The AI Race Isn’t Real

As one of his first acts in office, President Trump rescinded President Biden’s executive order imposing basic safety reporting requirements on artificial intelligence (AI) companies. A few days later, Trump issued an executive order directing federal officials to change existing policies that might inhibit AI development. Six months later, the White House released its AI Action Plan, titled “Winning the Race,” that called for removing barriers to innovation, cutting “red tape,” and revising federal safety frameworks to eliminate references to climate change, misinformation, and diversity. 

The notion of an AI race against China is at the center of the administration’s deregulatory turn. The AI Action Plan begins by stating: “The United States is in a race to achieve global dominance in [AI]. Whoever has the largest AI ecosystem will set global AI standards and reap broad economic and military benefits.” It sets “dominance in this global race” as its primary goal. Vice President Vance has raised concerns that, if America pauses AI development to focus on safety issues, we could “find ourselves all enslaved to P.R.C.-mediated A.I.” Secretary of Defense Pete Hegseth recently ordered the integration of AI into a variety of military systems, asserting that “[m]ilitary AI is going to be a race … and therefore speed wins” and stating: “We must accept that the risks of not moving fast enough outweigh the risks of imperfect alignment [between humans and AI].”

In Congress and the private sector, an AI race against China is also frequently invoked as a reason to favor light regulation of domestic AI companies, even in the presence of serious threat models. Announcing a Senate hearing on AI in May 2025, Sen. Ted Cruz (R-Texas) emphasized that “the way to beat China in the AI race is to outrace them in innovation, not saddle AI developers with European style regulations.” In a rare instance of bipartisanship, Senate Minority Leader Chuck Schumer (D-N.Y.) has warned that “if America falls behind China on AI, we will fall behind everywhere: economically, militarily, scientifically, educationally.” The leader of a tech industry trade organization argued that a proposed 10-year moratorium on state regulation of AI was “important in the race against China.” A tech lobbyist made similar arguments, asserting that “[w]e’re in a race with China to lead the future of artificial intelligence. A fractured, state-by-state system will only slow us down.” Ads in favor of the moratorium argued that “we can’t let China get the upper hand.”

In a forthcoming article, we critique the idea of an AI race that must be won at all costs. We think it is fundamentally flawed as a descriptive matter and normatively defective.

Start with the descriptive problems. The race metaphor implies a finish line, but AI competition has none. Military AI development is an ongoing accumulation of dangerous capabilities, with no fixed endpoint. The metaphor also gets the competitive dynamics backward. AI knowledge is leaky. It is rapidly copied, distilled, and reverse-engineered, and as a consequence, racing ahead often accelerates competitors rather than leaving them behind. And in the economic domain, the case for racing is weaker still. AI products show few of the network effects that would let a first mover maintain market dominance.

Normatively, the race framework corrodes safety by rewarding speed over caution, incentivizing risk-taking and shortcuts. Moreover, a race to build AI weapons risks destabilizing the current era of great-power détente and nuclear deterrence, potentially leading to more—and more deadly—conflicts between large nations. And it sharply cuts the United States’ ability to come up with solutions, by treating every policy choice as a zero-sum sprint. American AI superiority can be beneficial on the margin, and there is much policymakers can do to protect and advance it. Yet that margin does not justify abandoning the cost-benefit regulatory analysis that every other technology receives. The race framing treats these marginal benefits as categorical imperatives. In the article, we propose an alternative concept of American AI leadership: not AI defined by brute capability alone, but AI that is genuinely more useful, capable, reliable, and safe.

The Descriptive Failure of the AI Arms Race Concept 

As described above, the race metaphor invokes, for most readers, some implicit assumption of a finish line. America’s AI Action Plan evokes this explicitly: “Just like we won the space race, it is imperative that the United States and its allies win this race.”

This metaphor may have been apt for the space race, which had a clear flag-on-the-Moon moment, but in AI, no such endpoint exists. Policymakers adopting the race metaphor seem to assume that either the United States or China will build a powerful AI system and wield enormous long-term influence as a result. But this is not how superpower arms races have historically functioned. 

