U.S. AI Export Regulation Update #2
Final Update before May 15th Deadline & Our Predictions
April 24th, 2025 – As our clients, prospective clients and partners are well aware, the deadline for compliance with the Biden administration’s Interim Final Ruling (IFR) “Framework for Diffusion of AI” export regulation is only fifteen working days away. It’s expected that the Trump administration will publish their modifications to the IFR by this deadline. In the last Aiolian International update, we outlined four key factors the Trump administration will have to take into considerations with their modifications. In this update, we describe two possible scenarios which we call “Shift Left” and “Wolf in Bear’s clothing”. “Shift Left” retains the three tier system, leaves Tier-1 nations unchanged, limits Tier-2 nations to performance-capped H20 (or similar) GPUs and – as announced last week (footnote 6) - limits Tier-3 nations (China, etc) even lower performance GPUs. However, we believe the most likely is the “Wolf in Bear’s clothing” scenario where the Trump administration will close several loopsholes in the Biden IFR but otherwise maintain the same framework - importantly, it will maintain TPP (total processing performance concepts) and NVEU concepts, (footnote 7). Of course, when announced, the Trump administration will claim that Biden was weak and created a mess, and that Trump had to fix it with something that really stands up to China. But for the most part, it will be the existing IFR framework – i.e. the wolf is stronger and meaner but it’s the same wolf, and it will be dressed up to look like a bear. We are fully aware that our clients have been stockpiling GPUs (per our council) which buys some time, but the NVEU application process timeline is unpredictable and we urge clients to start planning now.
As described in our last update, some of our clients might have been heartened to know that in addition to Nvidia, both Microsoft (footnote 8) and Oracle (footnote 9) were lobbying the Trump administration and calling the IFR as a “strategic misstep in the global AI race” in that it created a strategic opening worldwide for China’s GPUs and AI influence. However, two events of the last two weeks show that the Trump administration wholly disagrees with this philosophy. On April 10th, at the Huawei Cloud Ecosystem conference, Huawei announced their Cloudmatrix 384 system (footnote 10). The system is designed to compete with Nvidia’s Blackwell-based NVL72 system and - leveraging system-level design to compensate for the lower per-chip performance of the Ascend 910C GPUs chip – delivers 1.7X floating point performance and 2.1X HBM (memory) bandwidth (footnote 11) versus the NVL72. While the Cloudmatrix 384 is sub-optimal on data center footprint, power and system cost (by 3-4X) and they still have to demonstrate that they can scale their manufacturing, it does demonstrate that Huawei is fully capable of delivering AI capabilities at scale… and more importantly, the cost/power disadvantage irrelevant to Chinese national security projects. Despite this announcement, in the days prior to April 16th (when Nvidia announced their financial hit), the Trump administration notified Nvidia that exports of H20 GPUs to China/etc will now be restricted. This notification was likely in the form of a “is informed” letter (footnote 12) as used by the U.S. Bureau of Industry and Security (BIS). The BIS has yet to make a formal announcement of their own, but we expect this shortly where the details of the new (lower) performance thresholds will be disclosed. No coincidently, Nvidia’s CEO, Jensen Huang visited Beijing on April 17th (footnote 13) to reassure Beijing and Chinese customers that Nvidia is already working on new chips compliant with the new thresholds.
Shift Left scenario - The level to which Nvidia was financially impacted last week demonstrates the extent to which the Trump administration is prioritizing their AI war with China versus local industry (Nvidia). As defined in the IFR, Tier-2 nations are inadvertently caught up in this conflict since the U.S. administration sees them as potential backdoors for China to access U.S. technology. Accordingly, the Trump administration may accept even further impacts to Nvidia’s revenues by limiting Tier-2 nations to H20 GPUs (or strictly speaking, the performance limits that the H20 was constrained to). This would be a dramatic additional restriction versus the existing IFR as shown below. Accordingly, we call this scenario the “Shift Left” scenario.
