DeepSeek shot to fame, what does the US government think?
2025-05-19
Reposted from 东不压桥研究院
With DeepSeek rapidly gaining widespread attention over the past few days, the trend of public discourse in the US has begun to shift. Initially marked by surprise, admiration, and praise, discussions have increasingly been accompanied by skepticism, resentment, and even hostility.
Alexandr Wang, CEO of data platform Scale AI, has firmly asserted that DeepSeek possesses 50,000 smuggled H100 chips but has refrained from publicly admitting it due to US export control violations. Given Scale AI's ongoing efforts to integrate itself into the US "defense technology" sector, his assertion is unsurprising. The company's former executive, Michael Kratsios, now serves as the Director of the Office of Science and Technology Policy (OSTP), and Alexandr Wang himself has strong incentives to demonstrate his patriotism. However, he has provided no evidence for his claims, and Nvidia has publicly refuted the allegation. Nonetheless, this conveniently explains DeepSeek's sudden rise in capabilities, leading many Americans to believe it.
Currently, the US Congress and White House remain preoccupied with the ongoing transition of government and have yet to issue a response to DeepSeek. Considering the high level of attention the topic has garnered, it is almost certain that DeepSeek has already entered the radar of the Trump administration. This low-profile company, primarily focused on model development, has now been involuntarily drawn into the geopolitics of US-China AI competition, whose decision-making and strategic moves can no longer ignore this factor. For Liang Wenfeng, DeepSeek's founder, and his team of just over 100 employees, this presents a significant challenge.
The rise of DeepSeek primarily impacts the US open-source model ecosystem and Nvidia's stock price. However, for the Trump administration, which is currently reviewing Biden-era AI policies and has recently announced a $500 billion AI infrastructure plan, DeepSeek's emergence represents a major disruption that could directly influence future policy directions.
While the stance of the new administration remains unclear, it has undoubtedly been forced to recognize both the intensity of US-China AI competition and China's formidable strength in this domain. As The Economist aptly titled its article, "Chinese AI is catching up, posing a dilemma for Donald Trump." Moving forward, a key question will be whether Trump prioritizes advancing US AI capabilities or restricting China's progress.
Against this backdrop, growing hostility within the US toward DeepSeek is unsurprising. In particular, an illustrative example of this sentiment came from Joshua Kushner, founding partner of Thrive Capital, also an early investor in OpenAI. In a post on X, he expressed strong concerns, "'Pro America' technologists openly supporting a Chinese model that was trained off of leading US frontier models," he whined, "with chips that likely violate export controls, and—according to their own terms of service—take US customer data back to China."
Three allegations—violating chip export controls, appropriating US models, and transferring American user data to China—are serious and warrant close attention. After all, such views likely represent a broader faction of "tech nationalists" in the US and could directly influence both Congress and the federal government.
It is highly probable that there will be an investigation into whether DeepSeek violated chip export controls of the US. But the greater issue is how its emergence will shape the Trump's evaluation of existing chip export policies.
One of DeepSeek's achievements is to demonstrate an alternative approach: producing cutting-edge models at minimal cost despite constraints on computing power. This directly challenges the previous assumption that model performance could only be improved through massive computational resources. This explains why Nvidia's stock has plummeted by 17% in recent days too. Naturally, some are now questioning whether high-end GPUs will remain essential in the future, and if not, what purpose US restrictions on China's access to advanced chips still serve.
Over the past few days, this issue has sparked intense debate within the US. Some analysts argue that DeepSeek's success proves that Biden's chip export controls have not only failed but may have backfired—forcing China to innovate more efficient methods of training powerful AI models with fewer chips. Others contend that export controls naturally have a delayed impact, and their effectiveness will only become apparent over time. The Biden administration's last-minute introduction of the "AI Diffusion Export Control Framework" was designed to address previous loopholes. Furthermore, even Liang Wenfeng has acknowledged that the biggest constraint on their operations remains the high-end chip embargo. This raises the question: if chip controls are ineffective, why does DeepSeek still face such limitations? Another perspective is that DeepSeek has been training its models using Nvidia's H800 GPUs, which were permitted for export to China at the time. Some now argue that the US should never have allowed even these chips to be exported, as it has effectively "nurtured a rival."
These differing viewpoints will inevitably influence US government decision-making. However, future policies will be adopted depending on which viewpoints are absorbed. American chipmakers and semiconductor equipment manufacturers or firms like Nvidia are likely to use DeepSeek as an example to argue for the removal of chip export restrictions.