Scientific research into weapons and related military technologies can confer a temporary battlefield advantage. But even before the digital era, weapons developed in one nation spread rapidly to others with the means to build them. The interconnected nature of the modern technology economy, coupled with the relative ease of copying or reverse-engineering digital technologies, indicates that AI-based capabilities will spread even more quickly than advanced weapons of the past. The competition between the United States and China will be just that: an ongoing competition with no clear end state, no finish line, and only minimal rewards for temporarily gaining the lead. The only way to “win” such a military race in any lasting sense would be to use the resulting advantage to permanently disable the competitor: to conquer it, destroy its ability to ever engage in AI research (which would require widespread destruction and continual monitoring), or credibly threaten annihilation sufficient to force permanent subordination. 

Needless to say, preemptively unleashing a superweapon against a peaceful competitor would be a geopolitical catastrophe and a moral abomination. It would violate international norms, shatter alliances, and likely provoke escalation, including nuclear escalation. But for now, our point is not normative but descriptive: Fighting a war of domination is a far cry from the natural meaning of “winning the race.” If policymakers are steering Americans toward that goal, intellectual honesty and democratic accountability demand that it be part of the public debate, not hidden behind euphemisms about “the end state” or how “speed wins.” Perhaps they have not thought so far. But they should, because short of preemptive conquest of a peaceful competitor, there is no finish line in an AI arms race.

Nor is the lack of a finish line the only reason why the race metaphor fails. Knowledge diffusion flips the very idea of a race on its head. Knowledge is leaky. Once discovered, it is extraordinarily difficult to prevent others from acquiring and using it. This has already played out a number of times in AI development, where acceleration by the leader often hastens catch-up by competitors rather than leaving them behind. Thus, rather than a race where the speeder accelerating widens a gap, here, it may well pull in competitors.

The pattern is visible throughout recent AI development. Voluntary publication is one channel: Google developed the transformer architecture, published “Attention Is All You Need” in 2017, and watched a small lab, OpenAI, use that architecture to overtake its lead and hold the frontier in generative AI. Open-source release is another. DeepSeek has built its distilled R1 variants on Meta’s Llama, while Meta reportedly assembled engineering war rooms to reverse-engineer DeepSeek’s techniques after R1’s release. Diffusion also operates in the shadows. Distillation, where one model is trained on the outputs of another, has proved hard to contain. DeepSeek’s V3 model repeatedly identifies itself as ChatGPT and reproduces GPT-4’s jokes. Anthropic has documented what it calls industrial-scale distillation campaigns by DeepSeek, Moonshot AI, and MiniMax, involving more than 16 million exchanges through roughly 24,000 fraudulent accounts.

But the most significant example is the release of OpenAI’s o1—its first reasoning model, breaking away with the older approach of generating tokens in a single forward pass. The literature had long speculated that test-time inference would unlock new capabilities, yet OpenAI’s September 2024 release still sent shockwaves. Even though OpenAI was tight-lipped about its methods, the release itself resolved the central question of whether the paradigm could scale. Seeing that it did, competitors redirected efforts toward a confirmed target. Four months later, DeepSeek released R1, exhibiting mastery of the new paradigm.

Not all knowledge is equally leaky, and the strongest counterexamples come from extreme ultraviolet (EUV) lithography and stealth aircraft. There, leads have been sustained for decades. But there are important domain differences. For these industries, advantages rest on industrial capacity, enormous capital investment, and tacit manufacturing knowledge. AI advances diffuse through papers, application programming interfaces (APIs), and researcher poaching. The consequences of this distinction can be counterintuitive: Before o1, much of the American lead was compute-bound, resting on chip and data access and the capital required to train at frontier scale. Once o1 revealed an algorithmic breakthrough, it did not just narrow the gap; it shifted the contest onto algorithmic terrain, where Chinese labs could innovate rather than follow.

