According to the report, U.S. officials are discussing a new regulatory framework for exporting AI chips and are thinking of requiring foreign countries to invest in U.S. AI data centers or provide security guarantees before allowing exports of 200,000 chips or more.
The regulations are subject to change and are not yet final. Since President Donald Trump’s administration claimed to have revoked its predecessor’s so-called AI diffusion restrictions, this would be the first attempt to control the flow of AI chips to U.S. allies and partners. These regulations aimed to maintain a large portion of AI infrastructure development in the United States and direct the majority of purchases via a small number of American cloud computing firms.
If accepted, the idea would provide the Trump administration with significant leverage to negotiate investments in the United States, one of Trump’s key priorities, as it determines how many AI chips to distribute to each country.
The guidelines differ significantly from former President Joe Biden’s approach, which was based on the assumption that close US friends should be immune from most limitations on exports of the coveted chips.
Blacklisted nations like Russia, who are prohibited from obtaining U.S. AI chips under Biden’s administration’s regulations, would not be impacted by the proposed regulations. Among those nations was China, which received approval in December to purchase Nvidia’s second-most sophisticated AI chips. National security concerns have delayed such shipments, which may persuade China not to proceed with the purchases.
Even limited chip installations with fewer than 1,000 chips would require a license. According to the document, in order to be eligible for an exemption, the exporter of the chips—such as Nvidia or Advanced Micro Devices—must keep an eye on them, and the recipient must consent to use software that prevents the chips from being connected to one another to form a “cluster,” which is the term used in the industry to describe large groups of chips.
According to the document , foreign companies seeking up to 100,000 chips would have to give government-to-government guarantees. It claimed that in order for Saudi Arabia to buy cutting-edge processors, the Trump administration already mandated that the country give such assurances.
Installations of up to 200,000 chips may also necessitate visits from US export control officials. The law might assist the US government address chip diversion to China and assure a more secure buildout of the most powerful AI supercomputers, according to Saif Khan, a former national security officer in the Biden administration who now works at the Institute for Progress, a Washington think tank. However, the license requirements are extremely broad and apply globally, creating worries that the administration plans to use the controls as bargaining chips with allies rather than for security.
In a statement on social networking service X, the US Commerce Department revealed it was discussing new restrictions, but claimed they would not be identical to the framework suggested by Biden’s administration, which it described as “burdensome, overreaching, and disastrous”.
The Commerce Department plans to transfer U.S. chips to Saudi Arabia and the United Arab Emirates, both of which have invested in the US. “The Commerce Department is committed to promoting secure exports of the American tech stack,” the department noted. We effectively advanced exports through our old Middle Eastern relationships, and there are current internal government debates about formalizing that approach.”
Exports of what are known as model weights—the crucial characteristics of an AI system that firms like OpenAI, Anthropic, and others fiercely protect as one of their primary competitive secrets—were not included in the draft. In an attempt to guarantee that the most cutting-edge AI was created and implemented in safe and reliable settings, the Biden rule imposed limitations on model weights.
A request for comment was not immediately answered by the White House. Requests for comment from Nvidia and AMD were not immediately answered.
