Perspective | A New Approach to US Export Controls on AI and Advanced Computing Integrated Circuits to China: Insights from the 'Artificial Intelligence Diffusion Framework'

(Image information source: U.S. Federal Register)
 

(Image information source: U.S. Federal Register)
 

According to this interim final rule, the Bureau of Industry and Security (BIS) of the Department of Commerce revised the Export Administration Regulations (EAR) control measures for advanced computing integrated circuits (ICs) and added controls over certain advanced closed-source dual-use AI model weights. In conjunction with these control measures deemed necessary by BIS to protect U.S. national security and foreign policy interests, BIS added multiple license exception clauses and updated the data center validation end-user authorization mechanism to allow for the export, re-export, and in-country transfer of advanced computing integrated circuits (ICs) to destinations that do not pose national security or foreign policy risks.

(Image information source: U.S. Federal Register)
 

Shortly after the release of the Framework, the Bureau of Industry and Security (BIS) of the U.S. Department of Commerce issued "Implementation of Additional Due Diligence Measures for Advanced Computing Integrated Circuits" (hereinafter referred to as the "Measures"), which took effect on January 16, 2025. The BIS's consecutive release of two major interim new export control regulations demonstrates its efforts to expand its jurisdiction in the semiconductor field under the existing Export Administration Regulations (EAR) framework.

(Image information source: U.S. Federal Register)
 

The Framework details the impact of AI diffusion on national security and foreign policy in the section "IMPACT OF AI DIFFUSION ON NATIONAL SECURITY AND FOREIGN POLICY", arguing that the global spread of advanced AI models will affect U.S. national security and foreign policy interests in three ways.

First, exporting advanced AI models or their means of production to foreign countries will increase the risk of technology being stolen or transferred to countries/end-users of concern. This risk exists even if the recipient is a verified entity and the destination is a U.S. ally with a robust export control system. This risk is particularly prominent in cutting-edge AI technology (i.e., the largest and most advanced models of each period).

Second, to realize the full economic and social benefits of advanced AI models, it is necessary to allow foreign verified entities to obtain a large number of advanced computing chips or the AI models themselves. These entities may discover innovative applications that American companies cannot independently develop, promoting global benefits. According to the Export Control Reform Act of 2018 (ECRA), export controls should be "precisely targeted at core technologies that may pose a significant national security threat" (50 U.S.C. 4811(2)(G) and (5)). When formulating control measures, BIS needs to balance national security concerns with the benefits of foreign entities obtaining technology. For regions where there is a risk of technology transfer, risks can be controlled through appropriate mitigation measures.

Third, maintaining global technological leadership in the AI industry is a core national security interest. ECRA requires continuous assessment of the impact of export controls on technological leadership (50 U.S.C. 4811(3)). BIS believes that to maintain its advantage in the AI field, U.S. developers may need to: ① build large advanced computing chip clusters abroad; ② store models overseas to ensure quality service for verified customers.

In summary, BIS has determined that it is necessary to establish a precise global regulatory mechanism for the diffusion of AI models, which must control national security risks, ensure that technological benefits are realized, and maintain U.S. technological leadership.

A careful analysis of the original text shows that the Framework summarizes at least three regulatory principles in this section: first, a "tripartite" approach, coordinating the three goals of national security, economic benefits, and technological leadership; second, precision policy-making, establishing a differentiated and dynamically adjustable export control system; and third, global cooperation, building an international technology governance framework that takes into account both security and development.

After reviewing the framework's table of contents, it can be seen that this Framework mainly includes two aspects: expanding the control of advanced computer integrated circuits and the new control of AI model weights.

