Technology Export Controls: US-China Chip War and Allies Compliance Pressure

Executive Summary

US semiconductor export controls, structured since October 2022 and substantially revised through mid-2026, are materially slowing China's ability to develop AI infrastructure and advance military modernization timelines, but the controls are leaking at multiple seams through stockpiling, circumvention via shell companies, and partial policy reversals under the Trump administration. The Bureau of Industry and Security shifted from blanket denial to case-by-case review for Nvidia H200 and AMD MI325X chips effective January 15, 2026, creating a more porous but more complex regime. Japan and the Netherlands have formally joined the coalition, South Korea faces escalating pressure, and electronic design automation software controls introduced in May 2025 are tightening the vice deeper in the value chain. The picture, however, is mixed: China's AI developers are finding algorithmic workarounds, state capital is absorbing market shocks, and Beijing is retaliating through critical mineral export restrictions that directly threaten allied defense industrial supply chains.

• Supply-chain/operations: Map your critical mineral inputs, specifically gallium, germanium, tungsten, and rare earth processing materials, against Chinese export control lists updated as of October 2025; the IEA reports prices for these inputs are now up to five times higher outside China than within.
• Risk officers/investors: The January 2026 case-by-case review regime introduces license uncertainty for any AI infrastructure project with China exposure; do not treat H200 access to China as a stable planning assumption.
• Technology/policy stakeholders: Monitor the Select Committee on China's August 2026 push to clarify the Foundry Due Diligence Rule, which, if resolved in favor of stricter interpretation, will close a material loophole in global chip fabrication controls.

US semiconductor export controls are delaying but not foreclosing China's AI and military chip ambitions, while allied cohesion on restrictions is fracturing unevenly across Japan, the Netherlands, South Korea, and Southeast Asia.

What Changed

The regulatory environment shifted substantially between January and August 2026. BIS moved from presumption of denial to case-by-case licensing for H200-class chips on January 15, 2026, applied a 25% Section 232 tariff on advanced chip exports on January 14, and simultaneously eliminated the Validated End-User program for foreign fabs in China, requiring Samsung and SK hynix to seek annual licenses for their Chinese facilities. In parallel, Chairman John Moolenaar of the House Select Committee on China wrote to BIS in August 2026 demanding clarification that the Foundry Due Diligence Rule, which governs chip fabrication for Chinese end users via third-country foundries, remains in effect, exposing a live enforcement gap.

The Chokepoints That Actually Matter: Eda, Hbm, And Fabrication Tools

The public debate on semiconductor export controls focuses almost entirely on finished chips, but the controls with the most durable strategic effect operate three layers deeper in the value chain. Western Electronic Design Automation software, specifically Synopsys, Cadence, and Siemens EDA suites, controls approximately 80% of China's EDA market, according to the IISS July 2025 analysis. When the US introduced EDA controls in late May 2025, the IISS assessed that Chinese domestic alternatives such as Empyrean Technology "significantly lag behind." Without EDA tools, Chinese engineers cannot design chips at sub-7nm nodes regardless of whether they can access fabrication equipment.

The second chokepoint is High Bandwidth Memory. The Department of Defense Annual Report documents that Huawei stockpiled approximately a year's worth of HBM chips before controls took effect, suggesting Beijing anticipated this as a critical constraint. Samsung and SK hynix, both Korean companies with Chinese production facilities, now require annual BIS licenses for those facilities following the January 2026 elimination of the Validated End-User program. Annual uncertainty, as Introl's February 2026 analysis notes, introduces a planning horizon problem for Korean chipmakers that effectively pressures them toward reducing Chinese fab exposure without requiring a formal prohibition.

The third chokepoint is front-end fabrication equipment. ASML's EUV and DUV lithography tools remain restricted for China, and Japan's controls target the five major Chinese fabs by name. The Asia Group noted in December 2024 that these controls were designed to receive reciprocal expansion from Japan and the Netherlands in exchange for carve-outs protecting their own industries. This translates directly into a commercial pressure on ASML and Tokyo Electron: the allied carve-out structure rewards compliance with tighter restrictions by preserving non-China market access, but creates an incentive to seek exemptions rather than enforcement.

