Winning the AI Competition: Why U.S. Chip Export Controls Need a Strategic Upgrade

Report Scope

Public sources anchor the analysis, including U.S. government documents, congressional investigations, academic research, industry filings, think tank analysis, and investigative journalism. Radiant Intel also contributes a proprietary operational case study from its intelligence platform. The case study matters because policymakers justified the 2022 controls largely as a way to preserve U.S. leadership in AI. Outside researchers cannot reproduce the case study, so readers should treat it as corroborative field evidence rather than standalone proof.¹

Radiant Intel has a commercial interest in an open and competitive AI model ecosystem. The commercial interest warrants disclosure but does not invalidate the evidence. Public evidence therefore carries more weight than the firm's internal evidence wherever possible.²

The paper focuses on the AI capability rationale U.S. policymakers publicly used to justify the controls. The paper does not attempt a comprehensive assessment of every military-specific semiconductor use case. A finding about AI competition alone does not resolve all questions about military electronics, sanctions signaling, or alliance politics. The paper instead tests the core benchmark set by the policy's architects: maintaining “as large of a lead as possible” in advanced computing and AI.³

Executive Summary

U.S. policymakers imposed the October 2022 semiconductor export controls to preserve and widen the U.S. lead in advanced AI. The controls restricted China's access to frontier chips, semiconductor manufacturing equipment, and related know-how.⁴ The policy imposed real friction. Chinese firms still face meaningful cost penalties, lower yields, persistent EDA dependence, and continued denial of EUV lithography.^5,6

Friction is necessary but not sufficient for strategic success.

The key question is whether the costs produced the policy's central strategic outcome: a durable and widening U.S. advantage in advanced AI. The evidence points to a weaker result. The controls imposed friction. They did not stop Chinese progress in advanced semiconductor production or halt the narrowing of the AI capability gap. The controls may also have accelerated the very outcomes the United States sought to prevent: supplier substitution, domestic equipment prioritization, and ecosystem adaptation inside China.^7,8,9

Five findings drive the assessment.

• First, the controls achieved partial success at the manufacturing layer. EUV denial remains real. Process and manufacturing penalties remain real. Serious assessments should acknowledge the manufacturing friction and the EUV choke point.^6,5
• Second, partial manufacturing successes have not translated into a durable AI advantage. Chinese firms continued to improve at the model layer. Public reporting indicates major Chinese models now launch inside increasingly domestic Chinese AI stacks, and the controls have not prevented large-scale circumvention or substitution.^10,11,12
• Third, the regime's enforcement problem is structural, not merely administrative. Chips are fungible, globally traded commodities. Smuggling, shell companies, and third-country transshipment have operated at a scale materially exceeding U.S. enforcement capacity.^13,14
• Fourth, the economic costs to U.S. industry are real and compounding. Public evidence shows market-value losses, revenue disruption, irreversible design-out risk, and lower expected R&D capacity for affected firms.^15,16,17
• Fifth, the strongest case for the current regime lies in the time the controls may have bought. Counterfactual analysis therefore deserves priority. Any resulting delay appears insufficient to count as strategic success on the policy's own terms, given the enforcement costs, industry impacts, and adaptive responses the regime has generated.^18,8

The evidence does not support a claim against all export controls. Broad performance-threshold chip controls do appear to underperform strategically relative to their economic costs, enforcement burdens, and second-order effects. EUV lithography denial remains the strongest exception because supplier concentration, physical scale, and maintenance dependence create an unusually enforceable chokepoint.⁶

The better path combines narrow, genuinely enforceable controls with a broader shift toward ecosystem competition. Software platforms, cloud distribution, talent, standards, and leverage-generating commercial relationships belong at the center of the strategy. The United States is unlikely to win the AI competition by trying to freeze it at the level of hardware access alone. The better instruments are available.

Introduction

On September 16, 2022, National Security Advisor Jake Sullivan outlined a new U.S. approach to advanced technology competition with China. Rather than accept a fixed generational lead, the United States would maintain “as large of a lead as possible” in technologies tied to economic and military power.³ Four weeks later, the Bureau of Industry and Security issued sweeping export controls on advanced computing chips, semiconductor manufacturing equipment, and related technologies.⁴

The strategic logic was simple. If the United States and its allies denied China access to the most capable chips and the tools used to make them, China would struggle to sustain the computational base needed for frontier AI development. Semiconductor denial would preserve AI advantage.