In addition, the Framework details in the section "MEANS FOR RESPONSIBLY CONTROLLING AI DIFFUSION": ① The nature of AI models, namely software programs composed of mathematical operations, whose output quality is determined by architectural design. ② Model development process, namely architecture design → code writing → data training → weight optimization. ③ Industry development trends: first, the surging demand for computing power, requiring advanced computing IC clusters to process massive parameters; second, the driving force, computing power/data volume is exponentially increasing (increased by several orders of magnitude in the past decade); third, future planning, leading companies plan to build computing power clusters tens of times larger than the current scale. ④ Key value judgments: pointing out that model weights have core intellectual property value, computing resources constitute fundamental strategic resources for AI development, and global diffusion control requires grasping both hardware foundations and software cores. ⑤ Control suggestions: proposing a dual control mechanism, on the one hand suggesting upgrading existing controls to restrict the export/re-export/in-country transfer of advanced computing ICs, and on the other hand adding new control targets, incorporating AI model weights into the scope of export controls.

After summarizing the structure of AI models and industry development trends, the Framework believes that the global diffusion of advanced AI models can be effectively regulated through the following methods: (1) expanding the existing control over the export, re-export, and in-country transfer of advanced computing integrated circuits needed to build large training clusters; (2) adding control over the export, re-export, and in-country transfer of the weights of the most advanced AI models.

The Framework details a customized strategic path for controlling the diffusion of advanced AI models in the section "Framework for the Diffusion of Advanced AI Technologies." The overall goal is to ensure that: the most advanced U.S. AI model weights can only be stored abroad under strict security conditions; and large advanced integrated circuit clusters needed to train these models can only be built in regions with lower risks of transfer/abuse. This section introduces the following three regulatory strategies:

(Image information source: U.S. Federal Register)
 

The Framework requires a license application for the export, re-export, or in-country transfer of advanced computing integrated circuits or the weights of the most advanced AI models to any end-user in any region. The U.S. government will review applications based on destination sensitivity, computing power level/model performance, and the recipient's security commitments. Global license requirements are crucial for preventing the flow of relevant items to regions or users of concern. The Framework also mentions that credible open-source reports show that Chinese companies purchase integrated circuits controlled by the Export Administration Regulations (EAR) through overseas subsidiaries established in non-controlled regions. Once model weights are maliciously stolen, they can be instantly copied and spread globally, posing an even greater risk. Therefore, U.S. national security and foreign policy require BIS to rigorously review transactions flowing to high-risk regions or users. To this end, BIS will adopt a "presumptive denial" licensing policy for large-scale advanced computing ICs needed to train advanced AI models. For low-risk regions or users, the global licensing mechanism can reduce overall risk through additional conditions. The Framework believes this is the most effective way to address the multiple risks of exporting advanced computing ICs and model weights.

(Image information source: U.S. Federal Register)
 

To reduce the compliance burden on low-risk regions/users and promote economic activity, BIS has implemented flexibility mechanisms in its licensing requirements. For example: providing licensed exceptions with security measures for verified end-users or very low-risk regions; allowing users in other regions to obtain "verified" qualifications through verification procedures to simplify the procurement process; exempting the export of weights of open-weight models (see the "Report on Dual-Use Foundational Models with Widely Accessible Weights," July 30, 2024) from licensing. These mechanisms ensure that controls target only the highest-risk areas of AI frontier development, without affecting the vast majority of AI technology circulation, aligning with the policy objectives of the Export Control Reform Act (ECRA), which seeks to restrict the export of items that "make a significant contribution to the military potential of countries that threaten U.S. national security" while maintaining U.S. leadership in the fields of science, technology, engineering, and manufacturing (50 U.S.C. 4811(3)).

(Image information source: U.S. Federal Register)
 

As the third part of the strategy, BIS will impose security conditions to protect the security of advanced models stored in high-risk regions and reduce the risk of advanced IC transfer. These conditions will protect U.S. national security and ensure that advanced models and large IC clusters stored abroad are protected against transfer/abuse.

To implement this three-step strategy, the Framework also introduces seven specific implementation paths, updating the ECCN system, adding 4E091 to specifically control AI model weights, revising control clauses in the 3A090/4A090 series, and optimizing procedures, expanding the ACA country list, and restructuring the NAC reporting process to improve efficiency. This part is omitted in this article.

