The Unconstrained Race

On January 20, 2025, a Chinese AI lab called DeepSeek launched a model called R1 on the Apple and Android app stores. Within a day, the S&P 500 had slipped more than 1.2 percent and Nvidia had lost over $590 billion in market capitalization. R1 demonstrated performance matching OpenAI’s most advanced reasoning model at approximately five percent of the training cost. DeepSeek’s model is open source: anyone can download it, modify it, run it on their own infrastructure, and use it for any purpose.

DeepSeek operates on servers in China. China’s National Intelligence Law requires companies to “support, assist, and cooperate” with state intelligence agencies. Any user who inputs sensitive data into DeepSeek is, under Chinese law, providing that data to any state actor with a lawful request — and in China’s legal framework, every intelligence request is lawful by definition. Security researchers have demonstrated that DeepSeek can be jailbroken to generate malicious code, fabricate identities, and produce targeted social engineering campaigns at scale. Unlike U.S. AI firms, Chinese companies rarely disclose the findings of their safety evaluations.

This is not a description of a hypothetical threat. It is a description of the current state of the race.

The regulatory environment is not keeping pace. More than 1,500 AI-related bills have been introduced in U.S. state legislatures in 2026 alone. The legislative cycle runs in years. The model release cycle now runs in weeks. By the time a regulatory framework is written, debated, passed, and implemented, the technology it was designed to govern has been superseded by two generations of its successor.

What one nation’s legislature restricts, another nation’s lab treats as competitive advantage. The GDPR governs European data privacy. It does not govern Hangzhou. The EU AI Act regulates AI development within member states. It does not apply to the People’s Liberation Army. What Congress legislates binds American companies. It does not bind Baidu, Alibaba, Huawei, or the Shanghai Artificial Intelligence Laboratory.

Nobody voted for nuclear either. The Manhattan Project scientists — Oppenheimer, Fermi, Szilard — circulated a petition in 1945 arguing the bomb should be demonstrated rather than dropped on cities. The military did not consult them on the final decision. Several spent the rest of their lives trying to contain what they had built. The logic that drove the project was the same logic driving this race: the alternative is that the adversary builds it first, and the adversary building it first is worse than us building it first, so we proceed.

The logic is internally consistent and entirely unaccountable to any democratic process. It was true in 1945. It is true now. The difference is that a nuclear weapon requires specific materials and manufacturing that takes years and is visible from satellite. The equivalent threshold in AI requires a server cluster, a training dataset, and the willingness to run the code. The proliferation problem is orders of magnitude harder.