US planning to ban DeepSeek?

A range of options is under consideration in Washington DC, but no clear strategy has emerged yet

Rumors are flying in Washington about how the new Trump administration will choose to respond to the challenge of Chinese AI upstart DeepSeek. The emergence of DeepSeek, as documented in this Substack, has roiled Washington and some other western capitals. Some governments, such as South Korean and Taiwan, have moved to ban downloads of the DeepSeek smartphone app, as have some US states including New York, Virginia, and Texas. Montana Department of Justice officials sent a letter this week to Congress calling for a ban on DeepSeek on government devices. But what are the authorities and options for US officials pondering what to do about DeepSeek?

“Congress should protect America’s national security by banning DeepSeek on government devices. If it has not already taken action to administratively ban DeepSeek, we trust that the Trump Administration would swiftly implement this ban to protect our national security from America’s ‘potent and dangerous’ adversary.” —Attorneys general from 21 states in letter to Congress.

The easy, the hard, and the really tough

The policy responses and authorities for dealing with DeepSeek as a national security problem are limited and complex: short- and long-term, executive, rule-making, and potentially legislative, depending on what goal the new administration is seeking to achieve. The Wall Street Journal reported this week that the administration was considering action against DeepSeek soon.

Given that other western governments have taken the minimal action of banning government workers from using the DeepSeek app, pressure is almost certainly growing on the new administration to do something. The easiest action to take would be an executive order (EO) banning downloads of the DeepSeek app by government employees. The US government has done this before for Chinese apps, such as TikTok, with a memo from the Office of Management and Budget. The OMB directive required all federal agencies to remove TikTok from government devices within 30 days, and also required agencies to prohibit internet traffic from government devices to TikTok. A similar approach could be used for DeepSeek. The impact of these types of bans is likely quite limited, as government employees are not likely downloading these types of apps in large numbers. The TikTok ban was almost certainly unnoticed by TikTok, with over 100 million active US users.

The next and more difficult step would be an executive order directing Google and Apple to remove DeepSeek from their app stores. This could be done in several ways, including using an EO based on the International Emergency Economic Powers Act (IEEPA), which allows the President to regulate international commerce after declaring a national emergency. This approach was used as the basis for executive orders targeting TikTok and WeChat in 2020, but these were successfully challenged in court and the requirements for executing the ban proved to be complex given the nature of the TikTok support system in the US. This would require a presidential declaration of a “national emergency” related to an “unusual and extraordinary threat”. This would seem hard to see a clear justification for in the case of DeepSeek. The President could also use an EO Based on National Security, citing national security concerns. This would likely require demonstrating a specific threat posed by DeepSeek, which is typically in the form of evidence of ties to a foreign adversary or evidence of specific harmful activities.

The EO approach is tricky, because unlike TikTok, which was already being used by tens of millions of US users and had a complex data and operational footprint in 2020, making the IEEPA argument of a “national emergency” at least plausible (that “national emergency continues to this day”), the case for DeepSeek posing either an economic emergency or a national security threat is pretty weak. DeepSeek’s V3 and R1 models are open source/weight models that have been released via a smartphone app and are widely hosted in the US by cloud service providers and open source model repositories such as Github and Hugging Face. As I have previously noted, this specific situation was not likely anticipated by US policymakers looking at China, technology controls, and AI.

Despite the slim case for DeepSeek being an economic or national security threat—the same holds for TikTok, of course—the administration could make the case based on the smartphone app being linked to inference servers in China. While there is zero evidence that DeepSeek, a small AI R&D organization, has any close ties to the Chinese government, as I have documented, the administration could use old justifications such as China’s National Security Law—which does not, as many claim, require Chinese companies to hand over data to the government—as sufficient to make the economic or national security claim. Here, unlike deep pocketed TikTok and TikTok influencers, DeepSeek would not be able to challenge such an action in US courts, nor would they have any interest in doing so.

There is also consideration being given to ordering cloud service providers (CSPs) to stop hosting Chinese open source AI models, including DeepSeek and Alibaba. More Chinese companies are open sourcing elements of their models. The concerns here are twofold but complex, particularly because they would require the US government to intervene in the open source community, which has warmly welcomed DeepSeek, as noted in earlier posts in AI Stack Decrypted. First, any national security argument around data transfer to China would not hold, as all of the CSPs and other players such as Perplexity are hosting the API on their US-based servers and there is no data or queries being transferred back to servers in China. Second, there are no easy authorities to use here other than the Commerce Department ICTS supply chain rule. But it is not clear that DeepSeek open source/eight models would fall under the jurisdiction of the rule, which has been used as the authority for rules on connected vehicles and drones. Because there is no connection between running DeepSeek open source models on US servers and China, similar justifications about a potential threat to US critical infrastructure would be hard to make here. In addition, any effort to restrict hosting of open source models on US servers would generate major pushback from the open source community, which so far has not been the subject of US export controls or technology restrictions. It remains unclear whether the Trump administration is willing to open up this difficult front, and it is not clear what the justification or the gain would be.