The Framework points out that by responsibly promoting advanced AI diffusion control in stages, multiple strategic goals can be achieved: First, reducing the access of countries/users of concern to the most advanced AI models or development capabilities; second, fully assessing the risks of each generation of AI and developing security measures before the global spread of model weights and large IC clusters; third, focusing on the precise control of frontier AI, maintaining the basic smooth flow of AI models and IC commerce; fourth, maintaining global access to the most advanced AI capabilities through user applications and API structured mechanisms (this rule does not set new restrictions).

By reading this chapter,the regulatory strategic framework of the Framework can be summarized as follows:

First, setting dual control objects: "weights of the most advanced AI models" for core assets and "advanced IC clusters required for training" for infrastructure.

Second, establishing review principles: First, introducing a three-element assessment, namely "destination risk level, computing power scale/model performance, and security commitment." Second, setting tiered licensing, namely "presumptive denial" for high-risk regions and conditional licensing for low-risk regions.

Third, policy flexibility design: First, introducing exemption mechanisms, namely licensing exemption for open-weight models (must meet the definition in the "Dual-Use Foundational Model Report"), licensed exceptions applicable to verified end-users/low-risk countries, and establishing an "verification qualification" application channel to optimize the compliance path. Second, setting exception categories, adding three exceptions: AIA/ACM/LPP to cover different scenarios, and splitting verification authorization to improve management granularity.

It can be seen that the regulatory strategic value concept of the Framework, on the one hand, conducts risk control, establishes a "buffer period" for technology diffusion to improve the security system, and maintains technological leadership through structured access (API/application layer opening), and on the other hand, also wants to maintain a balance of interests, precisely targeting frontier risks, maintaining the free flow of more than 90% of AI technology, and balancing national security and commercial interests.

In the "Expanded Advanced Computing Integrated Circuit Controls" chapter, the Framework introduces the following three regulatory measures:

1. New Global License Requirements

(Image information source: U.S. Federal Register)
 

BIS implements global license requirements for ECCN 3A090.a, 4A090.a, and their corresponding .z items (see the added clause § 742.6(a)(6)(iii)(A)). The global license requirements for these items (including items covered by the EAR's "Advanced Computing Foreign Direct Product Rule (FDPR)") aim to protect U.S. national security and foreign policy interests by: (1) allowing BIS to review transactions that pose a risk of transfer or misuse; and (2) enabling the U.S. government to track the end use, end user, and geographic location of advanced ICs. In addition, through data center VEU authorization, global companies can allocate AI computing resources in a verified secure environment.

The regulatory approach reflected in this section of the Framework is:First, refinement of ECCN classification, dividing 3A090.a/b and 4A090.a/b by TPP and performance density, clarifying the control levels corresponding to different performance thresholds. Second, rule adjustment, splitting regional stability control clauses, strengthening global regulation of high-performance ICs (3A090.a), canceling some license exceptions (such as GBS, LVS), and tightening the export of sensitive items in Group B countries. Third, FDPR expansion, expanding the scope of application of the Foreign Direct Product Rule from specific country groups to the world, strengthening supply chain monitoring. Fourth, supporting measures, allocating AI computing power within a secure framework through data center VEU authorization, balancing technology output and risk control.

2. License Exceptions

(Image information source: U.S. Federal Register)

BIS has added three license exceptions to the EAR applicable to advanced computing integrated circuits (ICs): § 740.27 Artificial Intelligence Authorization License Exception (AIA), § 740.28 Advanced Computing Manufacturing License Exception (ACM), and § 740.29 Low Processing Performance License Exception (LPP). In addition, BIS has updated the target scope of the "Advanced Computing Authorization License Exception (ACA)" and optimized the approval process for "Notification-type Advanced Computing (NAC)", requiring NAC users to provide more information to improve processing efficiency. A brief summary of AIA, ACM, and LPP is as follows:

In summary, the main content of AIA is to allow exports to low-risk countries (such as Australia, Canada, Japan, etc.), and for exports of AI-related advanced computing ICs, the end use and user compliance must be strictly reviewed; the main content of ACM is to support the private enterprise manufacturing supply chain and restrict flow to high-risk regions; and the main content of LPP is to allow limited exports of low-performance computing ICs, setting an annual TPP upper limit.

