US China Chip War: Export Controls Simulation 2026

Executive Summary

A 15-agent multi-round simulation of US China semiconductor export controls finds a 45% probability that restrictions only partially work, leaving China two or more generations behind at the frontier but unable to prevent 7nm domestic production at scale. The MiroFish simulation framework modeled 10 rounds of strategic interaction between Jensen Huang, Senator Tom Cotton, Huawei executives, ASML advisors, and 11 other key players in the global chip supply chain. The single most important finding: the DUV lithography loophole, with 1,200+ installed deep ultraviolet systems in China, is the hinge on which the entire control regime turns. SMIC has already demonstrated 7nm capability. Huawei's Ascend 910C is in mass production. The controls are self-defeating.

Background and Context

The US China semiconductor conflict has reached a critical inflection point in April 2026. Multiple rounds of US export controls have restricted advanced chip sales to China, targeting Nvidia GPUs, ASML lithography equipment, and chip design software. China has responded by accelerating domestic chip development, with Huawei's Ascend chips now reportedly capable of running DeepSeek's AI models at scale.

Nvidia CEO Jensen Huang warned on April 17-18 that Huawei chips running DeepSeek could shift AI advantage to China, calling it "horrible for the US." An AEI report on April 17 revealed that the DUV lithography loophole allows China to use older-generation deep ultraviolet equipment to print advanced chips, circumventing EUV restrictions entirely. US lawmakers scaled back the MATCH Act on April 16, a bill targeting Chinese chipmaking, signaling that political will for the hardest restrictions is weakening.

Chinese chip firms hit record revenue driven by the AI boom and US curbs. ASML stock sank 18% amid tightening China restriction signals despite strong earnings. The Council on Foreign Relations called the new AI chip export policy "strategically incoherent and unenforceable." Chip smuggling through third countries remains rampant.

This simulation builds on Zeki's prior geopolitical analysis. Our April 13 Hormuz crisis simulation found 62% diplomatic off-ramp probability. Our April 16 Trump-Xi summit simulation modeled US-China economic dynamics. Our April 18 Iran ceasefire simulation identified the Iran war as a critical distraction factor undermining US enforcement capacity.

Methodology

The simulation used the MiroFish multi-agent framework with 15 agents across 10 rounds of strategic interaction. Each agent operates with defined objectives, constraints, and information sets. Agents react to each other's moves, form coalitions, and adjust positions as rounds progress. The simulation converges toward equilibrium outcomes rather than assuming rational actor models.

Agent Personas:

Agent	Role	Primary Objective
Jensen Huang	Nvidia CEO	Preserve China market access, prevent Huawei substitution
Senator Tom Cotton	US Hawk (R-AR)	Total chokehold on China chip access
Rep. Lisa Blunt Rochester	US Moderate (D-DE)	Targeted controls, not blanket bans
Huwei Wei	Huawei Semiconductor VP	Close the chip gap, achieve self-reliance
Dr. Liu Yang	SMIC Chief Engineer	7nm production without EUV
Peter Wennink	ASML Former CEO/Advisor	Balance US pressure with European business
Wang Hao	Chinese Ministry of Commerce	Calibrated retaliation, preserve trade leverage
Maj Gen Scott Berri	US Cyber Command	Monitor and interdict chip smuggling
Satya Nadella	Microsoft CEO	Preserve cloud access, avoid blanket bans
Kim Jong-Soo	South Korea Ministry of Trade	Protect Samsung and SK Hynix China operations
Kayla Smith	BIS Export Enforcement Director	Enforce controls with limited resources
Professor Zhang	Tsinghua AI Lab Director	Prove software efficiency overcomes hardware limits
Tanaka	Japan METI Official	Maintain alliance signal, minimize economic damage
Sheikh Al-Maktoun	Dubai Free Zone Authority	Preserve trading hub model
Deepak Patel	Indian Semiconductor Mission Director	Attract displaced semiconductor investment

