Power or Peril: U.S. Foreign Policy in the AI Arms Race

Henry Strahm

Introduction

Nearly four years after OpenAI released ChatGPT, the capabilities and adoption of advanced AI models have grown exponentially, paralleling past revolutionary technological advancements. As AI entered the workforce, productivity surged economic development. Despite the economic benefits of AI, it carries complex financial and national security risks. Investment is risky and potentially unsustainable. While AI agents are able to more efficiently find and exploit cybersecurity weaknesses, if an enemy of the U.S. possessed a strong enough model, it could cripple our nation’s cyber infrastructure.

Why AI?

If the risks associated with AI are so severe, then why do the markets and the U.S. government continue to invest in its development? First is the obvious economic growth that comes with investment. AI infrastructure: data centers, electrical networks, and computers are projected to reach 3.63% of the U.S. GDP by 2032, the largest for any industry in the history of the nation, beating the 19th-century railroad boom, the 1970s highway expansion, and most recently the telecom and fiber boom. Although adoption is still in its early stages, productivity is predicted to bring long-term growth, raising GDP by 5.8%. Out of the top 20 companies by market cap, over half of them are AI technology companies like Google, Meta, and Microsoft, or hardware manufacturers like Nvidia. According to Goldman Sachs, nearly half of all S&P 500 growth in earnings per share is coming from AI investment. As for non-publicly listed AI companies like OpenAI, privately valued at $852 billion, and Anthropic, privately valued at $965 billion, each with its own initial public offering (IPO) approaching, which may rocket their valuations to some of the largest IPOs ever seen. Anthropic’s valuation may be close to two trillion USD, beating out the previous record of SpaceX earlier this year. The impact of AI on both the U.S. economy and stock market is clear, and it follows that the U.S. government has an incentive to deregulate and invest in the industry to sustain this growth. All this investment, however, is not risk-free; Anthropic’s own IPO prospectus revealed that the company is operating at a worsening operating loss, with billions in deficit doubled from 2024 to 2025. Other major companies have taken on huge amounts of hidden debt in order to maintain their AI investment. The riskiness of AI investment demonstrates that companies and the government are making large bets on the future of the industry under the belief that, as AI adoption grows, their revenues will increase as well. If the industry slows or stalls, companies risk defaulting on their loans, potentially sending the economy and stock market spiraling. To mitigate risk, the U.S. government is supporting AI by promoting AI abroad and deregulating the industry.

The U.S. is not the only nation with advanced AI models, however. China’s AI capabilities have grown significantly in recent years, closing the technology gap between them and the U.S. to about 8 months of development time. Part of its ability to close this gap lies in the distillation of U.S. models; according to the National Coordinator for Critical Infrastructure Security and Resilience, China is training its most advanced models on less advanced U.S. ones to speed up development. Distillation won't fully close the gap, but it gives Chinese companies a great starting point, which has allowed some of their models to match or even surpass their U.S. counterparts temporarily. The Chinese Ministry of Commerce refuted the U.S. allegations, saying distillation is a normal part of model development and is practiced by many companies. China also stated that if the U.S. uses distillation as a pretext to suppress Chinese AI companies, it will enact countermeasures. 

Recently, Anthropic released a report detailing and documenting uses of its models in aiding in various plots dealing with conventional and biological weapons, hacking, influence campaigns, and surveillance at the request of malicious actors often based in adversarial nations. Actors used a U.S. developed and hosted model that contained safeguards designed to prevent such actions.  The threat of AI also comes from rogue AI agents, which are increasing in frequency. OpenAI’s agents recently attempted to breach the Australian government’s Medicare program and were partly successful, though they did not gain access to any patient information. Another notable instance involved OpenAI’s agents going rogue, probing and attempting to hack U.S. government websites, specifically the Department of Education, Census Bureau, and the Securities and Exchange Commission. These examples show that the threats posed by advanced U.S. AI models, whether they’re acting on their own or on behalf of another organization, and the malicious capabilities of foreign actors will only increase if they have their own advanced models without any safeguards in place. If China were able to develop its own advanced model without safeguards, it could use it to scale up the operations laid out in the Anthropic document, threatening U.S. security. 

