Discuss the three "guiding opinions" on the global ban of Huawei's Ascend chips in the United States.

Source: Dong Bu Ya Qiao Research Institute

Today, BIS announced three "guidances"; after waiting for a long time, I still haven't seen the specific text, so I'll share my views based on the existing information.

Firstly, the BIS is quite careless; such an important notification actually contains typos: it mistakenly wrote "推理" (inference) as "interference". ( seems to set a very unscientific tone for the relevant rules? )

The reason for issuing this notice first is mainly because the "AI Diffusion Rules" officially came into effect on May 15. American companies are asking: Do we still have to comply with the set of regulations established by the Biden administration? Can controlled chips still be sold to Tier 2 countries?

The BIS clearly stated in this notification: Biden's AI diffusion rules are to be abolished, and BIS Deputy Secretary Kessler has already instructed that this rule should no longer be enforced.

Of course, the most important are the three "guiding opinions":

In addition, the U.S. Department of Commerce's Bureau of Industry and Security announced today actions to strengthen export controls on artificial intelligence chips overseas, including:

• Issued guidance stating that using Huawei's Ascend chips anywhere in the world violates U.S. export control regulations.
• Issued guidance, warning the public of the potential consequences of allowing American AI chips to be used for training and inference of Chinese AI models.
• Issue guidance to American companies on how to protect their supply chains from the impact of transshipment strategies.

The first "guiding opinion" is indeed quite domineering, essentially forcing everyone worldwide to choose sides between Huawei's H cards and Nvidia's N cards. The BIS (Bureau of Industry and Security, U.S. Department of Commerce) directly stated: if you use Huawei's Ascend chips, no matter where you are, that is a violation of U.S. export controls.

For third countries outside of China, this statement seems more like a heads-up or a warning—after all, the H card is currently in short supply domestically and there is simply no surplus for export. So this is more about setting rules and making things clear from the outset.

But for China, this is simply outrageous. You say that Chinese companies using their own domestically produced chips also counts as "violating U.S. laws"? Then can they still be used? If what you're saying is true, does that mean the U.S. intends to let Nvidia sell graphics cards in China again? Otherwise, what do you expect everyone to use?

Legally speaking, the "guidance" issued by the BIS actually falls into two categories within the U.S. administrative law system:

One type is called "interpretive rules", which explain how existing laws or regulations should be understood; the other type is "general statements of policy", meaning that law enforcement agencies inform everyone "how I intend to use my enforcement discretion".

The content released by BIS this time is actually its understanding of the Export Administration Regulations (EAR) and compliance recommendations, aimed at informing businesses: "You better do as I say, or I might trouble you."

This guidance itself is not law, nor does it need to go through the "public notice and comment" procedure required by the Administrative Procedure Act (APA). It is more like a signal or reminder issued by enforcement agencies. If it truly wants to have legal binding effect, it needs to be attached to specific EAR rules, or one day simply write this content into the formal EAR provisions, completing all regulatory procedures.

For example, the BIS issued a guidance on "deemed reexport" as early as 2013, indicating that providing U.S. technology to a third country person while abroad may be considered as "reexport." However, this guidance was not incorporated into the EAR until 2016, and prior to that, it was merely of a "for your reference, I suggest" nature.

Although the full text of the "guidance" released by the BIS has not been published yet, from the wording ("Using Huawei Ascend chips anywhere in the world violates U.S. export controls"), it may be related to the "Foreign Direct Product Rule" (FDPR) targeting Huawei, as well as the General Prohibition Ten under EAR Section 736.2(b)(10).

EAR No. 736.2(b)(10019283746574839201条,如果有大白话解释,是:如果你知道某项exit或使用行为已经违反了美国出口控制,或者即将发生违规行为,你就不能参与其中,无论是卖货、转运、 帮忙出钱、"暂时保管"、还是其他方式提供协助,都算协助行为,你也会违反出口控制。

The U.S. government had long accused Huawei of using "front companies" to have TSMC manufacture Ascend chips (which used U.S. tools), believing this violated the FDPR. As a result, any "knowledge" that Huawei previously did not issue this "guidance" might allow you to argue that you were not "aware". However, now that the BIS has clearly stated that using Huawei Ascend chips violates U.S. export controls, if you continue to use them, it is obviously considered "knowledge". Thus, the "guidance" actually provides a more substantial basis for the BIS to further impose penalties on companies using Huawei Ascend chips for violating U.S. export controls.