Beijing's Three-Track Response And Its Military Implications

China has responded to US semiconductor controls through three concurrent tracks, each of which has direct military implications. The Department of Defense Annual Report to Congress (December 2025) documents all three.

The first track is AI accelerator stockpiling. Beyond the DeepSeek and Huawei examples, the DoD report indicates that Chinese AI companies are broadly building GPU reserves to buffer against future tightening. This translates into a short-term military AI development window during which Chinese defense researchers can draw on pre-control compute stocks. The West Point Modern War Institute assessed in October 2022 that China's installed AI infrastructure lagged the US by a factor of roughly ten data centers to one, but that structural energy advantage, with China's installed generation capacity more than double that of the United States and growing fifteen times faster, means China can sustain the power demands of successive AI training runs in ways constrained by US grid capacity.

The second track is Huawei's Ascend ecosystem. Huawei used shell companies to procure approximately two million Ascend 910B dies from TSMC before TSMC suspended production of AI semiconductors at 7nm or below for Chinese clients in November 2024. The American Affairs Journal reported in November 2024 that TSMC's suspension, likely driven by Commerce Department pressure, represented TSMC "acting as an arm of US export control policy," but simultaneously "raises the geopolitical risks around the complex relationship between the United States, China, and Taiwan." The military implication is that Huawei's Ascend 910B, while not NVIDIA-class, is now embedded in Chinese military-civil fusion procurement chains in large enough volume to sustain domestic AI development at a reduced but functional pace for several years.

The third track is algorithmic efficiency investment. The West Point analysis notes that DeepSeek's Kimi K3 model, released July 16, 2026, landed during a global semiconductor rally and triggered a market reaction precisely because it demonstrated competitive model performance from a constrained compute base. This is the pattern that the IISS flagged as early as 2022: export controls impose costs and delays but do not eliminate indigenous development capability, particularly when Chinese firms are incentivized to compete on efficiency rather than raw compute.

What is not being reported: Chinese AI capability assessments in Western press coverage systematically undercount algorithmic advances that reduce hardware dependency. The focus on GPU headcounts misses the strategic adjustment Beijing has already made: trading compute scale for algorithmic density, a trade-off that partially defeats the logic of hardware-focused export controls.

The Allied Pressure Map: Who Is In, Who Is Hedging, Who Is Exposed

The US-Japan-Netherlands tripartite core has been the backbone of semiconductor export control multilateralism since January 2023. CSIS confirmed all three coordinated on equipment controls, with Japan specifically restricting chipmaking equipment to 14nm and below processes. Taiwan cooperates implicitly through TSMC's compliance posture, including the November 2024 suspension of sub-7nm AI chip production for Chinese clients. South Korea faces the most acute tension: Seoul's Samsung and SK hynix have significant Chinese manufacturing facilities, and the Annual Validated End-User license structure introduced in January 2026 places both companies in annual licensing negotiations with BIS, creating recurring commercial uncertainty that functions as soft enforcement pressure.

Germany and the broader European Union are surveyed by CSIS as possessing the legal authority to implement aligned controls but have moved more slowly. Southeast Asia presents the most serious gap. Singapore, Malaysia, the Philippines, and Thailand have implemented strategic trade management frameworks, but these predate the current AI chip control regime and do not systematically address the GPU-class hardware that US controls target. The Brookings semiconductor supply chain analysis explicitly identifies Shenzhen and Hong Kong broker networks as continuing to deliver chips to Chinese buyers, suggesting the geographic perimeter of the control regime is porous at the southern tier.

South Korea's domestic political environment, while running approximately 80% negative sentiment toward China per the Brookings transcript, does not automatically translate into tighter restrictions, because Samsung and SK hynix's Chinese revenue represents a material share of their global operations. The annual license regime for their Chinese fabs thus creates a structural pressure that is more durable than diplomatic appeals but less certain than a statutory prohibition.

The broader geopolitical implications connect the trade and security domains: as allied governments accelerate domestic fab investment under the CHIPS Act framework and European equivalents, the commercial case for maintaining Chinese fab exposure weakens, which gradually aligns economic incentives with security policy. The US CHIPS Act's $52.7 billion funding package conditions recipients on not expanding advanced manufacturing in China, directly tying subsidy access to compliance with the control architecture.