The regulatory approach reflected in this part of the Framework is mainly as follows:First, strengthening certification and reporting. Both AIA and LPP require end-users to submit compliance commitments, and BIS implements mandatory reporting for high-TPP transactions; ACM requires companies to establish supply chain tracking systems. Second, optimization of existing license exceptions. ACA is extended to a global scope (excluding high-risk regions), and NAC adds technical parameter review and user background investigation to accelerate the approval process.

3. Data Center Verified End Users

(Image information source: U.S. Federal Register)
 

In October 2024, the Bureau of Industry and Security (BIS) of the U.S. Department of Commerce revised the Export Administration Regulations (EAR), extending Verified End-User (VEU) authorizations to data center VEUs (DC VEUs). As stated in this rule, data centers are critical to global AI development, and the United States is committed to promoting international AI cooperation while controlling national security risks. This revision further subdivides DC VEUs into general VEUs (UVEUs) and national-level VEUs (NVEUs). Enterprises whose headquarters or ultimate parent company are located in the countries listed in Supplement No. 5(a) to Part 740 may apply for UVEU authorization; enterprises whose headquarters or ultimate parent company are located globally (excluding Macao and countries in Group D:5) and require large-scale import/export/re-export of advanced computing integrated circuits may apply for NVEU authorization under restrictive conditions.

According to Supplement No. 10, paragraph 6, to Part 748, all UVEUs must comply with AI computing power geographic distribution restrictions (measured by the total processing performance TPP of chips that meet the 3A090.a standard): ① UVEUs headquartered in countries listed in Supplement 5(a): The proportion of their global AI computing power (including affiliated enterprises) transferred/deployed to non-5(a) countries shall not exceed 25%, and the proportion for a single non-5(a) country shall not exceed 7%; ② UVEUs with headquarters in the United States: The proportion of deployment outside the United States shall not exceed 50%.

The regulatory approach reflected in this section of the framework is primarily:

First, an upgrade to the policy framework, adding a data center VEU classification, namely UVEU (general purpose) and NVEU (national level). UVEUs apply to enterprises headquartered in "whitelisted" countries (Supplement 5(a)), while NVEUs apply to enterprises headquartered outside of Group D:5 countries and Macao.

Second, strengthened compliance requirements: First, the owner of the computing assets bears the primary application responsibility and must disclose information about cooperating entities; second, when the asset owner and operator are separate, both must obtain VEU authorization; third, a multi-layered written assurance mechanism is established.

Third, setting geographic restrictions on computing power, including cross-border deployment of UVEU computing power, i.e., the total amount in non-whitelisted countries ≤ 25% (≤ 50% for US-headquartered companies), and a single non-whitelisted country ≤ 7%, and the implementation of a per-country quarterly quota system for NVEUs (2025-2027).

Fourth, setting technical control standards, using TPP (Total Processing Performance) as a quantitative indicator, controlling items that meet the 3A090.a/4A090.a standard for advanced computing chips, and setting quotas based on technology that lags behind the leading-edge AI training clusters by one generation.

Fifth, setting a dynamic adjustment mechanism, including an annual review system for quotas after 2027, and joint assessments by multiple departments (State Council/Ministry of Energy/Ministry of National Defense).

Sixth, setting special exemption clauses, including quota reductions for equipment losses (damage/scrapping/resale, etc.), and allowing multiple NVEUs (independently calculated quotas) in the same country.

In summary, in this section, the framework attempts to establish a long-term mechanism for controlling national security risks while safeguarding the competitiveness of the US AI industry through tiered authorization, quantitative control of computing power, and technological generation delays.

In the section "Overview of New Controls for AI Model Weights," the framework introduces the following regulatory measures:

1. Technical Control of AI Model Weights

(Image information source: U.S. Federal Register)
 

In the Commerce Control List (CCL), this rule adds ECCN 4E091 to control the "parameters" of advanced AI models. These models are defined as those trained using 1026 or more "operations." "Parameters" refer to any values learned during training (such as network weights, biases, etc.). "Operations" include subsequent training (such as fine-tuning pre-trained models), but do not include the collection and organization of input training data.