These are the forms of diffusion that occur in the open. A true race between the United States and China would likely involve espionage. AI models, sitting in networked data centers owned by private companies, present softer targets than nuclear secrets stored in classified government facilities. Cybersecurity measures at American AI companies are often inadequate to prevent intrusion by sophisticated state-sponsored actors, especially given race pressures to speed ahead. There has already been at least one known breach of a major AI company’s systems and theft of proprietary data, and other breaches may have gone undetected. The faster America develops advanced AI, the sooner China will develop comparable systems—and vice versa.

The race metaphor is even less persuasive in the context of describing economic competition. The notion that “winning” the AI race will confer durable economic dominance misunderstands both the structure of technology markets and the specific characteristics of AI as a product.Various industry commentators, including Andrew Ng, have noted that “models are commoditizing the foundation-model layer.” A senior Google engineer drafted an internal memo titled “We Have no Moat, and Neither Does OpenAI.” The claim is that users increasingly treat models as interchangeable, switching between them based on cost and convenience rather than brand loyalty. Peter Salib and Simon Goldstein further argued recently that AI simply does not lend itself to becoming a natural monopoly. To be sure, there are strong arguments on both sides, and it is hard to know whether sustainable moats would emerge through differentiation on latency, cost, capability, or autonomy. However, in the present reality, where the labs reshuffle their positions on the leaderboards on a monthly basis, the potential commodification of AI should serve as caution against over-investment in an expropriable benefit.

One potential case for a durable lead rests on recursive self-improvement (RSI): An AI system advanced enough to improve itself would compound advantages quickly. But even this case requires that growth will be exponential for a durable advantage to emerge, and there are real reasons to doubt that this will be the case. Diminishing returns are the norm in technological development, and scaling laws show log-linear, not exponential, returns. The ultimate bottlenecks will likely lie in hardware, industrial capacity, and, as Sam Altman puts it, “electrons.” Here, competitors may have comparative advantages over American labs. In 2024, China added more generation capacity than the entire U.S. grid, not to mention its lead in solar panels production. Thus, unless one assumes that RSI on its own will have exponential returns, the competitors will soon catch up, the moment their systems enter the RSI cycle.

A final consideration is that users do not universally flock to the “smartest” system. Users also care about cost, latency, style, autonomy, and reliability. This gives Chinese labs a significant opening, precisely because they are pressured to innovate on compute efficiency. Rather than a single race that can be won, AI development is likely to be an ongoing competition along various dimensions, with no single settled winner. 

In short, an American lead in AI is worth having, but it is not qualitatively different from a lead in any other globally competitive industry. Policymakers may correctly see the path to a sustainable lead as going through chip design, advanced lithography, hyperscaler data centers, and power generation: that is, an industrial policy that emphasizes energy and an industrial base. What they should not do is treat the pursuit of that lead as a singular ideal that forecloses normal regulatory analysis. Every advanced economy balances the benefits of technological leadership against the costs of unregulated markets, whether for prescription drugs, coal-fired power plants, toxic chemicals, or cars. The race framing short-circuits that analysis by treating deregulation as necessary rather than as one policy option among several. A more appealing alternative is one where the United States positions itself as providing the highest-capability, safest, and most reliable AI, capturing the returns to leadership without sacrificing the regulatory tools that allow a modern economy to function.

The Dangers of the AI Race Concept

Even if the concept of an AI race were coherent, treating AI development as a winner-take-all sprint could be disastrous. A race posture toward AI development carries substantial strategic risks. For one, race dynamics can destabilize existing deterrence structures and create conditions for catastrophic miscalculation. If one nation appeared to be on the verge of developing a superweapon that would allow it to neutralize its rival’s defenses or subjugate its population, the threatened party might conclude that a preemptive strike using existing weapons, perhaps including nuclear weapons, offered better odds than waiting. Second, even if one nation develops what it believes to be decisive AI-enabled military capabilities, it is far from clear that even radically advanced AI systems could neutralize an adversary’s nuclear deterrent. Like the U.S. and Russia, many of China’s nuclear systems are not networked nor easily disrupted by cyber operations or AI-directed attacks. The development of advanced AI military applications also threatens to spread dangerous capabilities to rogue states and non-state actors, much as nuclear weapons technology eventually diffused to North Korea and Pakistan. Unintentional harms to civilian populations from autonomous AI weapons also become more likely as such systems proliferate. 