What Has The Government Done to Promote or Restrict AI Abroad

The current U.S. administration has taken a policy of promoting U.S. developed AI models abroad, laid out in the third pillar of a July 2025 action plan. The administration plans to export American AI models to allies and partners abroad and pursues this by setting up a program through the Department of Commerce to promote and streamline the financing and authorization of full-stack AI exports, complying with Executive Order 14320. Full-stack is a bundle of AI resources including the hardware needed to run a model, the model itself, user interfaces, and any other required infrastructure. The program aims to support trusted nations' AI independence, allowing them to maintain sovereignty over their own infrastructure and models while simultaneously supporting America’s exports and AI industry. Promoting exports is the primary way the U.S. government has aimed to grow the AI industry. The rest of the plan lays out ideas on how to control AI and protect the U.S. from the risks that come with AI. The administration has already approved chip sales to Saudi Arabia. Congress has introduced legislation with the goal of promoting AI exports in competition with China. Namely, the Full AI Stack Export Promotion Act was introduced, which orders the Department of Commerce and the Department of State to remove barriers for full-stack exports as well as codifying the actions of Executive Order 14320 into law. A 2019 congressional reauthorization of the Export-Import Bank of the United States (EXIM) ordered the EXIM to establish a “Program on China and Transformational Exports” in order to help American exporters compete with government-backed and financed Chinese companies, including AI. 

The action plan states that one of the goals of the U.S. is to counter Chinese influence in international organizations. Specifically, the U.S. State Department is campaigning to re-elect Doreen Bogdan-Martin as the secretary-general of the International Telecommunication Union (ITU). Her vision document highlights the need to promote public-private partnerships for emerging technologies like AI. The U.S. is also campaigning for re-election for another term on the ITU council in part to counter China, which also has a seat on the council. The United Nations has also recently been looking into how to regulate and control AI, in part by setting up a scientific panel to determine the impact of AI. The General Assembly established this panel by a vote of 117-2, with the U.S. and Paraguay as the only countries to vote no, while China and Russia, along with many U.S. allies, voted in favor. 

The action plan’s next goal is strengthening U.S. export control enforcement. Export controls are laws administered by the Bureau of Industry and Security (BIS) that create regulations to export certain products out of the country for the purpose of protecting national security. Nvidia chips fall into this category and have largely been banned for export to China with a few exceptions. In December 2025, the FBI uncovered a smuggling ring that was mislabeling controlled chips in order to avoid export control restrictions and sell them to China. The smugglers attempted to export $160 million worth of chips to China. In keeping with the strengthening of BIS, the action plan seeks to close export control loopholes. BIS closed a loophole that allowed Chinese-based chip fabricators to purchase U.S. equipment and materials without having to apply for a BIS license to export the equipment from U.S. territory. These actions are part of the government's broader push to create new controls to limit China’s access to components used to develop new industrial capacity, slowing China’s semiconductor manufacturing growth. The U.S. also seeks to lead a global push to align export controls across the world to avoid routing U.S. goods through countries with weaker controls. Currently, the U.S.'s trade influence is so strong that it is able to enforce its export laws outside of its own territory, but it has trouble doing so due to a lack of export control officers. To solve this, it has pressured other countries through trade deals to update or enact export control schemes to align with U.S. controls. An example of this is the October 2025 U.S.-Malaysia Reciprocal Trade Agreement, which requires Malaysia to align its export controls with the U.S.’s. The U.S. also enacted a novel export control on Anthropic’s Fable and Mythos models in June 2026, preventing them from being accessed by any foreign person, which caused Anthropic to take the models down for every user. BIS cited a potential universal jailbreak that could be used to bypass safety protocols. The controls were lifted shortly after due to cooperation between Anthropic and the White House, but also partly because there was a concern that a model isn’t considered an export, which cast doubt on the legality of the controls. Congress has also played a role in regulating and limiting Chinese development. The Comprehensive Outbound Investment National Security Act (COINS) expands a prior outbound Investment and Security Program ordered by the White House by limiting U.S. persons from investing in industries related to military development, such as advanced computing and manufacturing in “nations of concern,” including China. 