Key Findings: China AI Chip Self-Sufficiency Assessment

Probability Distribution

Outcome	Probability
China achieves 7nm+ domestic production at scale by end 2026	30%
US controls partially work (China stays 2+ generations behind)	45%
US controls largely succeed (China stuck at 14nm+)	5%
Wildcard disruption resets the board	20%

The DUV Lithography Loophole is the Hinge

The single most important factor in the simulation is the DUV lithography loophole and the installed base of 1,200+ DUV systems in China. As long as DUV multi-patterning enables 7nm-class production, China has a viable path to domestic AI chip manufacturing at a "good enough" level. SMIC has demonstrated 7nm capability with yields at 55% and improving. Dr. Liu Yang confirmed that if the DUV window closes in six months, the installed base is sufficient to sustain production. The real bottleneck is below 5nm, not 7nm. As Wennink warned: closing DUV would crater ASML revenue, destabilize Europe's semiconductor ecosystem, and create a $50B demand subsidy for Chinese domestic lithography programs.

Export Controls Are Self-Defeating

The simulation identified a self-reinforcing cascade that undermines the restriction regime from within. US restrictions trigger Chinese domestic investment. "Good enough" domestic chips emerge. US allies lose market share. Allied commitment to restrictions weakens. Enforcement becomes more unilateral and less effective. The US doubles down with more restrictions. The cycle repeats. Wennink's contingency planning, Kim's patience limits, Tanaka's enforcement stall, and Huawei's market capture all feed this spiral. Each escalation produces the conditions for its own failure.

Huawei Ascend 910C Changes the Game

Huawei's Ascend 910C entering mass production was the pivotal moment in the simulation. Nvidia's China market share dropped from 90% to 55% in two years. The chip achieves 80% of A100 performance on inference workloads. SMIC produces 500,000 units per quarter with yields at 50% and improving. BIS Enforcement Director Kayla Smith acknowledged that domestic manufacturing is beyond the regime's scope: "We can restrict imports, but we cannot restrict a country's domestic manufacturing without a full economic embargo, which no one is proposing."

DeepSeek Proves Software Can Substitute for Hardware

Professor Zhang's persistent argument about algorithmic efficiency reframed the debate from "can China get the chips?" to "does China need the chips?" DeepSeek achieved GPT-4-level results with 1/10th the compute of GPT-4's training run. The efficiency multiplier is compounding. Each generation of Chinese models requires proportionally less hardware. Huawei's Ascend 910C running DeepSeek-style architectures achieves 80% of A100 performance on inference. The US is optimizing against one variable in a four-variable equation: hardware times software times data times talent.

The Nvidia China Export Ban Backfires

Jensen Huang's position crystallized across the simulation: controls are self-defeating because Nvidia loses $5B+ in China revenue while Huawei fills the void. "We're subsidizing our own competition." The CFR analysis corroborates this, calling the policy "strategically incoherent and unenforceable." Nvidia's market share decline is not a sign the controls are working. It is a sign that Chinese domestic alternatives are replacing American supply.

Market Implications

Semiconductor equities: ASML faces a lose-lose scenario. Closing the DUV loophole costs 25% of revenue. Keeping it open allows China to scale 7nm production. ASML stock has already declined 18% on restriction signals. Nvidia's China revenue collapse from $5B+ transfers to Huawei, not to zero demand. Samsung and SK Hynix face customer attrition as Chinese AI companies switch from Samsung memory paired with Nvidia GPUs to Samsung memory paired with Huawei Ascend, with long-term risk of full domestic substitution via Yangtze Memory's 232-layer NAND.

Rare earths and critical minerals: China controls 60% of rare earth processing and 90% of magnetic materials. Gallium and germanium prices rose 40% in April 2026 on restriction fears. Draft executive orders for comprehensive export controls on these materials await Politburo approval. China has calibrated its response carefully, keeping rare earth retaliation sheathed, but the threat remains credible.

Oil and energy: The Iran war continues to distract US enforcement resources and political attention. US Cyber Command reallocated 60% of its China-monitoring cyber personnel to Middle East operations. This enforcement gap is a tailwind for Chinese chip procurement networks.

