By Alexandra Alper and Stephen Nellis
SAN FRANCISCO, March 5 (Reuters) – U.S. officials are debating a new regulatory framework for exporting artificial intelligence chips and are considering requiring foreign nations to invest in U.S. AI data centers or security guarantees as a condition for granting exports of 200,000 chips or more, according to a document seen by Reuters.
The rules are not yet final and could change. They would be the first attempt to regulate the flow of AI chips to U.S. allies and partners since President Donald Trump’s administration said it rescinded its predecessor’s so-called AI diffusion rules. Those rules sought to keep a significant amount of AI infrastructure buildout in the U.S. and route most purchases through a handful of U.S. cloud computing companies.
If adopted, the proposal would give the Trump administration ample leverage to negotiate investments in the U.S., one of Trump’s top priorities, as it decides how many AI chips to give to each country.
The rules depart markedly from former President Joe Biden’s approach, which was based on the premise that close U.S. allies should be exempted from most restrictions on exports of the coveted chips.
The proposed rules would not affect blacklisted countries such as Russia, which cannot obtain U.S. AI chips under rules set by Biden’s administration. China, which was among those countries, got a greenlight in December to receive Nvidia’s second-most advanced AI chips. Those shipments have been held up by national security requirements that could convince China not to go through with the purchases.
According to a document seen by Reuters, even small chip installations of less than 1,000 chips could need a license. The document said that to qualify for an exemption, the exporter of the chips such as Nvidia or Advanced Micro Devices would have to monitor them, and the recipient would have to agree to use software that would not allow the chips to be linked to other chips to form a “cluster,” the industry term used to describe large groups of chips.
Foreign firms that want up to 100,000 chips would need to provide government-to-government assurances, according to the document seen by Reuters. It said the Trump administration already required Saudi Arabia to provide such assurances in order to purchase advanced chips.
Installations of up to 200,000 chips could also require visits from U.S. export control officials, the document seen by Reuters said.
“The rule could help the U.S. government address chip diversion to China and ensure a more secure buildout of the most powerful AI supercomputers,” said Saif Khan, a former national security official in the Biden administration now at the Institute for Progress, a Washington think tank. “But the license requirements are overly broad, applying globally, raising concerns that the administration intends to use the controls as negotiation leverage with allies rather than for security.”
In a statement on social media service X, the U.S. Commerce Department confirmed it was debating new rules, but said they would not be similar to what it described as a “burdensome, overreaching, and disastrous” framework proposed by Biden’s administration.
Instead, the Commerce Department said it would follow the mold of deals to send U.S. chips to Saudi Arabia and the United Arab Emirates, where both countries agreed to invest in the U.S.
“The Commerce Department is committed to promoting secure exports of the American tech stack,” the department wrote. “We successfully advanced exports through our historic Middle East agreements, and there are ongoing internal government discussions about formalizing that approach.”
The draft seen by Reuters did not touch on exports of what are known as model weights, the key parameters of an AI system that companies such as OpenAI, Anthropic and others closely guard as one of their key competitive secrets. In the Biden rule, restrictions were placed on model weights in an effort to ensure the most advanced AI was developed and deployed in trusted and secure environments.
The White House did not immediately respond to a request for comment. Nvidia and AMD did not immediately respond to requests for comment.
(Reporting by Alexandra Alper in Washington and Stephen Nellis in San Francisco; editing by Chris Sanders and David Gregorio)
Comments