What Does the Future of AI Policy Look Like?

Peter Harrell is a visiting professor at the Georgetown Institute of International Economic Law and former Senior Director for International Economics for the White House and a member of the National Security Council and the National Economic Council from 2021 through 2022. He laid out a series of options that the government has to counter Chinese AI companies. First, the government could cast doubt on the security and efficacy of Chinese models to discourage their use in the U.S. in order to protect domestic AI companies. A step up from this, he states that the Pentagon could ban Chinese AI from being used on government contracts, which would close the entire government contracting industry off from China. Banning Chinese AI is similar to the Pentagon’s designation of Anthropic as a supply chain risk. Next, he identifies that the government could place Chinese AI labs on the U.S. entity list, which would establish significant barriers and authorizations for those companies to buy U.S. products. Notably, it is not an import ban into the U.S., but it does make it difficult for these companies to do business in the U.S. The U.S. has already placed some Chinese AI companies on the entity list but held off on placing major companies like Deepseek and over 100 others on the list. So far, none of these options have banned the import of models, but Harrell lays out a few ways that the White House could ban Chinese models using the International Emergency Economic Powers Act (IEEPA). First would be using the Information and Communications Technology and Services rule, which is grounded in the IEEPA, to ban the import of software and technology that poses a national security risk. Next would be using IEEPA to place sanctions on Chinese AI firms and ban all U.S. transactions with them or restrict Chinese AI without sanctions. It should be noted that IEEPA requires a state of emergency, and Harrell raises a legal question as to whether IEEPA can control open-weight or free models from China without a transactional element. Harrell then states that congressional legislation could be passed to regulate AI, chip exports, and chip imports. 

There are multiple bills floating in Congress currently that will potentially regulate AI and Chip exports and imports. The proposed AI Overwatch Act effectively bans chip exports to China by requiring congressional authorizations and oversight when BIS issues any export licenses for chips to “nations of concern,” including China. It strips the White House of its authority to trade and authorize foreign chip sales. The Multilateral Alignment of Technology Controls on Hardware (MATCH) Act aims to strengthen the export control schemes of our allies to prevent Chinese chip manufacturers from procuring materials from nations with weaker controls than the U.S., creating stronger multilateral export controls. The Chip Security Act would target smuggled chips sold on the secondary market, which China has used to train its models, by requiring location verification, safeguards, and requiring chipmakers to report potential diversions. The Semiconductor Technology Resilience, Integrity, and Defense Enhancement (STRIDE) Act directs the Department of State to lead a multilateral effort to coordinate export controls and protect the global technology and equipment supply chain. All of these bills have broad bipartisan support. 

Conclusion

The primary goal of the U.S. government is figuring out how to protect against the national security and economic threats posed by cheap and effective Chinese AI without overregulating and smothering its own domestic chip manufacturing and AI industries. The U.S. government is attempting to cut China off from resources and trading partners through multilateral efforts aimed at slowing its AI development to maintain the technology gap and protect U.S. companies from foreign competition. In order to protect against the cybersecurity threat of AI, the U.S. must strengthen its cybersecurity systems by incorporating AI into its own cybersecurity development in order to find and fix any vulnerabilities before an adversarial agent does. A caveat to this is that the U.S. must maintain superiority over China in terms of model capabilities because, if not, China's models will be able to exploit vulnerabilities in U.S. models. In order to stay ahead, the U.S. has promoted AI development and supported AI integration into cyberdefense systems. The risk of this strategy is that models can grow so advanced they escape human control, now with the training to bypass any security system. It is clear that the U.S. executive branch believes that the economic rewards and the national security strengthening outweigh the financial crisis risk and rogue AI risk, as it has committed itself to a policy of deregulation and promoting integration into government systems in order to stay ahead of China. Many AI industry executives and other political leaders disagree and believe that the U.S. has a responsibility to slow down development and limit the strength of AI models. When it comes to its allies, Washington is pushing for more cooperation and trade in order to close export loopholes and increase its global leverage over international organizations. While the U.S. has already taken many steps to achieve its goals, both the White House and Congress have numerous options available to them that they can enact in the future, the arms race does not seem to be ending any time soon.