
The Zhitong Finance App learned that US President Trump said at the Asia-Pacific Economic Cooperation (APEC) Informal Leaders' Meeting held in Shenzhen on Thursday that the US opposes global regulations that may hinder the development of artificial intelligence, while acknowledging that AI technology has certain risks and requires careful management.
Trump said at the conference that artificial intelligence “is almost all good, but it also has some bad aspects,” stressing the need to properly address related risks, but opposes any attempt to control or stifle its development.
It is worth noting that when Trump delivered the above speech, the US government had just further tightened export controls on advanced AI chips and overseas computing power services. The US Department of Commerce recently introduced new regulations to extend some export license requirements to remote computing power access to overseas data centers in an attempt to restrict companies in restricted destinations such as China from obtaining advanced AI computing power through data centers in third countries.
This means that on the one hand, the Trump administration wants American companies to speed up the development and commercialization of AI technology, and on the other hand, it is also strengthening control over end users of advanced AI computing power. The two policies target different fields, but together they reflect the US government's policy trade-off between promoting the development of the AI industry and maintaining technological advantages.
Meanwhile, reports surrounding Nvidia's former CEO Huang Renxun and his relationship with a Chinese businessman also brought further attention to discussions on advanced chip export controls in the US. However, the relevant espionage allegations have yet to be independently verified.
Trump opposes regulating “stifling AI,” why is the US tightening overseas computing power licenses at the same time?
Trump said at the APEC summit in Shenzhen on November 19 that the risks posed by artificial intelligence require careful management, while warning that excessive regulation may hinder technological development. This statement almost coincides with the US government's recent measures to tighten exports of advanced AI chips and overseas computing power licenses.
The White House hopes that US companies can quickly develop and deploy cutting-edge models, while the Commerce Department wants to restrict certain overseas users from obtaining sufficient US computing power to train and run these models. The former is concerned about whether regulation will slow technological innovation, while the latter is concerned about who the chips designed by the US will ultimately provide computing power to.
The differences over AI regulation in the US mainly focus on who should assess the risk of cutting-edge models and when the assessment should be involved.
The White House previously promoted the establishment of a relevant evaluation mechanism, requiring the federal government to evaluate cutting-edge models that meet specific network capability thresholds up to 30 days in advance before opening them up to other trusted partners.
Meanwhile, an expert proposal announced by California on November 18 suggests that the state government establish an AI security bottom line and require enterprises to undergo independent testing during the model training, development, and deployment stages. However, these recommendations are still in the policy discussion stage and cannot yet be considered an implemented regulatory requirement.
For model developers, the difference between the two ways of thinking is whether the safety assessment mainly focuses on before the model is released, or whether it needs to go through multiple aspects of model training, development, and deployment.
The overseas computing power licensing rules recently promoted by the US government target another technical reality: even if advanced AI chips remain in overseas data centers, customers can still remotely call the entire GPU training model through the Internet, or continue to call the GPU to process user requests after the model is launched.
As a result, US controls on advanced AI technology are being further extended from physical exports of chips to end users of computing power services.
From where to ship chips to whom to rent computing power
The new regulations announced by the US Department of Commerce on November 16 and scheduled to take effect on November 19 set two licensing requirements.
The first focus is on the destination of the equipment: in principle, the export, re-export, or local transfer of designated advanced chips and related equipment to Singapore, Malaysia, Thailand, Indonesia, and Brazil must obtain permission from the US government.
The second one focuses on end users of computing power, and its scope of application is not limited to the five countries mentioned above. If the US side knows that the relevant equipment will be used in data centers outside the US and will provide remote computing power that exceeds the specified threshold to enterprises with headquarters or final parent companies located in restricted destinations such as China, permission must also be obtained for related equipment transactions.
In addition, the US government has set up a 90-day transition arrangement for some eligible transactions, and some eligible commercial recommendation systems can also be exempted.
This means that even if GPUs remain in overseas computer rooms such as Singapore and Malaysia, as long as the relevant data centers provide restricted end users with computing power services that meet the specified threshold, related equipment transactions may still be subject to the US export licensing system.
The impact of this rule on large cloud computing operators and data center operators in Asia is particularly noteworthy. For data centers that have already been built and electrified, if the original customer's use of computing power is restricted, operators may need to find new compliant tenants, and it will take time to re-sign and deploy the business, and computer room utilization rate and rental revenue may be affected during this period.
Nvidia and data center operators face different impacts
This also explains why the profit impact of the same new regulation on Nvidia and overseas data center operators may be significantly different.
For chip suppliers, if some customers are unable to continue to obtain advanced AI chips or computing power services, there is still an opportunity to resell related devices to other buyers that meet US regulations. As long as global demand for AI computing power remains strong, suppliers will not necessarily have to bear the full loss of demand.
Bank of America believes that Nvidia's global demand still exceeds supply, and some orders are expected to be accepted by other compliant buyers. Goldman Sachs also pointed out that after considering the new leasing restrictions, Nvidia raised its growth expectations, indicating that the market still has some confidence in its overall demand prospects.
Nvidia's median revenue guidance for the fourth fiscal quarter was $130 billion, of which data center calculation revenue from China has been excluded. For investors, the key is whether export restrictions will substantially reduce the global demand that Nvidia can serve, or whether they will mainly change the final flow of chips and customer structure.
In contrast, the risks faced by overseas data center operators may be more immediate.
Data center construction involves many upfront investments such as land, power supply, computer rooms, servers, and financing. Even if the original customer is unable to continue using the relevant computing power, the operator will still have to bear depreciation, space and financing costs. If new compliant customers are not found in a timely manner, the decline in GPU billable time may directly affect rental revenue and project returns.
Based on this, Morgan Stanley is concerned about the risk on the other side: model developers that rely on overseas clusters may delay training schedules, while companies that rent their own GPUs and provide overseas inference services may also be more directly affected by the new regulations.
On November 17, the stock prices of Malaysian data center infrastructure companies MN Holdings and Critical Holdings both fell nearly 20%, and Nvidia's local partner Yang Zhongli Electric Power also fell nearly 10%. Relevant market performance reflects investors beginning to assess the potential impact of the tightening of overseas computing power licenses on local data center project construction and future rental demand.
Overall, the tightening of overseas computing power licenses by the US does not mean that global demand for AI computing power itself has declined, but has further raised the compliance threshold for the cross-border flow of advanced AI computing power. For chip vendors such as Nvidia, whether global demand can be handled by other compliant customers will determine the actual impact of export restrictions on revenue; for Southeast Asian data center operators, being able to find replacement tenants in a timely manner and maintain computer room utilization may be a key factor affecting project returns.