This ECCN, due to regional stability (RS) reasons, controls exports, re-exports, and in-country transfers to all destinations globally, enforced through the addition of §742.6(a)(13). When the destination is a country requiring an AT:1 license (i.e., Iran §742.8, Syria §742.9, or North Korea §742.19), anti-terrorism (AT) controls will also apply. See also other control measures in Parts 744 and 746 of the EAR. For license applications under RS controls, all destinations other than those listed in Supplement No. 5(a) to Part 740 will be reviewed under a "presumption of denial" principle; destinations listed in Supplement No. 5(a) to Part 740 will be reviewed under a "presumption of approval" principle. This control does not apply to deemed exports/re-exports by employees of companies headquartered in the United States or in the countries listed in Supplement No. 5(a) to Part 740.

It is important to note that ECCN 4E091 exempts open-weight models. Note 1 states that the parameters of AI models that have been "published" (as defined in §734.7(a)), or additional training on published parameters that does not exceed 2 × 1025 operations (or 25% of the training operations defined in Note 2, whichever is higher), are not subject to control. Note 2 clarifies that if the training operations of an AI model are less than the training operations of the most advanced published AI model determined through comprehensive benchmark testing, it is also not subject to control.

In summary, the framework in this section identifies the controlled objects as AI model parameters (including network weights, biases, etc.) with training operations ≥ 1026, including fine-tuning and other subsequent training, but excluding the data collection phase.

The framework defines the scope of control as: ① Global RS control (due to regional stability); ② Additional AT control for Iran/Syria/North Korea (due to anti-terrorism); ③ License review: "presumption of denial" for non-designated countries, "presumption of approval" for designated countries (Supplement No. 5(a) to Part 740); ④ Exemptions: Deemed exports by employees of US and designated country companies.

The exceptions stipulated by the framework include: open-weight models, parameters of "published" models, additional training on published models (≤ 2 × 1025 operations or ≤ 25% of the original training amount), performance below the strongest open-weight model (determined by the AI Security Institute and the Department of Energy), and training volume below the benchmark of the most advanced published model.

2. Foreign Direct Product Rule for AI Model Weights

(Image information source: U.S. Federal Register)
 

To address national security and foreign policy risks associated with the production of AI model weights outside the United States, this rule adds a Foreign Direct Product Rule (FDPR) in §734.9(l). Under this AI model weight FDPR, items classified under the new ECCN 4E091 will be subject to global jurisdiction under the Export Administration Regulations (EAR) based on product scope standards. Licensing requirements are consistent with ECCN 4E091, meaning that the establishment of jurisdiction solely due to this FDPR does not automatically imply a license requirement; the specific requirements must be considered in conjunction with the licensing requirements of ECCN 4E091 and the end-use and end-user requirements of Part 744 of the EAR.

The product scope requirements of this FDPR are: Item 4E091 must be produced by a complete factory located outside the United States or its "major components," and the factory or its "major components" (whether produced in the United States or foreign countries) are subject to EAR and belong to items listed in ECCN 3A001.z, 3A090, 4A003.z, 4A004.z, 4A005.z, 4A090, 5A002.z, 5A004.z, or 5A992.z. Integrated circuits, servers, and other electronic devices in these ECCNs are considered "major components" of the factory and are crucial to the production of the 4E091 model weights. The rules specifically state: ECCN 4E091 includes foreign-produced items that have undergone subsequent training through fine-tuning, quantization, etc. However, if subsequent training is conducted using items not listed in the aforementioned ECCNs, it will not change the EAR jurisdiction status of item 4E091.

In summary, in this section, the Framework establishes that the controlled object is the AI model weights under ECCN 4E091. The core mechanism is to establish a new FDPR global jurisdiction system covering AI model weights produced overseas using US technology, applying the principle of tracing "major components" (including servers, chips, and other key equipment).