Looking back at the recent statements from key figures in U.S. AI policy, it seems that the current Trump administration's strategic containment of China's AI is gradually taking shape with three policy pillars: first, relaxing regulations domestically and increasing investment in AI infrastructure to support local companies in staying "far ahead" in the AI race; second, promoting the widespread adoption of the "U.S. tech stack" globally, ensuring that countries use American chips, models, and tools rather than those from China; third, implementing more precise and high-pressure export controls against China to block its access to American "choke-point" technologies.

The first guideline threatening countries to abandon the use of Huawei chips clearly reflects the aforementioned second pillar. As David Sacks mentioned in his comments on AI diffusion rules:

"The U.S. still has the opportunity to widely deploy the 'American tech stack' globally, while we are still ahead, we need to seize this opportunity quickly."

Michael Kratsios, Director of the White House Office of Science and Technology Policy, also emphasized:

"This government must ensure that America's partners and allies use America's excellent technology."

This strategy relies not only on "pushing American technology" but also on "blocking Chinese technology." Looking globally, the only country that can pose a systemic challenge in AI infrastructure is China. Therefore, the U.S. is intensifying the promotion of its own chips and cloud computing services in third countries while also attempting to keep China's AI solutions out of the market through export controls and investment restrictions.

This is also one of the few areas where the U.S. government and businesses have reached a consensus. Take NVIDIA as an example; Jensen Huang has publicly stated, "Huawei is our biggest competitor." Since the enemy is clear, it is natural to concentrate all policy ammunition on this opponent. You see that in this "guidance document," only the words "Huawei Ascend" are specifically mentioned, without even mentioning Cambricon, which indicates the clear direction and policy priority of this document.

In fact, some American companies have also made similar suggestions to the government: the gist is that Huawei's 5G arrived late in the U.S., and in the end, they still have to go through a lot of trouble persuading allies to remove Huawei equipment, so as not to lose again in AI.

For example, AMD's CEO Dr. Lisa Su recently testified before the Senate Committee on Commerce, Science, and Transportation.

The United States should also promote the widespread application of American artificial intelligence technology globally, while ensuring national security. Specifically, the U.S. needs to establish a clear policy framework through government and industry collaboration to avoid international partners turning to alternative technologies due to past restrictions on the dissemination of artificial intelligence technology.

Microsoft President Brad Smith, who also attended the hearing, emphasized:

The competition between the U.S. and China is not only about technological innovation but also about who can more quickly promote technology to other countries. The outcome of this part of the competition depends on first-mover advantage. The United States needs to devise a smart export control strategy that ensures national security while guaranteeing other countries reliable and continuous supply of artificial intelligence components and services. Moreover, we must continuously maintain the international community's trust in American products, companies, and the country itself.

In the comments submitted to the White House AI Action Plan, the American company Eclipse pointed out:

"至关重要的是,美国的领先公司应与美国政府策略保持一致,抓住主动权,为盟国提供可行的人工智能基础设施解决方案。 在 5G 竞争中美国难以提供比华为产品更具效能的替代方案,而美国公司在人工智能基础设施领域是世界领先者,能够向盟国提供可重复、可扩展的能力体系,以构建由美国开发的AI智能基础设施的全球网络。 与美国在 5G 竞争中不得不艰难说服盟国拆华为设备不同,美国目前有机会率先为盟国提供可行的人工智能解决方案。 此外,利害关系更大 —— 与5G 电信设备或对石油等自然资源的依赖不同,一旦一个国家采用了基于敌对架构构建的人工智能基础设施,几乎不可能liberation那些已经依赖这些系统的思想和社会系统。 ”

If foreign companies are not allowed to buy Huawei, they can only turn to buy NVIDIA, which may be the core strategic goal of the Trump administration's "guidance" this time.