However, the pro-export-control camp remains strong. Nvidia's recent response emphasized that DeepSeek's "dynamic inference-time scaling" approach still requires large numbers of GPUs. Currently, GPUs such as the H20, which are not subject to export controls, have demonstrated superior inference performance compared to the H100. If the DeepSeek incident prompts the US government to consider not only restricting China's model training but also its inference capabilities, it is conceivable that a ban on exporting H20 chips to China could follow. During the final months of the Biden administration, Zhipu AI became the first Chinese large model company to be added to the US Entity List. Given Trump's enthusiasm for sanctions lists during his first term, it would not be surprising if more Chinese AI companies were subjected to similar measures.
The so-called "training on cutting-edge US models" may also prompt a broader reassessment of the US open-source model governance, which is a particularly complex issue. The US tech industry has a deep-rooted culture, and prevailing opinion holds that open-source development has contributed to America's technological leadership. The Biden administration considered imposing export controls on open-source models but ultimately abandoned the idea, both due to legal constraints and because US judicial precedents have repeatedly affirmed that code constitutes an expression of ideas and is therefore protected under the First Amendment. Any attempt to impose restrictions on open-source code would lack legal basis and could even be deemed unconstitutional. As a result, apart from encryption software with military or security significance, US export control laws generally do not regulate open-source software.
Although DeepSeek and Llama employ different model architectures and it was not directly developed using any specific US open-source model, some believe that DeepSeek has built upon Llama's foundations. This has sparked debates over whether Meta should have released an open-source model in the first place—after all, it enabled Chinese researchers to not only learn from it but surpass their American counterparts. While such concerns may seem unreasonable from an open-source community perspective, they could carry weight in policy circles. The US government may now reconsider whether to restrict China's access to open-source models.
At present, leading global platforms for hosting open-source model code and weights, such as Hugging Face and GitHub, are operated under US legal jurisdiction. While GitHub's open-source model code is governed by its respective licenses, it could also be subject to US export controls and other regulations affecting upload and download behavior as information in GitHub.
While it is unlikely that the US would go as far as restricting Chinese developers' access to open-source platforms since doing so could push them toward Chinese alternatives like OpenI to enrich the open-source ecology.
It is worth noting that such precedents do exist. In July 2019, GitHub restricted Iranian users' access to private repositories and certain public repositories to comply with US sanctions, even suspending some developer accounts, leading to many programs being interrupted. This incident drew widespread criticism from the global open-source community, forcing GitHub to seek export control exemptions from the US Treasury Department. But the lesson is not far from the truth, and it is worthwhile for Chinese open-source model developers to be vigilant.
Joshua Kushner's final reference to "taking US customer data back to China" is essentially the same old issue of cross-border data flows between the US and China, and one level up is the issue of Chinese big model apps being available in the US app stores for American users to download and use. Trump's first term, had tried to use the International Emergency Economic Powers Act (IEEPA) as the basis to force the US app stores to take down Chinese apps, and ultimately subject to the Berman Amendment in the legal proceedings lost, and end up with nothing definite because of the subsequent election defeat. Whether Trump will continue his previous policy of targeting Chinese apps this term is unclear.
Like Xiaohongshu (rednote), DeepSeek's AI apps remotely provide services to American users, falling under the jurisdiction of the Protecting Americans from Foreign Adversary Controlled Applications Act, the Department of Commerce's Final Rule on Information and Communications Technology and Services, and other laws and regulations. The US could theoretically use these legal tools to restrict and take down these apps. It would be difficult for these apps to serve US users without generating some cross-border interactions of data. However, the Provisions Pertaining to Preventing Access to US Sensitive Personal Data and Government-Related Data by Countries of Concern or Covered Persons (PPPAUS) issued by the US Department of Justice in December last year, will come into effect on March 27 this year, at which time cross-border interactions of data generated by the operation of these apps in the US will also be directly subject to the supervision of this federal regulation.
For DeepSeek, both the US and China would do well to cool their reactions moderately. For China, it is important to celebrate and be proud, but at the same time alert to the crisis, to prevent "destroying it by heaping excessive praise on it" from the other side of the ocean. For the US, it is important to look at DeepSeek's success objectively and rationally, avoiding geopolitization and national security. But perhaps in any case, the past few days have had a far-reaching impact on the next stage of the AI competition between China and the US.
