Nvidia and Oracle pushing for full repeal of AI diffusion rule
Companies must start complying with global restrictions May 15
Nvidia argues restrictions will push the world toward alternative technologies with limited benefits to US national security.
Senior foreign officials and major tech companies are pushing the Trump administration to rethink the country’s global semiconductor strategy, as the US prepares a controversial framework for controlling artificial intelligence development around the world.
The so-called AI diffusion rule, which restricts the number of AI processors that can be exported to most nations, prompted an outcry from tech giants like Nvidia Corp. after Biden officials unveiled it during their last week in office. US allies such as Israel and Poland also have chafed at the rules, which they worry will threaten their supply of the precious chips or make their countries less attractive for AI investment.
Now a slew of governments and many companies are trying to persuade President Donald Trump’s team to loosen some of the regulations before the deadline for compliance arrives in less than two months. Administration officials are nowhere near a consensus on how to proceed, and it’s still unclear which voices will carry the most weight in the debate.
Spokespeople for the White House and Commerce Department, which oversees chip export controls, didn’t respond to requests for comment on this story, which is based on interviews with more than a dozen people involved in or briefed on the negotiations. All of them requested anonymity to speak candidly and emphasized that the discussions are fluid.
One option not currently being considered at the staff level, two people said, is a wholesale repeal. Whether more senior officials change course, however, remains to be seen.
The goal in Washington is to ensure that AI development remains concentrated in the US and close partner nations. For data centers built elsewhere — from Malaysia to Brazil to India — American policymakers want AI infrastructure to align with US security standards. That includes things like implementing cybersecurity protocols and stripping Chinese hardware out of data center supply chains.
To that end, the AI diffusion rule divided the world into three categories of chip access: Nearly 20 close US partners — located in Europe and East Asia — get mostly unfettered access to AI chips. At the other extreme, adversaries like China and Russia are still effectively barred from importing the technology.
And, for a vast group of countries in the middle, the rule establishes caps on the total computing power available for export. Those restrictions affect nations with significant AI ambitions in the Middle East and Southeast Asia, many of which hadn’t been affected by previous rounds of semiconductor rules.
Companies seeking to build data centers in places like the United Arab Emirates or India can apply to bypass those national limits by agreeing to US guardrails. But there are still restrictions on how much capacity firms can put in any one country.
If Trump officials choose to maintain the framework as written, it could reshape digital infrastructure plans in regions key to the US-China competition, and pose particular challenges for companies like Oracle Corp. that are expanding aggressively in foreign markets. It also could turn access to US hardware into a geopolitical bargaining chip, in one of the most significant examples of American technological statecraft in the AI age.
Key Trump officials have indicated they want to streamline and strengthen the rule, Bloomberg has reported, but what that means in practice remains in flux.
Some officials, for example, have floated the idea of scrapping the three tiers and associated compute caps — but sticking with the baseline requirement of export licenses for most of the world, which builds on previous limits for China and some 40 other countries, people familiar with the matter said. Some are focused on narrower changes like requiring licenses for smaller shipment sizes, instead of the framework’s current standard of government notification, Bloomberg has reported.
Maintaining the curbs in any form would be an unwelcome outcome for Oracle and Nvidia, which have both pushed the Trump administration to fully roll back the rule and start the regulatory process from scratch, people familiar with the matter said. Nvidia, which makes the world’s most popular AI accelerators, argues the restrictions will push the world toward alternative technologies with limited benefits to US national security.
“We should continue to interact with the government,” Chief Executive Officer Jensen Huang told reporters last week, so that “whatever policy they believe is best, is informed.”
Huang was asked generally about export controls, not specifically about the AI diffusion rule. A representative for Nvidia declined to comment.
Oracle, meanwhile, is concerned about a provision that prohibits US companies from housing more than 7% of their computing capacity in any one second-tier country. That’s a major problem for the Texas-based hyperscaler, which is planning a $6.5 billion data center investment in Malaysia that will likely blow past the 7% limit, according to the research firm SemiAnalysis. Oracle didn’t respond to a request for comment.
The two companies recently joined the Information Technology Industry Council, a trade group representing most major tech firms, in part to push for an unambiguous stance against the AI diffusion rule, as opposed to an approach that would suggest amendments but not oppose the framework’s broad strokes. ITI’s official position, adopted earlier this month, promotes withdrawing the rule in full — with an asterisk that some members consider the existing architecture a useful starting point. An ITI spokeperson declined to comment.
Google, for example, complained that the rule disproportionately burdens US cloud service providers, but stopped short of calling for a repeal. The AI company Anthropic, meanwhile, says Trump officials should make the framework tougher by reducing the number of chips that can be exported to second-tier countries without a license — a threshold currently set at around 1,700 of Nvidia’s H100 accelerators. Allowing smaller shipments with only a government notification, rather than a formal approval, could “address legitimate commercial purposes” but also risks allowing bad actors to smuggle chips in small batches, Anthropic says. Some Trump officials agree.
Google and Anthropic didn’t respond to requests for comment.
Comments