According to a document seen by Reuters, U.S. officials are discussing a new regulatory framework for the export of artificial intelligence chips and are considering requiring foreign investment in U.S. AI data centers or the provision of security guarantees as a condition for allowing large-scale chip exports.

These rules are not yet finalized and are subject to change. This would be the first attempt to regulate the flow of AI chips to U.S. allies and partners since President Donald Trump's administration announced the repeal of what his predecessor's so-called "AI proliferation rules." These rules aimed to keep a significant amount of AI infrastructure construction in the United States and direct most procurement to at least a few U.S. cloud computing companies.

Involves small chip installation projects

A document seen by Reuters shows that even small chip installation projects with fewer than 1,000 chips may require a license. The document states that to obtain an exemption, chip exporters (such as Nvidia or AMD) must monitor their projects, and recipients must agree to the use of software to prevent the chips from connecting with other chips to form larger "clusters" (a term used in the AI industry to describe large numbers of chips).

According to sources familiar with the matter, the proposed regulations would require companies to obtain a U.S. license when exporting virtually all AI accelerators from companies like Nvidia and AMD. This marks a global expansion of the restrictions, which currently cover approximately 40 countries. These chips are among the most sought-after components in the tech industry. Companies like OpenAI and Alphabet purchase these chips in bulk for installation in data centers running services like ChatGPT and Gemini.

The impact of this news

Following this news, shares of both Nvidia and AMD fell to their intraday lows on Thursday. Nvidia's stock price dropped as much as 1.9%, while AMD's fell 2.3%.

President Donald Trump's team has repeatedly stated their desire for the world to use American artificial intelligence technology, and these draft rules are not intended to ban Nvidia's exports. Instead, the regulations aim to establish the U.S. government as the gatekeeper for the AI industry: companies—and in some cases, even their governments—must obtain approval from the U.S. Department of Commerce to purchase these valuable accelerators. How the Trump team issues these licenses will determine whether countries can build critical digital infrastructure, a technology that many world leaders consider key to economic growth, corporate competitiveness, and military sovereignty.

Sources familiar with the matter said the specific approval process will depend on the computing power a company requires. These sources requested anonymity due to the ongoing policy debate. Orders of no more than 1,000 of Nvidia's latest GB300 graphics processing units (GPUs) will have a relatively simple approval process and may enjoy certain exemptions. Companies building larger clusters will need pre-approval before applying for an export license. Depending on the specific data center, they may need to meet conditions such as disclosing their business model or allowing the U.S. government to conduct on-site inspections.

Large-scale deployment

For truly large-scale deployments—such as a company owning more than 200,000 Nvidia GB300 GPUs in a single country—host governments must be involved. Sources familiar with the matter said the U.S. will only approve exports of such products to allies, provided those allies make stringent security commitments and "matching" investments in U.S. artificial intelligence. They also noted that the draft rule does not specify investment ratios. For reference, the 200,000 GB300 GPUs are what the British company NScale plans to provide to Microsoft, a company that specializes in leasing AI chips to third parties, and plans to deploy them in four locations in the U.S. and Europe. NScale called the deal "one of the largest AI infrastructure contracts ever signed."

The Bureau of Industry and Security of the U.S. Department of Commerce, which oversees semiconductor export controls, Nvidia, and AMD did not respond to requests for comment.

The sources emphasized that the Trump team's framework is not yet finalized, and officials from various federal agencies are currently providing feedback. The draft may undergo significant revisions or be shelved due to other priorities.

Nevertheless, this marks the most substantial step in pursuing a global chip export strategy since the Trump administration abandoned President Biden’s plan in May. Trump administration officials have consistently criticized the previous administration's so-called "AI proliferation rule," which restricted the sale of AI chips to most countries and set export caps, deeming it overly cumbersome. They have also repeatedly emphasized encouraging the world to use American, rather than Chinese, AI technology and have begun developing initiatives to promote the export of American AI products, particularly to the Global South.

Whether Trump's approach is ultimately more restrictive or restrictive will depend on how officials utilize the global licensing requirements they are currently considering.

If Washington can quickly approve chip sales with few strings attached, global AI infrastructure development can continue smoothly—it just requires more paperwork. Meanwhile, bureaucratic delays or protracted negotiations will hinder project planning. Last year, licenses were only issued months after the U.S. announced a chip export agreement with the UAE, but only on the condition that the Gulf state invest $1 in the U.S. for every $1 it invests domestically.

A significant unknown is how much funding the U.S. will expect from countries like France or India, which also have ambitious plans to build large data centers of 1 gigawatt or more. Another factor is how Trump will leverage chip restrictions in broader diplomatic negotiations, especially as he adjusts his tariff strategy. Last year, Trump threatened to impose controls on semiconductor exports in retaliation for digital services taxes imposed by the EU and elsewhere.

Equally uncertain is how the U.S. will handle restrictions on model weights. Model weights are numerical parameters used by AI software to process data and make decisions, making them one of the world's most valuable intellectual property. Biden's global chip framework imposed general restrictions on where companies can host cutting-edge model weights, while Trump's proposal addresses this through individual licenses.

World leaders are generally reluctant to put their nation's technological future at Washington's mercy. But when it comes to computing power, they have little choice. Countries must either import chips from Nvidia, the undisputed U.S. leader, or from Chinese companies like Huawei Technologies Co., Ltd. Huawei produces weaker, less efficient chips in much smaller quantities, but has ambitions for global expansion. To prevent them from considering the latter, Washington has warned that using Huawei's AI accelerators anywhere in the world could violate U.S. trade restrictions.

China's AI ambitions are a core factor in Washington's AI policy. One key focus is limiting China's production of AI chips, a goal the US has achieved by restricting exports of semiconductor manufacturing equipment—and, under Trump, allowing Nvidia to return to the Chinese market and compete with Huawei. Sources familiar with the matter say the latest global rules won't change Washington's strategy on chip exports to China, as Trump's team is weighing how many Nvidia chips are needed to both curb Huawei's rise and prevent China from gaining excessive additional computing power.

However, this global framework could have significant ripple effects on China's AI industry. One direct consequence is that the US government will have better control over global chip flows, which some officials have long considered crucial for combating semiconductor smuggling. Trump's team planned last year to impose chip export controls on Malaysia and Thailand to curb chip inflows to China, but ultimately decided against targeting any single country in their global strategy.

The US can also use export licenses to restrict Chinese companies' access to AI chips overseas, as long advocated by some national security hawks. In some cases, they have already done so: according to sources, Trump's team linked some of Nvidia's chip deliveries to the UAE to a ban on those companies providing computing services to any Chinese AI companies. It remains unclear whether they will implement similar licensing terms in other regions, such as Southeast Asia.

The U.S. Department of Commerce issued a statement on the social media service platform X, confirming that new rules are being discussed, but stating that these rules will not resemble the "cumbersome, excessive, and disastrous" framework proposed by the previous administration.

The Department of Commerce stated that it will follow the model previously used for exporting U.S. chips to Saudi Arabia and the United Arab Emirates, both of which have agreed to invest in the United States.

"The Department of Commerce is committed to promoting the secure export of U.S. technology," the department wrote. "We have successfully advanced exports through landmark Middle East agreements, and there are currently discussions within the administration to formally establish this approach."

The White House, Nvidia, and AMD did not immediately respond to requests for comment.

Source: Semiconductor Industry Observer (compiled)

View more at EASELINK