3. Exception License Eligibility for AI Model Weights

(Image information source: U.S. Federal Register)
 

For the newly established ECCN 4E091 category, the only applicable license exception is the newly established "AIA License Exception" in Section 740.27. The AIA License Exception authorizes the export, re-export, and in-country transfer of items under ECCN 4E091 category to entities located in the destinations listed in Supplement No. 5(a) to Part 740. This exception also allows the export, re-export, and in-country transfer of artificial intelligence model weights (including the most advanced AI models specified in ECCN 4E091) to entities that meet the following conditions: the entity is headquartered in or the ultimate parent company is headquartered in the destinations listed in Supplement No. 5(a), and the entity's location is not in Macao or belongs to Country Group D:5. The AIA License Exception also applies to certain eligible commodities, software, and technology. However, this exception may not be used to export, re-export, or transfer in-country items under ECCN 4E091 and other designated ECCN categories to entities whose headquarters or ultimate parent company headquarters are not located in the destinations listed in Supplement No. 5(a).

In summary, in this section, the Framework establishes that the exception license eligibility for AI model weights applies only to items under ECCN 4E091 category and covers the entire export/re-export/in-country transfer process. The core authorization is to allow the export of AI model weights, including the most advanced AI models, to entities in countries listed in Supplement No. 5(a), and it also applies to related commodities/software/technology. In addition, the Framework further clarifies geographical restrictions in this section, namely that the entity must meet a) headquarters located in the authorized country list or b) ultimate parent company headquarters located in the authorized country list, explicitly prohibiting the application areas of the Macao Special Administrative Region and countries listed in Country Group D:5. This license exception does not apply to entities whose headquarters and parent company headquarters are not on the authorized country list, maintaining export controls for non-listed countries.

4. Red Flag Guidance for AI Model Weights

(Image information source: U.S. Federal Register)
 

As a supplement to the "Know Your Customer" guidelines and red flag provisions in Supplement No. 3 to Part 732 of the Export Administration Regulations (EAR), BIS has added a 28th red flag through an interim final rule to provide compliance guidance for exporters, re-exporters, and transferors involved in AI model weights. This warning aims to help "Infrastructure as a Service" (IaaS) cloud service providers in the United States identify the following risk situations: when training advanced AI models (under the control of ECCN 4E091) for foreign entities' subsidiaries in the United States and transferring the generated model weights to the customer, there is a substantial risk that the model weights may be exported without authorization.

BIS recommends that exporters, re-exporters, transferors, and IaaS providers serving domestic customers add the following measures to their compliance plans: verify whether the relevant model weights will be exported to countries requiring licenses; if there are license requirements, apply for a license or inform the customer that they need to complete the export license application process in advance.

It can be seen that this clause reflects the strict control of the United States over the output of core AI technologies, focusing on preventing technology transfer behaviors that circumvent export controls through a subsidiary structure.

Through the above brief梳理 of the main regulatory strategies of the Framework, and in combination with the full text of the Framework, it can be seen that the Framework issued by the United States lists China as "Tier 3 (fully embargoed country)", that is, all exports of high-end AI chips, model weights, and cloud computing services to China are completely blocked. The United States is attempting to reshape the global AI order through "allied grading + computing power quotas".

In terms of controlled objects, the Framework has constructed a tripartite technological regulatory system, implementing a three-dimensional blockade from the physical layer (chips), the algorithm layer (models), and the service layer (cloud computing power), forming a control chain that runs through all elements of the AI industry:

First, a full-domain blockade at the hardware layer, establishing a computing power threshold of 990 TFLOP/s, systematically blocking the supply channels of high-end AI acceleration chips to China. By activating the "foreign direct product rule" (FDPR) and the "extension of technological sovereignty" mechanism, the control boundary is extended from physical chips to the global semiconductor industry chain. Any chip manufacturing behavior containing US-based technology is included in the scope of jurisdiction, achieving extraterritorial legalization of technological blockade.