The scatter above illustrates the core policy tension: nations with the highest economic exposure to China (South Korea, Singapore, Malaysia) tend to show the lowest export control alignment scores, while the US-Japan-Netherlands core anchors the upper-left cluster, demonstrating that lower economic exposure correlates with stronger alignment. South Korea sits in a structurally difficult middle position, and its trajectory in the next 12 months is likely the single most important variable for the coherence of the allied control architecture.

Counterarguments

1. The "controls are working" case overstates hardware dependency: The strongest challenge to this assessment's primary finding is that China may be adapting faster at the software layer than hardware restriction models anticipate. DeepSeek's ability to produce competitive AI models from constrained compute stocks, documented in the West Point Modern War Institute analysis of July 2026, suggests that the GPU-centric model of AI capability may itself be strategically miscalibrated. If algorithmic efficiency improvements continue to outpace hardware-dependency assumptions, then export controls slow China's AI development at a significantly lower rate than the headline capability gap implies. The DoD's own report notes that export controls create delays, not ceilings, and Chinese firms are demonstrably investing in the path around the hardware wall.

2. The allied coalition is more fragile than its core suggests: Presenting the US-Japan-Netherlands arrangement as a stable multilateral architecture overstates its durability. CSIS explicitly notes that Japan and the Netherlands were not even publicly acknowledged as partners until 2023, and the carve-out structure of their cooperation means their compliance is commercially conditioned. The Asia Group's analysis shows Japan and the Netherlands expected carve-outs in exchange for their participation, creating a transactional rather than principled alignment. Any US policy shift that appears to prioritize bilateral trade deals over control integrity, as the Congressional Research Service documents with the H200 licensing decision, degrades partner confidence in the US as a reliable multilateral anchor.

3. Chinese state capital can absorb market shocks that controls rely on to inflict pain: The Foreign Policy analysis of August 2026 makes the most substantive challenge to the economic pressure logic. Beijing's Politburo communique for the second half of 2026 confirms the state has explicitly chosen not to prioritize household consumption, meaning the transmission mechanism of tariffs and export controls, namely, consumer pain generating political pressure on the CCP, is severed by design. State funds, including a central bank relending facility established October 2024 specifically to finance technology-sector equity purchases, can absorb market shocks by buying technology stocks directly. When Moonshot AI's Kimi K3 triggered an 8% drop in the Shanghai STAR Market in July 2026, central state-capital operators deployed approximately 60 billion yuan within two days. This is not a government that will be moved by commercial pressure on its semiconductor sector.

Indicators To Watch

Near-term watch list: (1) BIS response to Chairman Moolenaar's August 10, 2026 letter on the Foundry Due Diligence Rule, expected within 30-60 days, will clarify whether front-end fabricators globally face worldwide licensing requirements for Chinese-linked customers; (2) Trump-Xi bilateral in the US, potentially September 2026, where chip access could surface as a negotiating variable, per Kharon's June 2026 reporting; (3) TSMC Q3 2026 earnings call, October 2026, where China revenue disclosures and capex guidance by node will reveal whether the sub-7nm compliance posture is holding under commercial pressure.

Decision Relevance

Scenario A (~55%): Porous status quo, controls degrade gradually but hold structurally. The current patchwork regime, with case-by-case H200 licensing, annual Korean fab licenses, TSMC sub-7nm suspension, and EDA controls in place, persists through 2027 without a major bilateral deal or coalition defection. China continues stockpiling, algorithmic workarounds, and Ascend development, while the allied core holds. If you advise on technology sector investment or hold positions in semiconductor equipment companies, maintain existing exposure to ASML, Tokyo Electron, and US equipment makers, which benefit from allied fab buildout, but flag annual Korean license renewal as a binary risk event. If you lack direct semiconductor exposure, monitor the BIS Foundry Due Diligence Rule clarification in Q3 2026 as the signal for whether the US is tightening or loosening the equipment perimeter.

