Quality / Technology (D5)L2 · 65 | Ascend 910C: ~1/3 BF16 throughput of Nvidia B200 per chip. Compensated by scale (SuperPod 384-processor systems). SMIC 7nm yields ~20% (well below commercial viability), offset by state subsidy. N+3 (5nm-class) yields ~30%. China’s volume-over-performance strategy is pragmatic but the quality gap is real. Domestic HBM is early-stage. EUV prototype not commercially viable until 2030. Software ecosystem (CANN, MindSpore) improving but still behind CUDA. The technology gap exists but the question is not parity — it is sufficiency. For many AI workloads, good-enough at volume may be better than best-in-class at constrained supply.[7][4] |
-- Silicon Sovereignty: 6D At-Risk Cascade
-- Semiconductor Cluster Case 3 of 4 (UC-103, UC-104, UC-105, UC-106)
FORAGE china_silicon_sovereignty
WHERE smic_advanced_node_production = true
AND duv_multipatterning_5nm = true
AND huawei_ascend_target > 500_000
AND nvidia_china_share_pct < 0.55
AND china_equipment_spend_yoy_increase > 0.50
AND domestic_equipment_mandate > 0.50
AND asml_china_servicing = true
AND hbm_domestic_production_active = true
ACROSS D4, D6, D3, D1, D2, D5
DEPTH 3
SURFACE silicon_sovereignty
DIVE INTO controls_paradox
WHEN export_controls_active AND domestic_self_sufficiency_accelerating AND western_revenue_declining AND ecosystem_bifurcating
TRACE at_risk_cascade
EMIT at_risk_signal
DRIFT silicon_sovereignty
METHODOLOGY 85 -- Export controls comprehensive: Entity List, EUV denial, DUV restrictions, AI Overwatch Act, VEU revocation, end-use certification, multilateral coordination with Japan/Netherlands
PERFORMANCE 35 -- SMIC 5nm production, Huawei 600K target, Nvidia share halved, $38B equipment purchases UP 66%, ASML still servicing, DUV loophole open, shell companies active, domestic HBM ramping
FETCH silicon_sovereignty
THRESHOLD 1000
ON EXECUTE CHIRP at_risk "SMIC 5nm-class via DUV multipatterning. Huawei 600K Ascend 910C target (2x output). 1.6M total Ascend dies. Nvidia China share: 90% to ~50%. $38B in allied equipment purchased 2024 (+66% from 2022). ASML still servicing DUV in China. 50% domestic mandate. CXMT ramping HBM. DeepSeek, Alibaba, Tencent migrating to domestic. CANN open-sourced. SuperPod 384 as GB200 alternative. EUV prototype exists but 2030. Export controls accelerating the self-sufficiency they were designed to prevent. Two semiconductor ecosystems forming. Revenue destruction weakening Western R&D funding. The paradox is structural: the tighter the controls, the faster the parallel stack."
SURFACE analysis AS json
Runtime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
Export controls were designed to maintain Western semiconductor advantage by denying China access to advanced technology. Instead, they removed China’s incentive to remain a customer and replaced it with an existential imperative to build a domestic alternative. The $38 billion in equipment purchases — up 66% since controls began — is the empirical proof. The tighter the controls, the faster China invests in alternatives. The paradox is structural, not fixable by better implementation.
China’s strategy is not to match Western chips on per-unit performance. It is to produce enough good-enough chips at volume to fill the domestic market. The Ascend 910C delivers one-third the throughput of Nvidia’s B200 — but Huawei compensates with SuperPod systems that aggregate 384 processors. For many AI workloads, the question is not single-chip speed but total system throughput. China is building ten-times more capacity at one-third the power to get three-times the aggregate compute. The math works.
Nvidia lost $5.5 billion. AMD lost $800 million. ASML is warning of China revenue decline. These are the companies whose R&D spending maintains the Western technology lead that the export controls are designed to protect. Every dollar of lost China revenue is a dollar less for next-generation chip development. The controls are simultaneously accelerating Chinese self-sufficiency and weakening the Western ecosystem’s R&D budget. This is a negative-sum outcome.
Once Chinese AI software is optimised for Ascend architectures, once data centres are built on Huawei hardware, once developers are trained on CANN instead of CUDA, the switching cost to return to Western chips becomes prohibitive — even if export controls were lifted. The software ecosystem lock-in is the same force that maintains TSMC’s moat (UC-103) and prevents Intel from winning customers (UC-104). In China’s case, it works in reverse: the ecosystem lock-in that forms around domestic chips may make the bifurcation permanent.
If something you are dealing with looks like this case, write.