Second, dynamic interception at the algorithm layer, creating a "supply chain of algorithms" regulatory paradigm, including closed model weights with training costs exceeding US$70 million in the list of strategic materials, clearly stipulating that they must not be transferred to China. A technology evolution synchronization mechanism is established, automatically triggering control upgrades when the training cost of open-source models exceeds US$30 million, and retaining the right to trace and supervise open-source models. Through a mandatory reporting system, cross-border cloud computing power calls are brought into the regulatory field of vision, forming a full-cycle technological isolation of algorithm research and development.

Third, intelligent blocking at the service layer, constructing an AI computing power firewall system, and forcing cloud service providers to deploy an intelligent monitoring matrix. This system integrates multi-dimensional verification mechanisms such as network fingerprint identification (based on IP geographic fencing), hardware identity traceability (chip DNA tracking, network latency, etc.), and real-time blocking of the cross-border flow of controlled computing resources. An accompanying "compliance black box" audit system implements penetrating supervision of cloud service providers, and any abnormal flow of computing power will trigger the qualification revocation process, undermining the possibility of technological breakthroughs from the server side. This three-pronged blockade strategy of "chip-algorithm-computing power" marks the entry of technological suppression into a new stage of the AI supply chain war.

In terms of control strategies, the Framework divides technology access levels and introduces a computing power quota mechanism, aiming to define the accessible boundaries of AI technology and attempt to control the growth rate of global AI infrastructure.

First, a tiered technology access system is set up, dividing it into three technology access tiers, namely:Tier 1This framework categorizes countries into three tiers: Tier 1 (core partners), Tier 2 (conditional partners), and Tier 3 (comprehensive embargo countries). Each tier has different access levels to AI technologies, chips, and computing power. For Tier 2 countries, there are technical access red lines, specifying that partners using Chinese AI chips or large language models will lose access to US technology supply chains. This design creates a "technology choice dilemma," forcing cooperating countries to make an exclusive choice between the two major technological systems, thus strengthening the centripetal force of the US technology alliance and consolidating US dominance in the global AI system.

Second, it introduces a computing power quota mechanism to control the growth rate of computing power: 1) Infrastructure deployment ceiling: Limiting the scale of computing infrastructure deployed by US companies such as Microsoft and Google in Tier 2 countries to no more than 7% of their global deployment total. This infrastructure layout constraint curbs the risk of technology diffusion and ensures that Tier 2 countries cannot obtain excessive high-end computing power; 2) Dynamic control of hardware imports: Implementing an annual import quota system for H100 chips in Tier 2 countries (100,000 units in 2025) and setting a computing power annual growth rate threshold (≤50%), forming a dual-dimensional control of "total quantity control + growth rate adjustment." This composite control mechanism guarantees the computing power supply of the technology alliance, while ensuring that the evolution pace of global AI infrastructure is synchronized with US strategic planning through a precisely designed growth curve.

Through the synergistic effect of technology access and computing power control, this system achieves three strategic goals: solidifying the technology alliance, containing competitors, and controlling developmental dominance. It not only builds an exclusive technological ecosystem but also uses quantitative control measures to bring global AI development into a predictable trajectory, providing institutional safeguards for the US to maintain technological hegemony.

In summary, from a policy perspective, the US government seeks to achieve a threefold balance in the Framework: preventing the security risks associated with technology diffusion, maintaining the economic benefits of technology application, and consolidating its global technological leadership. However, there may be a discrepancy between policy design and actual results—excessively strict computing power restrictions may suppress global market demand, and cumbersome compliance requirements may create an innovation-inhibiting effect. More noteworthy is the systemic contradiction between this framework and the existing Export Administration Regulations (EAR) in terms of regulatory logic and operational aspects. This overlapping regulatory system may trigger rule conflicts in the international trade system.