Given all this, it is still likely that the administration will take some combination of the above measures, and could even consider working with Congress on legislation around these issues. One reason for doing so would be to avoid having a situation where Chinese AI frontier model API hosting on US servers becomes normalized and even creates new dependencies as more US companies leverage the model, undermining one of the overall justifications for controlling GPUs that the Biden administration made for more than two years. The Biden administration was vocal in justifying extensive export controls that have also had major collateral damage for US technology leaders, which I have documented frequently over the past two years, on the basis that they would slow the ability of Chinese companies to develop frontier AI models. If, after all this effort, billions in lost revenue, and disruption of supply chains inviting Chinese retaliation against US firms in China, Chinese frontier models are routinely hosted and used to build applications in the US, this would appear to undermine the entire premise of the export controls. This alone will likely propel a strong US response to DeepSeek and other Chinese open source AI models.

In addition, the growing association of DeepSeek with Chinese technology giant Huawei could also provide a rationale for the Commerce Department to put DeepSeek on the Entity List. Previously, companies with only minimal and often undocumented ties to Huawei have been put on the Entity List with the justification that the listed firm could be engaged in supplying Huawei with technology support. This would be a bit of a stretch, as the collaboration between DeepSeek and Huawei appears to be limited and revolve around optimizing Huawei hardware support for running DeepSeek’s models. As I noted in a recent post, many other Chinese GPU hardware design firms (MetaX, Moore Threads, Intellifusion, etc.) are doing the same. If this route is also chosen, it is likely that the Commerce Department would place a much broader set of Chinese companies on the Entity List. This opens up another can of worms though, as major e-commerce and social media companies such as Alibaba and Tencent have so far not been targeted for Entity List action, and many US companies rely on their applications and support for business operations in China. Alibaba has also recently open sourced some of its leading QwQ AI models that are also hosted on US systems such as Github and Hugging Face.

Where do we go from here?

Finally there is the H20 control issue which keeps coming up (and which I addressed to some degree here). Interestingly, an anonymous post on Substack recently attempted to make the case for banning the H20 because 1) with emphasis in the industry shifting from training to inference, US controls need to be adjusted, and there is a need to consider “restricting chips like the NVIDIA H20 before their strategic importance escalates further,” and 2) in the wake of DeepSeek, the US must consider a new approach to open source/weight models “to ensure that its diffusion benefits reinforce, rather than undermine, U.S. national AI leadership.” I had a lengthy discussion with ChatGPT-4.5 about who the anonymous author could be and why he/she would choose to be anonymous, and I and others have our small list of candidates for the identity of the anonymous commentator and the reasons for posting anonymously which reveal the individual’s real motivations.

As one official from a leading industry player confided to me, the nameless author “acts like the importance of inference is a “complete surprise”, when the transition to inference has been the point of AI the entire time. No point in training super expensive models if you can’t get users to use them.” So why didn’t anyone think of this in October 2022? Yes, post training and compute (test) time scaling are now much more important, and DeepSeek’s innovations include leveraging these trends, as I noted here. But after multiple senior US officials have told us repeatedly that the goal of the export controls is to prevent China from training frontier models, moving the goal posts, yet again, to using inference, is just a bit of a stretch—and almost certainly too late to prevent the widespread use of the H20 for inference in China. And again, the vast majority of that inference compute will be for consumer applications, chatbots, coding, healthcare platforms, etc. Is the US government really trying to both further undermine Nvidia’s China market access to prevent the average person in China from using AI on their smartphones or as part of business operations or healthcare consultations? The same holds for open source: taking on the open source community because Chinese companies have developed very capable open source/weight models is highly problematic to justify, as noted above. Even the usual “mil civ fusion” standard throwaway justification does not hold any serious water here. But an anonymous technology guy thinks these are great ideas.

Clearly this situation has become complex and will require a nuanced response from the US government to avoid creating new and unexpected collateral damage to industry and to US China relations. Whether those within the Beltway are consulting experts across the AI technology stack before issuing new rules or policies remains unclear. Given the importance of open source models and the open source community, the administration will need to be careful in considering potential options to respond to the DeepSeek Effect.