Scenario B (~30%): US-China bilateral deal partially suspends controls on H-series chips. A September 2026 Trump-Xi meeting produces an exchange of chip access for trade or geopolitical concessions, further relaxing the case-by-case regime and effectively signaling to Japan and the Netherlands that the US will not maintain control discipline under commercial pressure. If you have supply-chain or sourcing exposure in the semiconductor sector, this scenario accelerates the timeline for Chinese domestic fab competition and reduces the strategic moat of allied equipment producers. If you are a risk officer in defense or national security sectors, this scenario requires reassessment of Chinese military AI development timelines, which could compress by 18-24 months if frontier compute access is restored.

Scenario C (~15%): Coalition hardens, Southeast Asian gap closes, full perimeter established. BIS clarifies Foundry Due Diligence Rule strictly, South Korea formalizes permanent restrictions rather than annual licenses, and Singapore/Malaysia implement strategic trade management for GPU-class hardware. If you have supply-chain dependency on Chinese-origin materials, including gallium, germanium, or rare earth processing, begin active diversification now; a hardening US coalition will high confidence be met by Beijing expanding critical mineral restrictions further, as the IEA's 2026 Outlook documents. If you operate in defense sectors dependent on GaN radar components, this scenario accelerates the urgency of domestic gallium refining investment.

Expert Integration

Expert Consensus Assessment

Government, academic, and think-tank references broadly agree that US semiconductor export controls have created measurable delays for China's AI and advanced chip development and that allied coordination, particularly with Japan and the Netherlands, has been achieved. Consensus is weaker on the magnitude and durability of these delays.

Expert Disagreement Areas

• Delay magnitude: The IISS assessed in 2022 that controls would set back China's semiconductor industry "by many years," while Brookings research documents active indigenization strategies that could compress that estimate; the American Affairs Journal argues that Chinese commercial AI developers will eventually be pushed "into hiding" in ways that reduce Western visibility rather than actual capability.
• Allied cohesion durability: CSIS acknowledges that the Japan-Netherlands coordination was commercially conditioned from the outset; the Brookings transcript documents allied frustration that US controls were "presented to allies as if, here are the rules, now everyone should fall in line," indicating the multilateral architecture carries latent tension.
• Algorithmic substitution: The West Point Modern War Institute's October 2022 analysis and the DeepSeek/Kimi K3 evidence of July 2026 suggest algorithmic efficiency improvements are materially reducing hardware dependency, a position that challenges the DoD's hardware-centric control framework.

Areas Of Expert Agreement

• China's chip indigenization is driven by business-continuity logic, not primarily by state mandate, per Brookings.
• ASML EUV and DUV restriction, combined with Japan's equipment controls, represents the most durable near-term chokepoint.
• Chinese critical mineral export restrictions are a direct and structurally symmetric counterweight to US semiconductor controls.

Systematic-Expert Alignment

Alignment: MIXED

This assessment aligns with expert consensus on the short-term effectiveness of hardware controls and the EDA chokepoint. It diverges from some think-tank analyses by weighting the algorithmic substitution risk more heavily and by treating the January 2026 H200 licensing relaxation as a coalition-cohesion risk rather than a manageable policy adjustment.

• Chinese AI military applications are not publicly documented in sufficient detail to assess how PLA-specific capability development has been affected by controls; the DoD Annual Report's military AI section relies on open-source and commercial assessments that may systematically underestimate classified program progress.
• The precise terms of the US-Netherlands-Japan coordination agreement remain non-public; CSIS notes the exact contours of the deal are not publicly known, meaning this assessment cannot confirm whether allied obligations are legally binding or political commitments.
• Chinese domestic semiconductor manufacturing data is self-reported by state-affiliated sources; if actual production capacity at SMIC and Hua Hong exceeds reported figures, the timeline compression implied by this assessment would accelerate further.
• The IEA critical minerals price data reflects market conditions as of early 2026; if alternative gallium and germanium production outside China scales faster than projected, Beijing's retaliatory leverage through mineral export controls could diminish, reducing one of the key feedback pressures on allied cohesion.
• This assessment draws on English-language analysis of Chinese capabilities; Chinese-language technical literature, CCP internal documents, and PLA procurement records, which would materially inform the military AI development timeline assessment, are unavailable.