By dividing technology access into levels, the Framework places China, Russia, and North Korea into the highest restriction level, Tier 3 (comprehensive embargo countries). Countries categorized in Tier 2 (conditional partners) must sign an agreement promising not to use Chinese AI technology, chips, or large language models in exchange for limited US technological support. This tiered system essentially forms a three-pronged technology blockade: Tier 3 countries are not only systematically isolated from key links in the global AI industrial chain but also face a multi-layered blockade of technology flow, from basic computing infrastructure (AI chips) and high-end computing resources to core algorithm resources (model weights). Cross-tier isolation is also implemented at the technological cooperation network level—that is, cutting off technical interaction channels with Tier 1 (core partners) and Tier 2 (conditional partners). This "digital iron curtain" strategy, with its Cold War mentality, aims to create asymmetrical technological advantages in key strategic areas by building multi-level technology isolation zones. Its intensity surpasses that of traditional trade embargoes, creating a comprehensive technological blockade covering hardware, algorithms, and cooperation.

The Framework not only establishes a technological blockade against China but also attempts to shape a highly exclusive global AI ecosystem. The implementation of the Framework will have a structural impact on the development of the Chinese and global AI industries, manifested in the following six dimensions:

Firstly, in terms of technological development, by including advanced computing integrated circuits and AI model weights in the scope of export controls, it creates technological barriers between most countries and cutting-edge computing resources. This will directly affect the ability of most countries to obtain advanced AI resources, thus slowing down the development of their AI technology.

Secondly, at the level of research and development innovation, hindered technological element flows may exacerbate theMatthew effectin the global AI research system. Countries lacking high-end computing hardware will face significant technological gaps in areas such as basic model training and complex algorithm development, potentially leading to structural fragmentation in the global AI innovation landscape.

Thirdly, in terms of supply chain restructuring, multinational corporations and economies constrained by technological dependence may face pressure to diversify their supply chains. Building alternative technology sources and fostering domestic technological ecosystems are becoming important strategic choices, which may propel the global AI and semiconductor industrial chains into a period of deep adjustment.

Fourthly, in terms of market landscape evolution, the "technology access levels" constructed by the US give its allies a first-mover advantage in the market, potentially strengthening existing technological monopolies. In terms of countermeasures, China and other restricted countries may accelerate the construction of independent and controllable technological systems, pushing domestic substitution into a strategic acceleration period.

Fifthly, at the level of enterprise operations, the Framework significantly raises the industry's compliance threshold. AI industry chain-related companies need to build multi-level compliance systems, involving the complete restructuring of technological R&D, supply chain management, and market expansion, significantly increasing compliance costs. In particular, multinational cloud service providers face new strategic constraints in terms of computing power deployment and service area selection, and the flexibility of their global business expansion is subject to institutional restrictions.

Sixthly, in terms of the international cooperation landscape, technology controls will reshape the global science and technology cooperation network. Traditional technological alliances are facing restructuring pressure, and new technological cooperation bodies may accelerate their formation, resulting in a multipolar trend in international science and technology competition. It is worth noting that the "boomerang effect" of technological blockades may become apparent, objectively stimulating more countries to increase investment in independent innovation, and the intensity of global scientific and technological competition will continue to escalate. In particular, for cooperation projects involving AI and advanced computing technologies, some countries may reduce cooperation with the US and its allies and instead seek other partners.

The introduction of this Framework may cause short-term pain in the global AI industry. However, in the long run, the history of technological development repeatedly confirms the law of "pressure breeds breakthrough." Historical experience shows that technology restrictions often act as a catalyst for self-reliant innovation, and this "blockade-breakthrough" dynamic game may lead to the formation of diversified technological development paths globally. Faced with the new industrial landscape, various countries will inevitably adopt combined strategies such as strengthening basic research investment, building new technology alliances, and optimizing the industrial ecosystem layout to meet the challenges.

Specifically for China, the solution lies in implementing an innovation-driven system project: Continuously increasing investment in key areas such as AI basic research and advanced computing integrated circuits, constructing a full-chain innovation system of "R&D-transformation-application"; accelerating the process of independent innovation of core technologies and overcoming "bottleneck" areas such as high-performance chips and AI development frameworks; and building a diversified development system for AI technology, while simultaneously promoting a global strategy for international market development. This multi-dimensional strategic layout will become the key path to breaking through the technological blockade and may even open up new value growth poles in the global AI industrial landscape.