The record is more mixed than either side of the debate usually acknowledges. The 2022 controls and later revisions imposed real friction on Chinese semiconductor development. But the evidence does not show a durable and widening U.S. lead in AI. A narrower conclusion fits better: the regime raised costs for China without producing the strategic return policymakers sought. Some of the regime's most visible second-order effects may also have weakened the long-run position of U.S. firms.

The paper does not argue the controls were a nullity. The paper argues policymakers should judge the regime by strategic return, not by inflicted pain alone. A policy can raise costs and still underperform. The current regime may fit the pattern.

How to judge the controls

Judge the controls by strategic outcome, not by narrow technical pain. The relevant benchmark is whether the controls preserved a meaningful U.S. lead in advanced AI.

Three questions matter. First, did the controls impose real friction on Chinese semiconductor and AI development? Second, did the controls prevent or materially delay the outcomes policy architects set out to prevent? Third, were the benefits of any resulting delay large enough to justify the economic costs, enforcement burdens, and adaptive responses the policy generated?

Export-control debates often collapse into a false binary: either the policy worked because China suffered, or critics are holding the policy to an impossible standard. Both framings evade the core question: were the imposed costs commensurate with the strategic result achieved?

The current regime looks weakest on the third question.

What the controls achieved

The strongest defense of the current policy is simple: the controls did impose friction, and some of the friction is material.

At the manufacturing layer, teardown and process analysis show meaningful cost and yield disadvantages for Chinese firms relative to the global frontier. TechInsights found Huawei's Kirin 9000s required a far more elaborate DUV multi-patterning flow than an EUV-based equivalent. CSIS noted launch-period yield estimates around 50% for SMIC's 7nm-class process, versus roughly 76% for TSMC's 7nm yields.⁵

EUV denial is the clearest success. ASML has never shipped an EUV system to China. The Dutch government blocked ASML's EUV export path years before the September 2023 Dutch measures expanded restrictions on advanced DUV immersion systems, and ASML remains the sole global producer.⁶ The distinction matters. EUV machines are physically large, serializable, heavily serviced, and dependent on ongoing support from a single supplier. Chips are small, fungible, and easy to reroute through intermediaries.

Chinese firms also remain dependent on foreign EDA and lithography. Even optimistic assessments of Chinese progress still show meaningful reliance on non-Chinese design tools and lagging capability at the most advanced lithography frontier.^19,18

EUV denial and manufacturing friction are real achievements. Any serious assessment should say so plainly. The regime's limitation is not a lack of friction. The limitation is that friction at the manufacturing layer has not translated into the durable AI advantage the policy was designed to produce.

What the controls did not achieve

The regime imposed real pain. The question is whether that pain converted into strategic gain at the AI capability layer. The evidence suggests the conversion has been weaker than the policy's architects intended.

Technological adaptation is the first problem. In September 2023, less than a year after the initial rules, TechInsights publicly confirmed Huawei's Mate 60 Pro contained a 7nm-class chip fabricated by SMIC using DUV multi-patterning rather than EUV lithography.⁷ By late 2025, TechInsights reported further progress through SMIC's N+3 pathway.²⁰ The Mate 60 result and later N+3 progress did not erase China's gap with the frontier. Both developments did undermine a central assumption of the denial strategy: blocking frontier tools would prevent meaningful movement at advanced nodes.

Progress at the model layer is the second problem. DeepSeek became the clearest example because the model showed how Chinese developers could remain highly competitive with constrained hardware and unusually efficient engineering choices.^21,9 Public reporting in February 2026 also suggested Chinese firms were launching increasingly important models inside a more domestic Chinese AI stack.¹¹ Readers should treat the reports cautiously. Even so, the reports fit the broader pattern: the controls did not freeze AI competition at the level of hardware denial.

Regulatory ratcheting is the third problem. The 2023, 2024, and 2025 revisions looked like refinements, but the cycle revealed a deeper dynamic: each round of tightening responded to the limits of the previous round.^22,18 BIS closed the A800 and H800 workaround only after the products had already become commercially relevant. U.S. officials let the H20 become Nvidia's final compliant product for China and then swept the H20 into licensing restrictions in 2025, prompting an initial multibillion-dollar charge estimate and, later, a $4.5 billion recorded charge.²³ The pattern does not prove futility. It does show a regime reacting to adaptation faster than policymakers can define a stable strategic end state.

