Quick Take: Commerce "Rescinds" AI Diffusion Rule, Replacement Coming
Guidance on Huawei Ascend may be more important, as the Geopolitics of GPUs heats up
[Updated 5/13/2025]
Today the Commerce Department’s Bureau of Industry and Security released a statement formally announcing it was rescinding the Biden era AI Diffusion Framework. That this revocation of the rule was inevitable, just before it was set to take effect on May 15, has been known for many weeks. Quick take on what will replace the rule.
As I have documented in this Substack, there was lot of confusion around the original AI Diffusion Rule, along with opposition from industry and countries relegated to Tier 2 status that eventually led to its pullback. In addition, the pullback was almost certainly driven by Trump administration priorities around AI and the Middle East.
The BIS announcement stressed that the rule would eventually be replaced. We can already see the outlines of what will replace it. First, there will be a series of bilateral deals, of the type Trump has already struck with Saudi Arabia and will shortly do the UAE this week, which involve access to advanced US origin GPUs in exchange for major investments in the US and some agreement around controls on AI data center infrastructure and things like KYC requirements.
Second, BIS issued three critical sets of “guidance” as part of the announcement:
Issuing guidance that using Huawei Ascend chips anywhere in the world violates U.S. export controls. Full guidance here.
Issuing guidance warning the public about the potential consequences of allowing U.S. AI chips to be used for training and inference of Chinese AI models. Full guidance here.
Issuing guidance to U.S. companies on how to protect supply chains against diversion tactics. Full guidance here.
All of these are part of the overall package of measures that will constitute the replacement for the AI Diffusion Framework. Let’s translate each of them briefly:
Huawei Ascend guidance: This is a clear warning to any country or company that if it chooses in the future to purchase Huawei AI hardware and associated software services that this could open up the risk of penalties under US export control laws. It remains unclear how this would be enforced, and what happens with existing companies in China, including all the leading AI developers, currently using Ascend 910X based clusters.
Using US AI chips guidance: This is part of a another warning that is attempting to get companies, particularly second tier AI cloud services (IaaS) providers, to be concerned that allowing Chinese companies to access remote hardware that includes GPUs from Nvidia, AMD, or Intel for training or inference of models, could come under scrutiny. The language is important here because it makes clear that there are currently NO RULES in place that would make this illegal under US export control law, so companies need to be aware of the “potential consequences” of this activity.
Diversion guidance: This is guidance without threatened consequences, rather offering help in protecting supply chains against “diversion tactics.” This suggests that the Commerce Department will offer to send teams to companies with complex supply chains to discuss methods used by middle men to divert shipments of advanced GPUs to China, or set up data centers in third countries, receive GPU servers, and then move them to China. As I have noted, there is no evidence that this type of diversion is happening at scale, for example, in the 10s of thousands of GPUs, and none of the leading Chinese AI companies are likely using diverted or smuggled GPUs. A serious number of GPUs, say 50,000, in any case would require support from Nvidia and the firm and its distributors would be aware of such a large stock of restricted GPUs ending up in China. The listing the second AI chips guidance, without an clear legal limitation, suggests that the Commerce Department now realizes that this is a much more likely scenario than diversion: Chinese using IaaS tools provided by AI data center operators outside of China, in numbers that would be way beyond those available by mere “diversion.”
Interestingly, for perhaps the first time (checking this), BIS stressed the use of GPUs for inference (misspelled “interference” in the document), indicating, as I have noted before, the expansion of original justifications for GPU controls focused on training frontier models, to inference operations, in the wake of technology developments such as test time scaling and reinforcement learning post-training. This was the reason that Commerce added the Nvidia H20 GPU and similar performance GPUs from AMD and Intel to licensing requirements via “is informed” letters in early April. This move of course cost Nvidia a whopping $5.5 billion dollars in a single write-off and likely $15-20 billion overall.