Another key problem with the race frame is that it may create the very dynamic it claims to describe. The race metaphor may operate as a “hyperstition”: a belief that makes itself true through the actions it inspires. By treating AI development as a race, policymakers do not merely respond to a competitive threat—they might unwittingly construct one. Each nation’s accelerated investment signals aggressive intent to the other, triggering the dynamic that international relations scholars call the “security dilemma”: Defensive measures taken by one state are perceived as offensive threats by another, prompting escalation that leaves both parties less secure. The race metaphor does not merely describe a security dilemma; in effect, it manufactures one. And once the dilemma takes hold, it becomes self-sustaining, as each round of investment provides fresh evidence that the competition is real and urgent.

But by far our greatest concern with the AI race as a working policy blueprint is how corrosive it is to safety. It promises better AI, but it produces more brittle systems. Races reward speed, and speed is incompatible with caution. In the absence of an industrial policy based on a race perception, labs would find more careful balance in frontier AI development between safety investments and the economic and strategic benefits that advanced AI would provide. Normal competitive pressures would continue to hold, development would not stall, but policy would create the conditions necessary for developers to take the time to test systems and implement safeguards, if only to avoid liability. Attempting to win an arms race disrupts this equilibrium, either by creating real pressure or by providing a convenient alibi to labs that deprioritize safety. 

The race dynamic can also lead to dangerous overdelegation of authority from humans to AI agents. Principal-agent problems are well understood in law. Delegation involves various risks, and principals often choose a mix of delegation and oversight, to ensure the benefits of delegation while minimizing its costs. In a race dynamic, especially one with military implications, human decision-makers will most likely be pressured to overdelegate to AI agents to exploit their superior speed and processing capacity, at the expense of effective oversight. A preview of this is already occurring in the Ukraine-Russia war, where both nations deploy drones capable of identifying and destroying targets without explicit human permission, prioritizing effectiveness over human judgment. 

We can’t conclude this discussion without addressing the elephant in the room: superintelligence. Many experts believe that AI may well reach levels of capability that far exceed those of most humans. The strategic advantages of superintelligence are potentially enormous. A nation that created an entity far more capable than any human might gain access to insights, solutions, and capabilities beyond human ability to conceive or counter. 

Yet deploying a superintelligent system before we have learned to reliably align such systems with human goals and interests would be extraordinarily reckless, comparable to stockpiling nuclear weapons before developing safety protocols to prevent accidental detonation. Known and unknown risks loom large. Further, despite some progress on current-level large language models, AI alignment has thus far proved brittle and incomplete. Racing to build systems that would, by definition, exceed human abilities in virtually every area, before solving alignment problems we cannot yet solve at far lower capability levels, risks catastrophic outcomes. Building superintelligent systems before we understand how to make them reliably safe is not competition—it is collective suicide with extra steps, a race toward a brick wall.

In short, the benefits of “winning” the AI race are far smaller and more temporary than those adopting the race metaphor appreciate. And the costs of racing are potentially massive, ranging from devastating attacks by rogue organizations, to an elevated risk of global war, to human extinction itself. The right response is not to run faster, but to recognize that the race concept itself is the problem.

Strip away the race metaphor, and options that were obfuscated come into clear view. Cooperation on safety standards, measured pacing of frontier development, catalytic regulation, international agreements on military AI: None of these is conceivable if the only question is who crosses the finish line first. But there is no finish line. Once that becomes clear, the full range of policy tools becomes available again. In our forthcoming article, we develop this alternative vision: American AI leadership defined by effectiveness and safety.