At the same time this opinion was announced, Trump made a high-profile visit to Saudi Arabia and the UAE, announcing that the two countries would make large-scale purchases of Nvidia GPUs, which is clearly a carefully arranged dual signal of politics and industry. The U.S. government has now issued guidance on a "global ban on the use of Huawei Ascend chips," likely to convey a clear message to other countries: as long as your companies are willing to exclude Chinese technology and adopt U.S. standards like G42, they can obtain the most advanced GPU supplies.

The direct effects of this strategy are threefold: first, it weakens the motivation of other countries to purchase Huawei products, creating a "side-choosing" pressure; second, it helps Nvidia explore new markets outside of China, increasing its global dependency diversity; third, it assists Nvidia in gradually reducing its revenue dependence on the Chinese market, establishing a more controllable strategic customer base.

这一战略安排,实际上延续了美国之前对半导体制造设备的管制方式——和聯合达成"共识"(如美日荷三方协议),在全球范围内构建对中国的技术封锁链。 但这很可能会重现FDPR的固有缺陷:要让一个主权国家,心甘情愿地长期执行美国法律,尤其是针对中国的部分,非常难。 从实际效果看,要达成让这些国家不再使用华为芯片的目标,美国很可能还是需要一国一策、挨个谈判。 这一过程的复杂性、diplomacy成本、所涉及国家之广,将远远超过当年三方密约谈判。

People can't help but think of the "Mutual Defense Aid Control Act of 1951" in the early days of the Cold War, which is also known as the "Battle Act" (named after its proposer, the United States and Laurence Battle). 该法案的核心条款规定:凡接受美国協助的国家,若不对苏联及其卫星国实施战略物资禁运,美国将中止对其協助。 当时所谓的"战略物资",不仅包括军火、石油,也涵盖了可能增强苏联军事能力的关键技术产品。 这一legislation,本质上是美国试图以協助为杠rod,迫使盟国在经济上与共产主义阵营彻底脱钩,通过控制贸易手段来服务于其地缘政治和国家安全目标。 这与today 美国以英伟达GPU、AI技术栈作为新"战略物资",迫使他国"选边站"的行为如出一辙。

However, the actual implementation of the "Bartel Act" was not smooth. European allies, especially the UK and France, were not willing to completely follow Washington's orders regarding trade with the Soviet Union, as they still relied on economic interactions with Eastern Europe to maintain domestic recovery in the early stages of the Cold War. In the face of allies' resistance, the United States once invoked the Act to exert pressure, but ultimately had to make concessions in its implementation, issuing a large number of exemption licenses, which significantly undermined the effectiveness of the policy.

For example, after the Korean War, U.S. Secretary of State Dulles advocated for the continuation of a comprehensive embargo against China and North Korea, urging allies to follow suit. However, the United Kingdom and France insisted that trade policies toward China should be consistent with those toward Eastern Europe and should not involve "selective decoupling." Under such pressure, although the U.S. relied on the "Battle Act" for leverage for a time, it ultimately compromised in 1957, allowing allies to resume some trade with China and North Korea. Meanwhile, the U.S. maintained its unilateral embargo policy against China until the Nixon administration took office in 1969, at which point the ban was gradually relaxed.

Compared to enterprises from third countries, I actually find it most awkward for those domestic Chinese companies that still rely on American technology but have to deploy Huawei Ascend chips. For example, some large AI companies, cloud computing platforms, and operators may not have fully achieved technological "decoupling" from the U.S., and still involve EAR-controlled items in their procurement or supply chain. In this case, if they continue to use Huawei Ascend chips, they may be deemed by the U.S. as violating the FDPR, thereby facing the risk of being added to the Entity List. Once on the list, even if they completely get rid of American technology in the future, they will encounter many obstacles in business operations, financing, and global cooperation. The difficulties involved may only be truly understood by experts in compliance/strategy within large enterprises.

The second "guiding opinion" is actually about how to prevent American AI chips from being used to help China train large models.

This is not a new issue raised by the Trump administration; similar restrictive measures began to take shape during Biden's term. At that time, two main paths were considered:

Firstly, through legislation, such as the "Remote Access Security Act," the Department of Commerce is granted the authority to control the export of IaaS cloud services, which means that U.S. cloud service providers are not allowed to sell computing power to Chinese users for training cutting-edge models, but this path ultimately did not materialize into law.