Circumvention is not peripheral

A common defense says the strategy is sound and enforcement is weak. Senate investigators and the Government Accountability Office found material weaknesses in BIS oversight, compliance review, and enforcement capacity.^13,22 But structure matters more than staffing.

Public reporting in late 2025 described Megaspeed, a Singapore-based Nvidia partner, as a major node in suspected China-bound GPU diversion. The reporting involved billions of dollars in hardware and data centers whose inventories did not line up cleanly with sales volumes.¹² Reuters also reported in 2025 that TSMC allegedly faced the prospect of a major penalty after Huawei-associated entities allegedly used shell structures to obtain chip production through the world's most sophisticated foundry.²⁴ In February 2026, Reuters further reported U.S. officials believed DeepSeek had trained a major model on Nvidia Blackwell chips already inside China despite restrictions.¹⁰

No single case settles the question. Together, the cases show why the problem is not simply administrative sloppiness. Chips are global commodities. Shell companies are cheap. Re-export routes adapt quickly. Smuggling can coexist with formal compliance by the original seller. Any regime tracking millions of high-value but transportable items across multiple jurisdictions will face a structural asymmetry between the speed of evasion and the speed of enforcement.

Institutional evidence points the same way. The Senate Permanent Subcommittee on Investigations described BIS enforcement as inadequate “at every level.” The Justice Department's Disruptive Technology Strike Force reported bringing 14 cases in its first year, a number small relative to the scale of documented circumvention.^13,14 The answer is not to abandon enforcement. It is to design the next iteration of the regime around what enforcement can actually deliver, rather than assuming enforcement can scale to match evasion.

Costs to U.S. industry and the adaptation problem

The strategic case would be stronger if the United States were imposing large costs on China at modest cost to itself. The public record does not show such a result.

CSIS and researchers at the New York Fed documented measurable valuation, profitability, credit, and employment effects on firms exposed to the export controls.^15,16 Nvidia's April 2025 H20 disclosure initially pointed to approximately $5.5 billion in charges tied to inventory and supply commitments. Nvidia later recorded a $4.5 billion charge in Q1 FY2026, still a large hit tied to the same policy shock.²³ Other major firms also flagged material exposure, especially in semiconductor equipment.²⁵

The longer-run issue is not just lost revenue. It is design-out. Once Chinese customers rebuild software stacks, data-center procurement plans, and process recipes around domestic or third-country alternatives, U.S. firms will struggle to recover the lost market position even if restrictions later ease.²⁶ The shift is not theoretical. It is already visible in the move away from Nvidia's China-tailored accelerators and in the push by Chinese fabs to diversify away from U.S. equipment where possible.

The policy also appears to have intensified China's domestic mobilization. Big Fund III added another $47.5 billion in registered capital to China's already substantial state-backed semiconductor financing effort.²⁷ CSET documented a broad post-2022 push toward equipment localization, domestic substitution, and expanded state support.⁸ Rhodium Group likewise concluded the controls accelerated stockpiling and indigenization rather than halting Chinese adaptation.¹⁸

China would have pursued self-sufficiency regardless. The question for U.S. policymakers is whether the current regime accelerated that timeline. The evidence suggests it did.

Radiant Intel case study: corroborative operational evidence

Radiant Intel's case study offers original operational evidence from the AI deployment layer the policy was designed to protect. The benchmarks are proprietary and readers should treat them accordingly.

Radiant Intel operates an AI-native intelligence platform for geopolitical, macroeconomic, regulatory, and sanctions-related analysis. The platform uses frontier models for client-facing analytical generation. It uses separate system-tier models for upstream tasks such as data triage, classification, extraction, signal scoring, and structured research workflows.¹

In early 2026, Radiant Intel's internal benchmarking found DeepSeek-V3.2, accessed through Microsoft Azure, performed exceptionally well on a set of system-critical tasks relevant to the platform. On some workloads, DeepSeek-V3.2 outperformed American open-source and smaller closed-source alternatives. The model also improved the economics of high-volume research operations on a cost-adjusted basis.¹ The results did not establish general AI superiority for China. The results also did not displace American frontier models from every layer of the stack. Radiant Intel still prefers American frontier models for some client-facing analytical generation. Even so, the firm's operational experience suggests the deployment gap at the applied AI layer is narrower than a hardware-denial strategy alone would predict.