[NEW ANALYSIS]
The full guidance that came out on May 13 adds a few new important wrinkles to the guidance:
The BIS guidance on Huawei processors/GPUs includes the 910B, 910C, and 910D. The guidance to “the public” is that it is possible that these semiconductors were developed or produced in violation of US export controls. “BIS is warning that, pursuant to GP10, the use of such PRC advanced-computing ICs risks violating U.S. export controls and may subject companies to BIS enforcement action. Here, GP10 is an arcane piece of the export administration regulations (EAR) that attempts to be a catch all for companies about potential knowledge that a violation “is about to occur” or is “intended to occur.” The issue here is the use of US semiconductor manufacturing tools by domestic foundry leader SMIC to produce the Ascend 9XX series, and the use of a third party company by Huawei to use TSMC to manufacture the Ascend last year, as documented in this Substack. How either US toolmakers or TSMC would have known that a violation “was about to occur” remains unclear. SMIC, for example, had all the tools it used for manufacturing some versions of the Ascend 9XX series before export controls were imposed on tools in October 2022. US toolmakers lost access to their tools in China after this date as personnel were forced to pull out, meaning their ability to know how the tools were used went away. This accounts for the major degree of unknowables here, as BIS acknowledges by saying "these chips were likely developed in violation of US export controls.” In fact we do not know, and neither does BIS know, which exact Ascend 910X semiconductors were produced in violation of anything. BIS could assert that the FDPR rule that applied to Huawei in 2020 would preclude SMIC from manufacturing Ascends on behalf of Huawei, but neither SMIC nor Huawei has acknowledged where the Ascends are manufactured, some industry sources have suggested Huawei itself has manufacturing capability, and applying FDPR to a Chinese company manufacturing on behalf of another Chinese company has not been tested in terms of applicability and enforcements. And the BIS language suggests some uncertainty here also. Nevertheless, despite the fact that a potential user of the Ascend 910X and systems based on it would have no way of knowing whether the specific ICs violated to export controls, the guidance is designed to instill doubt about how advisable this is and threaten unspecified penalties, even though BIS admits that it does not know for certain whether a specific Huawei IC was produced in violation of US controls.
GP10 states, in relevant part: “You may not sell, transfer, export, reexport, finance, order, buy, remove, conceal, store, use, loan, dispose of, transport, forward, or otherwise service, in whole or in part, any item subject to the EAR and exported or to be exported with knowledge that a violation of the EAR… is about to occur, or is intended to occur in connection with the item.”
For the guidance on use of US advancing computing ICs to train AI models, BIS notes that it has “determined that access to advanced computing integrated circuits (ICs) and commodities subject to the EAR for training AI models has the potential to enable military-intelligence and weapons of mass destruction (WMD) end uses.” The thrust of the guidance is companies that are providing Infrastructure as a Service (IaaS), in other words data center providers, and again the key is that the IaaS provider has “knowledge” that the IaaS provider will use these items to conduct training of AI models for Chinese companies. Again there is a knowledge gap as not IaaS provider will be in a position to “know” what the potential end use of an AI model will be. In addition, the most likely users of IaaS services outside China are in fact companies that are developing AI models specifically for civilian, not miliary or WMD related end uses. So the IaaS service provider, on the basis of knowing the business model of a specific Chinese company, could argue that there is no reason to believe that a model being trained on its infrastructure by a private sector commercial Chinese company could be linked to a military end use. Nevertheless, BIS has included an interesting twist in the guidance, threatening that “In addition, foreign parties acting contrary to U.S. national security and foreign policy interests, including by training AI models that could support WMD or military-intelligence end uses for or on behalf of parties headquartered in Country Group D:5 (including China) or Macau, may be added to the Entity List, even where no violation of the EAR occurs.” This is quite a threat, which in essence means, BIS itself will somehow determine whether an AI model being training via an IaaS cloud services providers could have a military end use and if it decides that this is the case, the IaaS provider risks being added to the Entity List. How BIS will determine this remains unclear.