Another path is the "AI Diffusion Rules" announced later, which directly sets limits on the training of large models themselves. As long as a closed-source model is trained using controlled hardware such as GPUs and servers, and the computational operations during training exceed 10 to the 26th power, even if this model is trained outside the United States, its "model weights" still belong to controlled items, and exports must comply with U.S. export controls. It also introduces a "red flag warning"—if a U.S. IaaS cloud service provider trains a large model for a foreign company's subsidiary established in the U.S. on its home soil, and this model is subsequently handed over to the foreign parent company of the customer, that action could constitute an export, and the U.S. company may be at risk of assisting in violations. Therefore, BIS recommends that cloud service providers clarify how customers intend to use the model weights before delivery, and whether a license is needed; if so, it must be obtained first.

However, the AI diffusion rules put in place by Biden have now been explicitly abolished by the Trump administration, and the entire regulatory logic surrounding model weights has suddenly lost its specific legal basis. So we are back to square one: how will the U.S. government prevent American AI chips from being used to train Chinese models? This remains a hotly debated topic within U.S. domestic policy circles.

From the wording of this second "guidance" document, it appears more like a "warning" aimed at the public rather than a hard regulation—informing you that using American chips to help China train models is a sensitive action that could cause trouble, and it’s best not to do so. However, it does not clarify whether doing so is illegal, whether there will be penalties, how much the penalty would be, or who would impose the penalties. From this perspective, it is unlikely to impose a clear legal obligation on American cloud service providers. After all, according to the principles of administrative law, such "guidance" needs to be attached to specific EAR official rules in order to have mandatory effect, and there are currently none.

In comparison, the third "guiding opinion" appears the least conspicuous, as it does not establish any new restrictions but rather reminds American companies to pay attention to supply chain risks and prevent chips from being transshipped. It is more of a "compliance reminder" rather than a legal obligation, but it could actually be used as a risk clue for enforcement or as a basis for post-event review.

The above is just a preliminary analysis based on the currently limited information; the specifics will still depend on the full text of the "guiding opinions," which is expected to be released soon.

Attachment: Full Text of the Ministry of Commerce Notification

The Ministry of Commerce revokes the artificial intelligence diffusion rules from the Biden era and strengthens export controls related to chips

Washington D.C. — Today, the U.S. Department of Commerce initiated the process to revoke the Biden administration's artificial intelligence dissemination rules, while also announcing additional measures to strengthen semiconductor export controls globally.

The artificial intelligence diffusion rules were released on January 15, 2025, with compliance requirements originally set to take effect on May 15, 2025. These new requirements will stifle innovation in the United States and impose heavy new regulatory burdens on businesses. The artificial intelligence diffusion rules will also downgrade dozens of countries to second-class status, thereby undermining diplomatic relations between the United States and these countries.

The U.S. Department of Commerce's Bureau of Industry and Security (BIS) plans to publish a notice in the Federal Register to officially withdraw the rule and will issue an alternative rule in the future.

The Deputy Secretary of Commerce for Industry and Security, Jeffrey Kessler, has instructed enforcement officials at the Bureau of Industry and Security not to enforce the Biden administration's artificial intelligence diffusion rules, stating:

"The Trump administration will implement a bold and inclusive strategy to collaborate with globally trusted foreign partners in artificial intelligence technology, while preventing the technology from falling into the hands of our adversaries. At the same time, we do not endorse the Biden administration's approach of imposing its poorly considered and counterproductive artificial intelligence policies on the American public."

In addition, the U.S. Department of Commerce's Bureau of Industry and Security announced today actions to strengthen export controls on overseas artificial intelligence chips, including:

1. Issued guidance stating that the use of Huawei's Ascend chips anywhere in the world violates U.S. export control regulations.
2. Issue guidance and warn the public about the potential consequences of allowing American AI chips to be used for training and inference of Chinese AI models.
3. Issue guidance to American companies on how to protect their supply chains from the impacts of transshipment strategies.

Today's actions ensure that the United States maintains its leading position in artificial intelligence innovation and sustains its dominant status in the global artificial intelligence arena.