The benchmarks speak more directly to system-tier performance than to frontier-model leadership. One firm's operational experience does not constitute a dataset. But Radiant Intel builds and deploys these systems professionally, and the results are drawn from production workloads, not synthetic tests.

The case study still matters for two reasons.

First, the case study offers real-world evidence from the workflows the policy directly targeted. If a Chinese open-weight model is already competitive in production intelligence work inside a U.S. platform, readers should view claims of a large and widening AI lead more skeptically.

Second, the case study highlights a second-order effect often missed in the export-control debate: open-weight economics can expand the feasible research frontier. In one recent workflow, Radiant Intel needed roughly 600,000 API calls to score an indicator framework for a conflict-related experiment.¹ Lower-cost open-weight deployment changed not just model choice, but the range of experiments the firm could run at all.

The paper uses the Radiant evidence only as corroborative field evidence. The case study helps show operational competition is closer than a hardware-centric theory of advantage would predict. The case study does not, by itself, settle the strategic question. It does suggest that the strategy's theory of advantage needs updating.

The counterfactual question

The hardest question is not whether China absorbed costs. The hardest question is whether the controls bought enough strategically meaningful time to justify the costs.

The wrong counterfactual asks whether China would have pursued semiconductor self-sufficiency without the controls. Of course it would have. Big Fund I, Big Fund II, and Made in China 2025 all predated October 2022.^27,18 Three narrower questions matter.

First, did the controls delay Chinese progress? Probably yes, at least in some areas. Evidence on EUV denial, higher process costs, lower yields, and continued dependence points toward delay.^5,6,19

Second, did they prevent Chinese progress in advanced semiconductors and AI? The evidence assembled here suggests no. Chinese firms continued moving down the manufacturing curve through DUV multi-patterning, producing competitive AI systems, and accessing restricted hardware through both lawful and unlawful channels.^7,10,21

Third, did any achieved delay generate enough strategic value to outweigh the costs and second-order effects? The case for the current regime is weakest there. The delay appears partial rather than decisive. Enforcement appears structurally porous. The policy also appears to have reinforced supplier substitution, state mobilization, and domestic ecosystem development inside China.^8,18,26

The conclusion is straightforward. Analysts should not treat the controls as a total nullity. Analysts should treat them as a policy whose strategic returns appear insufficient relative to their economic costs, enforcement burdens, and counterproductive adaptive effects.

Policy options

Policymakers face at least three plausible options.

Option 1: Preserve and tighten the current regime

Option 1 would keep broad performance-threshold chip controls in place, expand enforcement resources, add more entities, and push for additional multilateral tightening. The option preserves friction and projects a visible posture of technological denial. Its weakness is the assumption that enforcement improvements alone can overcome the structural porosity of broad chip controls. The evidence reviewed here suggests otherwise.

Option 2: Narrow and redesign the regime

Option 2 would retain only controls exploiting unusually favorable enforcement conditions, above all EUV denial, while reducing reliance on broad chip controls targeting small, fungible, globally traded items. Its strength is the alignment between policy and the distinction between enforceable and weakly enforceable chokepoints. Its weakness is the political concession: a broad denial architecture is easier to sell.

Option 3: Pivot from denial-centric policy to ecosystem competition

Option 3 would treat hardware denial as a supporting instrument rather than the center of gravity. The strategic emphasis would shift toward software platforms, cloud distribution, talent, standards, open-model strategy, and commercial relationships creating long-run leverage. Its strength is simple: durable advantage is more likely to emerge there. Its weakness is political. Sustained policy work is harder than announcing another visible prohibition.

Recommendation

The United States should combine Options 2 and 3.

The United States should retain narrow controls where enforcement conditions are unusually favorable, most importantly EUV lithography denial. It should end the current overreliance on broad performance-threshold chip controls as the centerpiece of AI competition with China. It should resist the temptation to expand equipment controls into every adjacent category simply because DUV adaptation and smuggling exposed the limits of the existing regime. It should redirect political capital toward ecosystem competition.

Five moves follow.

• Preserve EUV denial as a narrow, time-limited hold while the chokepoint remains real.⁶
• De-emphasize broad chip controls with a record of porosity, reactivity, and economic cost.^16,15
• Focus on software and cloud ecosystems, where U.S. firms still hold stronger global positions.
• Treat talent policy as a strategic variable rather than a domestic afterthought.
• Evaluate technology restrictions by strategic return, not by the volume of pain imposed.