Finally, the guidance on diversion contains lots of details that BIS advises companies to consider in conducting due diligence on customers, essentially a know your customer recommendation. Significantly, BIS sets a threshold of 10 MW as a trigger for determining whether a particular data center could contain advanced ICs that support training advanced AI models of concern. BIS suggests sellers of advanced GPU based systems “Evaluate data centers to determine whether they have the infrastructure to operate servers containing advanced ICs greater than 10 megawatts. Data centers at or above this threshold merit additional scrutiny as they may be able to provide access to a large quantity of advanced computing ICs for training AI models for or on behalf of parties headquartered in countries of concern, where such activities may support WMD or military-intelligence end uses/end users.” Here 10 MW is not a particularly large data center, and there are many other potential uses of cloud services in such a data center beyond training models that could “support WMD or military intelligence end use/end users”. How the seller is supposed to determine the risks here also remains unclear. As I have previously noted, it seems unlikely that the diversion/smuggling of advanced GPUs is a significant issue, as none of the major Chinese commercial private sector companies leading AI development in China are using such GPUs. The guidance appears primarily directed at cases such as the one media reported on in Malaysia, where some number of AI servers were installed in a data center in Malaysia, and then re-exported to a China based end user. This case did not involve large numbers of GPUs and it is not clear how common such diversions are, and in this case, it is also unlikely that the leading Chinese AI developers are using this method to acquire export restricted GPUs.
In sum, BIS is using all of this new guidance to threaten companies with unspecified administration actions and penalties for conducting what are legal transactions with customers based on unclear criteria for knowing about the end use of AI models, which companies involved in such transactions and services will have no way of knowing. This is an attempt to extend “long arm jurisdiction” on advanced compute for AI well beyond existing regulations. How any of this will be enforced and on what legal basis determinations will be made around “knowledge” of end uses remains unclear.
Middle East deals presage new AI rule
At the same time as the BIS guidance was released, major deals around access to advanced GPUs were being signed in the Middle East between companies from Saudi Arabia and the UAE, and US technology leaders.
As part of the $600 billion deal Trump inked with Saudi Arabia this week, new Saudi AI company Humain will gain access to Nvidia’s most advance GPUs, and Saudi firm Datavolt will invest $20 billion in the US in AI data centers and associated energy infrastructure. The Nvidia deal is particularly notable, and involves the say of 10s of thousands of advanced GPUs—the first sale will be a cluster of 18,000 Blackwell GB200 GPUs headed for Humain, just established this week, presumably for this precise purpose. Humain is headed by Crown Prince Mohammed bin Salman (MBS), and falls under the Public Investment Fund (PIF), and will provide a range of AI related services, including AI data centers and other AI infrastructure, cloud services, and development of advanced AI models, according to Saudi media sources.
In addition, Amazon inked a deal with Humain worth some $5 billion to build an AI Zone in the Kingdom, and GPU major AMD announced a $10 billion deal with Humain to supply its GPUs for an AI data center in the country. Finally, US networking giant Cisco announced a deal with AI UAE AI and technology leader G42, to help the company develop its AI sector—more such deals are likely later in the week. No Tier 2 countries here.
The AI infrastructure deals were negotiated by a complex team from the Trump White House and key regional power brokers. The US side is being led by White House AI and crypto czar David Sacks and AI advisor Sriram Krishnan, who are pushing for major sales of advanced U.S. GPUs to firms such as Humain and G42. A major role has also been played by Emirate sovereign wealth fund Mubadala chief Khaldoon Al Mubarak.
Sacks’ approach has already run afoul of China critics in Congress. Concerns were raised last summer when G42 was seeking to invest in AI infrastructure and acquire advanced U.S. AI hardware. Of course it was specifically this dynamic that led the Biden administration to craft the AI Diffusion Framework, to erect guardrails around sales of large numbers of GPUs to countries outside the U.S. Both Saudi Arabia and UAE were consigned to Tier 2 status under the Biden rule, and it was no coincidence that the Commerce Department this week announced it was pulling back the rule. Sacks is also increasingly at loggerheads with national security circles within the administration, as he moves forward with talks, in part over disputes about whether Saudi Arabia can continue to use Huawei equipment in facilities where advanced U.S. hardware would be installed. More on this later.
Looking ahead, it remains unclear how other countries once relegated to Tier 2 status under the now defunct AI Diffusion Framework will fare with bilateral agreements and quid pro quo investments in the US. The new rule, likely to come out in the next several months, will almost certainly cover the terms of both a global licensing issue, potentially with some compute caps, and what provisions bilateral agreements will be required to have, along with enforcement measures. We are definitely in a whole new world of AI diffusion….more on this as details become clearer around new rule and bilateral deals.