The emergence of DeepSeek's highly efficient and cost-effective AI model represents a potential inflection point in the global AI landscape and challenges prevailing assumptions about the resources required to develop and train frontier AI models. From a Washington DC policy perspective, we view the development as providing additional near-term political momentum behind the strengthening of the U.S. export control regime — both through a potential expansion in the scope of the controls and through enforcement. Our note goes through the likely implications from the policy perspective, including sticks (enhanced tech restrictions) and carrots (support for domestic AI investment).

What is DeepSeek? Chinese AI startup DeepSeek has garnered significant attention with the release of its open-source large language model (LLM) that reportedly achieves performance comparable to or surpassing leading models like OpenAI's GPT-4o and Claude 3.5 Sonnet. Headlines suggest that the latest model was trained in 2 months using ~2,000 Nvidia H800 chips, which cost roughly ~$6 million.
Sticks: implications for tech restrictions.
The short-term limitations of the current U.S. export control regime were clearly demonstrated by the DeepSeek release — opening new questions about the next steps for U.S. tech restrictions. While we expect to see a narrative emerge that DeepSeek proves the ineffectiveness of current export controls, it will most likely be interpreted by the Trump administration as a reason to tighten controls and further limit and track who has access to leading-edge technology.
We have long maintained that the new administration is likely to ramp up the aggressiveness of existing tech export controls, and DeepSeek provides a clear catalyst for an acceleration in both 1) scope and 2) enforcement, especially given that the new administration is already conducting a review of existing export control policies with an April 1 deadline. This announcement also challenges the assumption that the AI diffusion rules, released in the last week of the Biden administration, establishing country-level chip restrictions will be rescinded or fundamentally altered by the Trump administration. The DeepSeek announcement will make it politically difficult to loosen chip restrictions.
The America First Trade Policy executive order (EO) orders the following actions:
Secretaries of State and Commerce to “review the United States export control system and advise on modifications in light of developments involving strategic adversaries or geopolitical rivals as well as all other relevant national security and global considerations.”
Assess and make recommendations regarding how to maintain, obtain, and enhance the U.S.’ technological edge and “how to identify and eliminate loopholes in existing export controls”, with a focus on those that enable the transfer of strategic goods, software, services, and technologies to strategic rivals/proxies.
Assess and make recommendations regarding export control enforcement policies and practices, and enforcement mechanism to incentivize compliance by foreign countries, including appropriate trade and national security measures.
DeepSeek’s release falls squarely under the “developments involving strategic adversaries” category, and we would expect that any incoming recommendations would seek to (at least in part) reflect this.
Expanded scope? Between the broader desire of the new administration to continue tightening U.S. export controls (a theme that is frequently underappreciated and overshadowed by Trump’s tariff agenda) and the EO directives, the imposition of additional restrictions that expand the scope of chips impacted by U.S. export controls (e.g., by imposing restrictions further down the performance scale) is one potential outcome.
While the H800 chips that DeepSeek claims to have trained their V3 model on have been subject to U.S. export controls as of October 2023, the resource effectiveness and use of relatively less advanced chips (i.e., not H100 chips) will still open new questions as to whether the scope of current controls is sufficient. The intent of the export control regime is to deny and restrict China’s ability to develop its domestic semiconductor and AI ecosystems over the longer term — a goal which the current restrictions should, in theory, support and compound over a longer timeframe.
However, even with the relatively short timeframe since the imposition of controls restricting the H800 chips specifically, we would expect the new administration to react with some form of tightened controls given the optics of the situation and the administration’s demonstrated desire to preserve the semiconductor/AI gulf between the U.S. and China (see the “carrots” section). Given the very recent imposition of the AI diffusion rule, FDPR, and other restrictions, the immediate and specific next steps are less obvious at this stage, but we will continue to monitor for updates on this front.
Expanded enforcement?
An acceleration in enforcement remains another likely avenue, given that the H800 chips are subject to restrictions and in light of external (and as-of-yet unverified) claims that DeepSeek has 50,000 Nvidia H100 chips — which have been restricted since October 2022.
Diversion remains a material concern from an export control perspective, and the December imposition of both the foreign direct product rules (FDPRs) and the January AI diffusion country caps represent a step-up in the U.S.’ enforcement of these concerns. Additional pressure on third countries would be a logical next step from this perspective, and the use of tariffs to achieve this goal (or at least increase pressure) is one option we are monitoring. Trump’s Sunday threats to impose a 25% tariff on Colombia are a clear example of his increased willingness to expand the use of tariffs as a negotiating tool, and the EO explicitly suggests that trade measures (i.e., tariffs) could be a remedy for enforcement.
