By: Jeffrey D. Bean
Last month, the U.S. Department of Commerce under the Donald Trump administration formally rescinded his predecessor’s Framework for AI Diffusion, which was set to go into effect on May 15. The rule divided the world into three tiers for access to advanced graphics processing units (GPUs) for artificial intelligence (AI) datacenters. Firms headquartered in Tier 1 countries – most close allies and partners of the United States – would have had unlimited access to advanced H100 or B100/B200 accelerators from Nvidia. Tier 2 countries – which included Singapore, Israel, Portugal, Switzerland, Poland, the United Arab Emirates, Saudia Arabia, and India – would have had only a limited ability to buy advanced GPUs for datacenter buildouts. Tier 3 countries included adversaries like China, Russia, North Korea and others that were banned from acquiring advanced GPUs already through export controls and sanctions.
The AI diffusion rule was widely criticized by the semiconductor industry, experts, and partner countries for lack of consultation and it reflected insufficient market understanding. The rule was rightly characterized by experts as challenging to enforce, overly broad, too rigid, and potentially pushing partners toward Chinese AI options while limiting U.S. chip firm sales. With the diffusion framework now out of the picture, the public now awaits the Trump administration’s crack at a policy that enables continued U.S. sales and leadership while preventing adversary access.
The rescinding of the AI Diffusion Rule may appear a respite for Tier 2 countries, including in the Middle East and India. Trump’s recent tour to the Middle East resulted in several large deals for AI chips for Saudi Arabia and the UAE, similar to the large Microsoft-G42 deal for AI datacenters in the UAE that the Biden administration brokered in 2023-2024. The prevalence of high energy sources, land, and massive investment capital in these two countries facilitated the agreements (and presumably overshadowed any human rights concerns). The agreements include sales of Nvidia Blackwell chips to Saudi startup Humain as well as a $10 billion collaboration between Humain and AMD. In addition, the UAE made a preliminary commitment to buy 500,000 NVIDIA GPUs annually through 2027 which will go to local firm G42, including a 5 gigawatt installation, as well as datacenters operated by Microsoft and Oracle. The details of commitments from Saudi and UAE to limit to Chinese access via export controls are unknown, but it is reasonable to assume that some are present.
These deals suggest a return to a landscape whereby the United States will consider sales of advanced GPUs on a bilateral basis, which may be initially greeted as good news by key partners like India. As part of the U.S.-India joint leaders’ statement in February, the two sides committed to a roadmap for accelerating the building out of AI infrastructure and datacenters which was totally incongruent with the diffusion framework.
Yet key issues remain. The bandwidth needed to not just negotiate several bilateral agreements but enforce their export control terms remains incredibly high for the U.S. Department of Commerce (and partner countries) and still imposes significant compliance costs on U.S. companies. Moreover, one of the unresolved challenges is the issue of Chinese actors paying for data center access to run training and inference via the cloud, because Chinese firms do not need to have a datacenter in their borders if they can pay for cloud access remotely – which is another challenge that the Biden team had attempted to tackle with their framework. There remain legitimate national security concerns with respect to both sales and cloud access, and reportedly there is significant disagreement in the Trump administration about how to manage this.
In addition, the Trump administration did underscore a clear articulation that utilizing competitor advanced AI chips designed by any Chinese actor, including a chip like the Huawei Ascend 910, is a violation of U.S. export control law. How this will be enforced effectively also remains an open question. But U.S. and global companies have been warned, and this may have an intended chilling effect of limiting collaboration with Huawei’s CANN AI stack. Know your customer requirements on AI chips and cloud services will likely be increasingly important for U.S. semiconductor and technology firms. All in all, the United States’ oscillating policy on AI diffusion reflects an ongoing struggle in how best to simultaneously retain U.S. leadership in semiconductors and advance compute for AI at both a market level and in national defense applications, while blocking adversaries’ access to advanced AI chips and the capability to manufacture them.
Jeffrey D. Bean is the Program Manager for Technology Policy and Editor at ORF America.