American AI Dominance

To win on AI, the USG needs a vision for leading the world

A purely defensive posture by the U.S. towards the AI competition with China is inadequate. DeepSeek’s accomplishment of achieving model performance that measures up to leading American model developers should serve as a reminder of this. A positive vision for American AI dominance will be critical to ensuring that state-of-the-art AI is beneficial for the world at large.

AI policy and technical discourse is highly susceptible to herd mentality. A couple of months ago, the consensus was that scaling laws no longer worked - until the details about OpenAI’s O3 and subsequent models proved that broad assertion to not quite be true. This past week, commentators have overplayed the accomplishments of DeepSeek as somehow rendering compute unimportant as a differentiator in AI development, while those in industry and policy who disagree with that take are fixating on export controls as the only important place to be focused for ensuring American leadership in AI. Neither point-of-view is fully accurate.

It’s been discussed ad nauseam by policy folks at this point that export controls are not to blame for the launch of DeepSeek V3 and DeepSeek R1 and that export controls are critical for U.S. national security interests, so I won’t spend too much time on those points. Dario Amodei, the founder and CEO of Anthropic, compellingly argued in an essay this week that export controls are vital to “keeping democratic nations at the forefront of AI development.” I agree, but also believe this is not the only important topic for the policy conversation to be focused on.

Export controls will stop some chips from getting from the U.S. to China and make it more difficult for China to take advantage of computational scale for training and inference in their AI ecosystem over time. Total accessible volume of compute for a given company/country will continue to be a key differentiator in speed of advancement in AI R&D. But export controls alone will certainly not make it impossible for China to leverage advanced AI in consequential ways or to continue pushing at various edges of AI research and development.

Over time, it is quite possible for China to continue to stay on the heels of the United States through prolonged campaigns to erode the willingness of countries to commit to such trade regimes (possibly in concert with subtle forms of economic and military coercion to reduce that willingness), domestic efforts to ramp up chip production and quality and data center construction, and continuing to foster the enormous amount of domestic research talent that will be optimizing for building AI in a resource-constrained environment. For the United States, complacency will be the enemy of success.

American AI dominance matters because the regime under which state-of-the-art AI models are developed and released matters. The Chinese government has consistently shown a willingness to be aggressive towards minority populations, threaten the sovereignty of other countries, and generally flout international law. If Chinese developers achieve the superintelligent AI capabilities that AI company leaders forecast to arrive within the next year or two, it increases their ability to carry out these disruptive activities. Achieving a sufficiently advanced level of general-purpose AI capabilities could then become an asset in allowing the Chinese government to then prevent others from achieving those same capabilities and perhaps be a key asset in ensuring China becomes the most influential nation in the world for a very long period of time.

A key question to be asking vis-a-vis US-China AI competition (and a question I would argue is just as important as that of export control regimes) concerns defining what a positive vision for American AI dominance actually looks like. There has been no significant government-led effort to outline a compelling vision of tangible outcomes from which to measure the success of U.S. policy. In lieu of this, past administrations adopted a defensive posture and an incredibly hard-to-measure goal of maintaining a technological lead over China. Fleshing out a positive policy vision with measurable goals towards keeping the United States at the frontier of AI development will increase the likelihood that such a vision is actually realized - while also serving as a powerful tool with which to attract collaboration and partnership for vital allies such as the Netherlands, Taiwan, Japan, and Korea, where key parts of the AI supply chain are based.

The Trump administration recently put out an AI executive order that stated: “It is the policy of the United States to sustain and enhance America’s global AI dominance in order to promote human flourishing, economic competitiveness, and national security.” The EO requires an action plan to be drawn up to meet these goals over the next 6 months.

This EO is an opportunity to set forth a real vision for American AI dominance. If the Trump administration’s AI strategy is intended to be impactful, that plan needs to consist of outcomes that are much more substantial than a set of reports and plans (the traditional outputs of many past federal AI initiatives). It has been quite exciting to see the launch of the Stargate initiative to dramatically increase the total amount of domestically-based compute. Here are some more initiatives that the U.S. AI strategy should include as part of a positive vision for dominance:

• Place a wide array of bets on potential future leverage points in AI R&D and highly secure the research happening in these areas: Government-backed investment programs (such as NSF, DARPA, In-Q-Tel, DOD OSC, and others) should be geared towards spurring entrepreneurship in potentially high-leverage areas of AI research and development, including exploration of alternative GPU designs, using quantum computing for AI model development and deployment, blue-sky research regarding LLM usage in developing new model architectures, and other areas. There are more than a few potential paths that exist towards securing differentiated and proprietary advantages in further advanced AI capabilities at a national level. Anyone who tells you they know exactly what the future paths of model development look like over the coming decades is lying, especially given the manner in which LLMs themselves are increasingly being used to create new approaches to chip design and model architecture development. From a national security perspective, placing a wide variety of bets is critical. Equally critical is strong government support for enabling organizations pursuing these areas to set up research security practices that prevent intellectual property from leaving the United States.

• A minimum viable federal AI framework that creates absolute clarity that the U.S. is the best place for AI talent, investment, and entrepreneurship: A White House-backed effort should create and pass a “minimum viable product” for federal AI legislation, which focuses on putting together a framework that solidifies America’s position as the most attractive location for AI investment and talent, while also putting in place the bare essentials required to allow the government to monitor and intervene in development and deployment when necessary and with appropriate justification of threatened national interests. Diehard opponents of any AI regulation say such legislation would hinder U.S. competitiveness against China. But the opposite can also be true - a vacuum of regulatory clarity can easily create confusion and avoidance that erodes the amount of talent and investment occurring within the U.S. I expect the struggle for AI talent to matter at a national level at the level of individual researchers within the next year (if it does not already), much as was true regarding the struggle for nuclear talent. This erosion of U.S. advantage is especially true if other countries explicitly legislate clear and hospitable environments for AI development in the meantime.

• A comprehensive strategy to iteratively introduce state-of-the-art AI capabilities into the American and allied cyber arsenal and military capabilities: Under the prior administration, some small programs were carried out to pilot the usefulness of generative AI in various dual-use AI/cyber use-cases, including leveraging LLMs for vulnerability scanning in federal and military networks. Given the rapid pace at which AI model capabilities are advancing, it is vital to institute an ongoing program of experimentation, evaluation, and eventual introduction of LLM capabilities at various levels of cyber and military operations. There should be a complementary ongoing program of transferring some of these learnings to allies - both to ensure allied contributions to joint cyber and military activities are as effective as possible, as well as to build confidence and closer security bonds that keep these allies farther away from Chinese influence. China has a track record of introducing AI capabilities quite early-on into critical military areas (including nuclear security) and, at the moment, chances are high that they are integrating LLM-enabled tools into their overall military at a much faster pace than the United States.

The ability to proactively take steps that advance American AI dominance in a manner aligned with classical liberal values is likely to get dramatically tougher over the course of the next two years due to the likelihood of a Chinese invasion of Taiwan and related disruptions, in addition to the increasing awareness of the high stakes of access to AI resources among a broader array of wealth individuals, governments, and corporations. The U.S. should start aggressively putting forth and pursuing a comprehensive positive vision for American AI dominance now.