Carrots: implications for the U.S. AI agenda.
DC to accelerate AI investment in near term, potential for policy realignment. In the near term, DeepSeek's achievement is likely to pressure the U.S. into increased support for domestic AI development, most likely leading to increased federal investment in AI research and infrastructure
(CHIPS and Science Act for AI?). A reevaluation of current AI development strategies is likely, emphasizing not just size and scale of investment, but also efficiency and innovation in AI architectures.
Potential changes may include:
Increased funding: There may be a push for greater government funding for AI research, development, and infrastructure projects.
Regulatory adjustments: Trump revoked Biden’s 2023 executive order that sought to reduce potential risks posed by artificial intelligence. We view DeepSeek’s announcement as an opportunity for Trump to provide further specifics on the direction of AI policy, including further rollbacks on regulations that could be hindering AI innovation and deployment.
Talent retention and attraction: Workforce development and training programs are likely the next area to be prioritized in building infrastructure for the evolving AI ecosystem. AI workforce development and training programs will likely require collaboration among the
public sector, academia, and the private sector given the rapid advancement of AI systems.
Hyperscalers keep investing: As Microsoft has already committed $80B to AI infrastructure capex and $60B+ from Meta, we do not expect the hyperscalers will pull back meaningfully this year — although it raises the bar for them to improve performance relative to DeepSeek, given their access to the latest cutting-edge GPUs.
Impact to Stargate. While the $500 billion investment planned for Stargate remains substantial, there may be a reevaluation of how these resources are allocated, with a potential shift towards funding more diverse and innovative approaches to AI development. The project's objectives
may be adjusted to place greater emphasis on developing efficient AI training and inference methodologies, and its timeline might be accelerated in response to perceived threats to U.S. AI leadership. Additionally, the scope of Stargate might broaden to include more initiatives focused on AI model efficiency, while also exploring opportunities for international collaboration with allied nations to pool resources and talent to counter technological advancements from China.
Compute matters. As discussed in our AI deep dive note, firms that have access to the greatest amount of compute resources at the lowest total cost when developing and training models will be at a distinct advantage. Advancements in AI capabilities have been closely tied to increases in compute resources, with tech firms investing billions in massive data centers and specialized hardware. Notably, DeepSeek was able to demonstrate that progress can be made through training methodologies, potentially reducing the reliance on absolute computing power. While initially this “more can be done with less” approach may seem like a net negative, we view this efficiency gain as actually serving as an accelerant for AI workloads.
We would highlight that Microsoft CEO Satya Nadella’s recent comments on a recent podcast regarding Jevon’s Paradox (which argues that increased efficiency leads to increased consumption) support our view. In this context, we view DeepSeek’s innovations as suggesting that efficiencies in model training and deployment could ultimately drive greater demand for compute resources as AI applications become more widespread and accessible.
Is bigger always better? DeepSeek's achievement also challenges the idea that bigger is always better in AI model development. By significantly improving data quality and model architecture, DeepSeek demonstrated an alternative path to efficient model training that contrasts sharply with
the current arms race among hyperscalers to build ever-larger clusters of high-end GPUs. This shift in the AI landscape suggests that frontier AI capabilities might be achievable without the massive computational resources previously thought necessary. We expect this achievement to only further the debate around the quantifiable AI return on invested capital (ROIC) for large language models (LLM) training.
DeepSeek a positive for AI apps and adopters. DeepSeek's breakthrough is likely a positive for application software companies and enterprises leveraging or adopting AI technologies. By potentially lowering barriers to entry in terms of cost, it could allow firms to more easily and affordably integrate AI into their products and services. This could lead to increased competition and innovation in AI-powered applications, benefiting end-users through improved functionality and lower costs. While there are still many unknowns, if DeepSeek does catalyze a push toward efficiency for training LLMs (a big if), then it should accelerate enterprise adoption of AI use cases and create more innovation — generally supportive of Jevon’s Paradox.
Big Data is a Big Deal. This achievement also underscores the increasing value of distribution and unique data. Companies with vast and diverse datasets, such as YouTube, Facebook, Instagram, and X, stand to benefit, in our view. The ability to train efficient models on proprietary data could
become a key differentiator in the AI landscape, and potentially reshape competitive dynamics in the tech industry.
What is DeepSeek? Chinese AI startup DeepSeek has garnered significant attention with the release of its open-source large language model (LLM) that reportedly achieves performance comparable to or surpassing leading models like OpenAI's GPT-4o and Claude 3.5 Sonnet. Headlines suggest that the latest model was trained in 2 months using ~2,000 Nvidia H800 chips, which cost roughly ~$6 million.
To download a PDF of this report and see important investor disclosures, please click here.
Comentários