In practical terms, an ecosystem strategy means reinforcing CUDA-adjacent developer lock-in where it still exists and preserving U.S. cloud platforms as the default distribution layer for global AI deployment. It also means expanding visa pathways for top technical talent, coordinating standards and safety frameworks with allies, and supporting open-model and applied-AI development inside the United States. Policymakers can announce the agenda less dramatically than a new control threshold, but the agenda fits the arena now deciding AI competition better.

The recommendation rests on a hard-headed assessment: hardware parity may prove difficult to prevent indefinitely, even with optimally designed controls. This is not defeatism. It is the starting point for a strategy that can actually win. The United States does not most need to win a static contest over access to a specific chip. It needs to win the broader contest over ecosystems, deployment, and institutional capacity.

Conclusion

The broad performance-threshold chip controls imposed in 2022 were not pointless. They imposed real costs on China and continue to constrain some parts of the Chinese semiconductor stack. But policymakers did not justify the controls simply as a way to make Chinese firms pay more. Policymakers justified the controls as a way to preserve a durable and widening U.S. lead in advanced AI.

Measured against the policy's own benchmark—a durable and widening U.S. lead in advanced AI—the controls have underperformed.

Chinese firms continued to advance in advanced-node manufacturing through DUV workarounds. Chinese firms also improved at the model layer. Circumvention and transshipment proved persistent. U.S. firms absorbed meaningful economic costs. Chinese policymakers intensified domestic substitution and equipment localization. Radiant Intel's operational experience likewise suggests the deployment gap is narrower than a hardware-denial strategy promised.

A sound judgment rejects both extremes. The policy was not a total failure, and not every export control is futile. The current chip-control regime does appear strategically underperforming relative to its costs, while narrow chokepoints such as EUV denial remain worth preserving.

The policy debate should now turn in a more serious direction. Hardware denial can impose pain. The real question is where the United States can still build durable advantage. Durable advantage is more likely to reside in ecosystems, talent, software, cloud distribution, standards, and commercial leverage than in any attempt to govern AI competition primarily through chip thresholds.

The United States can win the AI competition. It is unlikely to do so by trying to freeze it at the level of hardware access. The better instruments are available. The question is whether policymakers will use them.