Indian semiconductor opportunity: India's planned 28nm and 14nm fabs gain commercial viability if "good enough" chips at 7nm can serve most AI workloads. However, execution remains the binding constraint. Micron's Gujarat fab is underway but insufficient to change the global balance.

Second-Order Effects

Alliance fracturing accelerates. The simulation converges on a clear warning: if the US does not share the economic benefits of controls alongside the costs, allied commitment will fray to ineffectiveness by 2027. Europe loses ASML revenue. Japan loses chemical market share. Korea loses fab customers. The asymmetry is politically unsustainable. Tanaka's Japan and Kim's South Korea are already signaling reduced enforcement enthusiasm.

Cloud access becomes the unaddressed frontier. Nadella's position throughout the simulation is that physical chip controls address a diminishing share of the threat surface. Chinese entities access frontier AI capabilities through APIs, model weights, and cloud partnerships regardless of domestic chip production. Senator Cotton's proposed cloud restriction bill would force Microsoft to terminate Azure contracts across Asia, losing both revenue and visibility.

Smuggling networks are uncontrollable at current resource levels. BIS has 200 enforcement officers tracking 50,000+ export license applications. The interdiction rate is under 15%. Chinese procurement networks acquire any chip within 90 days through third countries at a 20-30% cost premium. Dubai reported a 300% increase in chip procurement by Chinese buyers since 2024. The Iran war has increased Middle East trade volumes, creating more places to hide semiconductor shipments.

The definition of success has been quietly downgraded. A fundamental redefinition occurred across the 10 rounds. The original goal, denying China advanced AI capability, has shifted to maintaining a 12-18 month lead at the frontier. BIS Director Smith offered a bureaucratic definition: making illicit acquisition "expensive, slow, and uncertain enough that it cannot support a strategic military AI program at scale." By that definition, controls are partially succeeding. But "partial success" is not what Congress was promised.

Risk Assessment

What the simulation may overestimate: China's ability to sustain yield improvement trajectories. SMIC reports 55% yields at 7nm, but independent verification is limited. If yields stall below 60%, "at scale" production becomes uneconomical and the probability shifts toward partial control success. The simulation also may underestimate US capacity for creative enforcement, including secondary sanctions on transshipment hubs and AI-powered trade monitoring.

What the simulation may underestimate: The probability of a Taiwan Strait crisis, assigned 20% in the wildcard category, could be too low given the current geopolitical environment. Any military escalation between China and Taiwan would immediately halt TSMC production, which manufactures roughly 90% of the world's most advanced chips. This would render current export controls irrelevant, replacing regulatory denial with physical supply chain destruction. The impact would be existential for the entire semiconductor ecosystem.

Uncertainty bands: The 45% probability assigned to "partial success" carries significant uncertainty. If the DUV loophole closes by late 2026 under US pressure, the installed base of 1,200+ systems means the impact is delayed, not prevented. If alliance commitments hold better than the simulation suggests, enforcement improves. If DeepSeek's efficiency gains hit diminishing returns faster than Professor Zhang projects, the hardware constraint tightens. The range of reasonable outcomes spans from meaningful delay (controls buy 18-24 months) to functional irrelevance (China achieves 7nm self-sufficiency by Q3 2026).

Conclusion

US semiconductor export controls are strategically relevant but strategically insufficient. They impose real costs and delays on China's AI chip development, but they cannot prevent 7nm domestic production at scale. The DUV lithography loophole, Huawei's Ascend 910C, DeepSeek's software efficiency gains, and enforcement incapacity at current resource levels collectively undermine the restriction regime. The most likely endpoint is not a decisive victory for either side but a fragmented global semiconductor ecosystem where China is self-sufficient at 7nm, the US leads at the frontier, and the rest of the world navigates between both poles with significant friction, inefficiency, and strategic instability as the permanent condition. The policy choice is no longer whether to contain China's AI rise, but how to manage a multipolar semiconductor world.