Wolf in Bear’s Clothing scenario – As mentioned in the prior Aiolian update, one of the benefits of the Biden IFR - that the Trump administrations appreciates – is the ability to blunt China’s Digital Silk Road initiative via the extraterritorial requirements of the NVEU program. Borrowed from the U.S. experience with G42 in the UAE, the program includes expulsion concepts that force NVEU-authorized companies to sever any ties to Chinese companies and eliminate any dependency on Chinese technology (e.g. Huawei, ZTE networking technology). So while the U.S. doesn’t believe their restrictive AI export policies are leaving a strategic opening worldwide for China’s GPUs and AI influence, realiable sources tell us that the Trump administration prioritizes this benefit of the current IFR. To maintain the motivation of Tier-2 AI cloud companies to “suck it up” and accept those expulsion requirement, the IFR would have to continue to offer high-performance GPUs (albeit with aggregate TPP limits). That’s why we believe that this scenario is more likely than the “Shift Left” scenario which would relegate Tier-2 nations to H20 levels of performance.
Of course, Trump will not want to be perceived as adopting Biden’s IFR unchanged. We expect that BIS will close several loopholes as shown above. One would be to close or limit the loophole on lower volume imports (i.e. the "License Exception Low Processing Performance, LPP") as advocated by Anthropic (footnote 14). Another would be to close the loophole that allows Chinese clients (or subsidiaries thereof) to use Tier-2 Infrastructure-as-a-service for workloads other than creating frontier-models (ECCN 4E091). The “Acceptable Use Policies” (footnote 15) described in the regulation currently doesn’t prohibit Chinese clients (e.g. Bytedance) from using NVEU company infrastructure for non-frontier workloads (e.g. fine-tuning, inference, test-time compute, post-training scaling). And while we don’t believe BIS will limit access to high-performance GPUs, they may lower some of the aggregate TPP limits. All of these additional restrictions would provide Trump with enough of a platform to loudly proclaim that Biden was weak and created a mess, and that he had to fix it with something that really stands up to China.
Finally, another sign that the BIS is moving ahead with the IFR is via Aiolian’s current engagement with the department as we move ahead with NVEU applications
In summary, at Aiolian International, our confident prediction is that the U.S. will move ahead with the “Wolf in Bear’s clothing” scenario as described above. We are fully aware that our clients have been stockpiling GPUs which buys some time, but the NVEU application process timeline is unpredictable and we urge clients to start planning now. We provide these update and analyses to enable our clients to do educated planning and scenario considerations. Our mission at Aiolian International is to help developing nations and their local companies create sovereign AI infrastructure while navigating the evolving US regulations for AI technology exports. We provide a unique three-pillar combination of AI cloud strategic planning, regulatory consulting and IFR authorizations, and systems for regulatory compliance. Please don’t hesitate to contact us if you have questions.
References:
The above content uses materials under fair use principles for commentary and criticism.
1 - For “License Exception Low Processing Performance (LPP)”. See the new section § 740.29 of the US Export Regulations
2 - Calculation based on Nvidia Blackwell B200 with 20,000 FP4 TFLOPs = 80,000 TPP. US AI export regulation limits low volume imports to 26.9M TPPs
3 - Per-country authorization and allocations as described in new section § 742.6 of the US Export Regulations
4 - Based on same B200 stats as above. US AI export regulation limits each country to 790M TPP cumulative over three years
5 - The National Validated End User (NVEU) authorization described in new section § 748.15 of the US Export Regulations
6 - Nvidia, caught in the US-China trade war, takes a $5.5 billion hit, April 15th, 2025, CNN Business
7 - § 748.15 Authorization Validated End-User (VEU), Bureau of Industry and Security
8 - “The Trump administration can avoid a strategic misstep in the AI global race", February 27th, 2025, Brad Smith, Microsoft.
9 - “Export Control Diffusion Confusion", Ken Glueck, Executive Vice President, Oracle, January 5th, 2025
10 - “Cloudmatrix 384 launch” Webull, April 10th 2025
11 - “Huawei AI CloudMatrix 384 – China’s Answer to Nvidia GB200 NVL72”, SemiAnalysis, April 16th, 2025.
12 - “Is-Informed Letter, BIS ISSUES NEW GUIDANCE ON DIVERSION RISK", Lewis Brisbois, July 26th, 2024
13 - “Nvidia CEO stresses importance of China market in Beijing visit, Chinese state media reports" , Reuters, April 17th
14 - “Anthropic’s Recommendations to OSTP for the U.S. AI Action Plan”, Anthropic, March 6th, 2025
15 - “Acceptable Use Policies, Certification and Policy Requirements Relating to National and Universal Validated End Users (VEU)", Bureau of Industry and Security