Notes

1. ↑ Radiant Intel internal benchmark results and deployment notes, January–February 2026, on file with the authors. These results are proprietary and not independently reproducible. They compare model performance on system-tier tasks including data curation, extraction, classification, signal scoring, and structured research workflows.
2. ↑ Radiant Intel also has a commercial interest in lower-cost, higher-performing model ecosystems, including open-weight deployment. The note discloses the commercial interest because the interest could shape incentives at the margin.
3. ↑ Jake Sullivan, “Remarks by National Security Advisor Jake Sullivan at the Special Competitive Studies Project Global Emerging Technologies Summit,” The White House, September 16, 2022, bidenwhitehouse.archives.gov.
4. ↑ Bureau of Industry and Security, “Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items,” 87 Fed. Reg. 62186, October 13, 2022, federalregister.gov.
5. ↑ TechInsights, “HiSilicon Kirin 9000s (SMIC 7nm, N+2) Process Flow Analysis,” September 2023, techinsights.com; Gregory C. Allen, “In Chip Race, China Gives Huawei the Steering Wheel,” Center for Strategic and International Studies, October 6, 2023, csis.org.
6. ↑ ASML, Annual Report 2024, 49, asml.com; Reuters, “ASML Has Had No EUV Machine Sales to China for Several Years, CEO Says,” January 25, 2023; Associated Press, “Dutch Set to Restrict More Exports to China of Advanced Chipmaking Equipment,” March 8, 2023; CNBC, “ASML Blocked from Exporting Some Critical Chipmaking Tools to China,” January 2, 2024.
7. ↑ TechInsights, “TechInsights Finds SMIC 7nm (N+2) in Huawei Mate 60 Pro,” September 2023, techinsights.com.
8. ↑ Jacob Feldgoise and Hanna Dohmen, “Inside Beijing's Chipmaking Offensive,” Georgetown Center for Security and Emerging Technology, December 19, 2025, cset.georgetown.edu.
9. ↑ John Villasenor, “DeepSeek Shows the Limits of US Export Controls on AI Chips,” Brookings Institution, January 29, 2025, brookings.edu.
10. ↑ Reuters, “Exclusive: China's DeepSeek Trained AI Model on Nvidia's Best Chip Despite US Ban, Official Says,” February 23, 2026, reuters.com.
11. ↑ South China Morning Post, “Zhipu AI Breaks US Chip Reliance with First Major Model Trained on Huawei Stack,” February 11, 2026, scmp.com; Z.ai, “GLM-5,” February 2026, z.ai.
12. ↑ Bloomberg, “Nvidia Partner Megaspeed Draws China Chip Smuggling Concerns in US,” December 22, 2025, bloomberg.com.
13. ↑ U.S. Senate Permanent Subcommittee on Investigations, “The U.S. Technology Fueling Russia's War in Ukraine: Examining the Bureau of Industry and Security's Enforcement of Semiconductor Export Controls,” December 18, 2024, hsgac.senate.gov.
14. ↑ U.S. Department of Justice, “Fact Sheet: Disruptive Technology Strike Force Efforts in First Year to Prevent Sensitive Technology from Being Acquired by Authoritarian Regimes and Hostile Nation-States,” February 16, 2024, justice.gov.
15. ↑ Kirti Gupta, Chris Borges, and Andrea Leonard Palazzi, “Collateral Damage: The Domestic Impact of U.S. Semiconductor Export Controls,” Center for Strategic and International Studies, 2024, csis.org.
16. ↑ Matteo Crosignani, Lina Han, Marco Macchiavelli, and Andre F. Silva, “Securing Technological Leadership? The Cost of Export Controls on Firms,” Federal Reserve Bank of New York Staff Reports, no. 1096, revised February 2025, newyorkfed.org.
17. ↑ Trelysa Long, “Decoupling Risks: How Semiconductor Export Controls Could Harm US Chipmakers and Innovation,” Information Technology and Innovation Foundation, November 10, 2025, itif.org.
18. ↑ Rhodium Group, “Slaying Self-Reliance: US Chip Controls in Biden's Final Stretch,” December 2024, rhg.com.
19. ↑ Karen M. Sutter, “U.S. Export Controls and China: Advanced Semiconductors,” Congressional Research Service, Report R48642, September 19, 2025, congress.gov.
20. ↑ TechInsights, “SMIC N+3 Confirmed: Kirin 9030 Analysis Reveals How Close SMIC Is to 5nm,” December 2025, techinsights.com.
21. ↑ DeepSeek-AI, “DeepSeek-V3 Technical Report,” arXiv:2412.19437, December 2024, arxiv.org.
22. ↑ U.S. Government Accountability Office, “Export Controls: Commerce Implemented Advanced Semiconductor Rules and Took Steps to Address Compliance Challenges,” GAO-25-107386, December 2, 2024, gao.gov.
23. ↑ Nvidia, Form 8-K, April 15, 2025, stating Nvidia expected first-quarter results to include “up to approximately $5.5 billion” in H20-related charges, sec.gov; Nvidia, quarterly filing noting the company “incurred a $4.5 billion charge” in Q1 FY2026 related to H20 inventory and purchase obligations, investor.nvidia.com; CNBC, “Nvidia Says It Will Record $5.5 Billion Quarterly Charge Tied to H20 Processors Exported to China,” April 15, 2025.
24. ↑ Reuters, “Exclusive: TSMC Could Face $1 Billion or More Fine from US Probe, Sources Say,” April 8, 2025, reuters.com.
25. ↑ Reuters, “Lam Research Warns of Up to $2.5 Billion Revenue Hit from U.S. Curbs on China Exports,” October 19, 2022, reuters.com.
26. ↑ William Alan Reinsch, Jack Whitney, and Matthew Schleich, “The Double-Edged Sword of Semiconductor Export Controls,” Center for Strategic and International Studies, October 4, 2024, csis.org.
27. ↑ Che Pan, “Tech War: China Doubles Down on Semiconductor Self-Sufficiency Drive with US$47.5 Billion Big Fund III,” South China Morning Post, May 27, 2024, scmp.com; Qin Min and Han Wei, “China Piles $47.5 Billion into ‘Big Fund III’ to Boost Chip Development,” Caixin Global, May 28, 2024, caixinglobal.com.