Middle powers are racing to procure and integrate artificial intelligence (AI) into their military operations. Almost all of the nations that are commonly considered middle powers have publicly announced deals to acquire or develop military AI. So far, most of those deals are for narrow applications, such as computer vision for drones and targeting. But as Anthropic’s decision to withhold its Mythos model over its offensive cyber capabilities shows, the military implications of frontier general-purpose models built by top American and Chinese labs are becoming unavoidable.
Where middle-power militaries source their AI is a matter of bipartisan concern in Washington. The White House’s “AI Action Plan” describes exporting American models to allies and partners as “imperative,” warning that failure would push other nations to “turn to our rivals,” namely China. The Trump administration’s Pax Silica initiative and AI exports program work toward these ends. Democrats such as former National Security Advisor Jake Sullivan have made similar cases.
But other nations are increasingly wary of overreliance on the United States—concerns that may heighten as the Trump administration considers taking stakes in the top AI labs. Convincing them to adopt U.S. AI poses challenges unlike any past arms export. Whether American technology underpins the next generation of military power—or whether its influence fragments—depends on navigating these risks.
Broadly, nations face three choices about which sort of AI to bet their military future on and how to access it. These decisions will be shaped, in part, by disparate national views on what defines AI, how transformative it is as a technology, and how imminent that transformation is.
The first bet is that frontier general-purpose AI, including popular large language models trained on massive datasets and designed to handle a wide variety of tasks, will dictate future military power, empowering everything from cyber operations and command and control to drone coordination and logistics.
The most direct route to obtaining that tech is to contract with the top frontier labs—OpenAI, Anthropic, or Google DeepMind—all of which are American-owned. Partnerships like this are already taking shape. Britain signed a national security agreement with DeepMind, OpenAI’s deal with Australia named defense as a target sector, and Anthropic’s role in the Maven Smart System may extend into NATO operations. But these models are closed-weight, meaning that the developer controls the parameters that determine model behavior. This inherently ties a military’s AI access to its relationship with labs—and the labs’ relationship to Washington.
A second, more modest approach is trusting that smaller-scale general-purpose AI will be sufficient for military operations. Some countries will try to build their own models. France is deploying Mistral, a Paris-based start-up, across its military, Canada is doing the same with Cohere, and India has committed billions of dollars toward AI sovereignty. Others could partner with an open model developer such as Meta in the United States, China’s DeepSeek and Alibaba, or Mistral. Open-weight models are publicly available for download, but military use is typically prohibited by acceptable use policies.
An official partnership could grant militaries legal, unrestricted use along with the operational and technical support that most nations lack. Meta has taken an aggressive step by extending access to its large language models to select allied militaries, though this offer comes without any deployment assistance. But today, China dominates the open model market in terms of popularity and performance, and France has pitched Mistral, in Europe and beyond, as a path to military AI that reduces dependence on the United States and China.
The third approach is the most limited: focus on tactical applications of narrow AI, such as computer vision for targeting, drone autonomy, and forward-deployed command-and-control nodes that operate in hostile environments.
These three approaches aren’t exhaustive. Nations can partner with specialized U.S. defense firms such as Palantir and Anduril, which could bundle frontier AI access with engineering, infrastructure, and operational support. Ukraine and Israel have started to do so, but allies in Europe are more hesitant and may opt for local competitors such as Germany’s Helsing. Sanctioned states and those most wary about overreliance can also illicitly use models outside their terms of service or distill the best AI—a process of querying commercial models millions of times to train replicas at “a fraction of the cost,” per a White House memo. Enforcement mechanisms against illicit use are essentially nonexistent.
They also aren’t mutually exclusive approaches. A country might partner with a U.S. lab for near-term capability while experimenting with open models or developing its own as longer-term hedges.
But experts on the diffusion of emerging technologies argue that military advantage lies less in access to technology than in the capacity to exploit it. Recent AI warfare supports these theories, as specialized AI models that are narrower in scope than those produced by top labs are already reshaping military operations.
In Ukraine, Russian drones face off against semi-autonomous Ukrainian interceptors while unmanned ground vehicles hold front-line terrain. Israel has used AI to conduct mass surveillance across Gaza, track individuals to their homes, and generate targeting recommendations at a scale that would be impossible without significant automation. These Israeli efforts, however, have reportedly led to mistaken arrests, wrongful interrogations, and civilian deaths, raising questions about the reliability of these technologies in life-and-death contexts and who is accountable when they fail.
The viability of any military AI strategy depends on how that nation weighs its military objectives against the domain-specific strengths of frontier and nonfrontier models. Highly compute-intensive frontier AI such as Claude Mythos, which requires tens of thousands of cutting-edge chips and hundreds of millions of dollars to train, may generate significant advantages for a nation such as the United States with sweeping military ambitions. But for those with limited objectives—defending one’s borders, say—narrower models with lower compute requirements may suffice.
Specialized models often outperform larger systems on the narrower tasks that dominate today’s battlefield AI, such as object detection. Even for broader language-based applications, one defense start-up, EdgeRunner, claims that its model—a fine-tuned version of an open-weight OpenAI model—runs on a single consumer graphics card while matching OpenAI’s flagship GPT-5 on military tasks. If a middle power can approach frontier performance by using distilled and fine-tuned versions of larger models, the case for costly models—and their inherent entanglements with the U.S. government—weakens considerably.
Still, general-purpose AI models with capabilities far beyond today’s applications could eventually become indispensable for militaries. A nation with privileged access to frontier coding and research assistants can iterate on drone autonomy, computer vision, and targeting systems faster than one without. The United States leads on every frontier AI metric that matters: model performance, investment, compute capacity, and talent. Its commanding position in the AI race raises the cost of every alternative pathway.
But it remains unclear whether the top closed-weight models will outperform what’s openly available enough to confer a battlefield advantage over lesser models that are adopted faster. Epoch AI estimates that since January, the capabilities of open models have trailed frontier proprietary models by roughly four months. If that modest gap holds, then battlefield advantages may turn less on temporal gaps in technical capabilities than on how quickly capabilities that are merely good enough reach the fight. And open models, even if they are American, leave suppliers with no meaningful leverage once their weights are downloaded—a clear draw for middle powers with sovereignty concerns.
The second major barrier to U.S. AI hegemony is the labs themselves, which were founded on ideological missions with no precedent in the defense industry. Claude’s constitution expresses concern about “unprecedented degrees of military… superiority.” DeepMind’s founders reportedly made a no-military-use commitment a condition of its 2014 sale to Google. In 2024, meanwhile, OpenAI quietly removed its “military and warfare” ban as it began work with the U.S. Defense Department.
These commitments shape the labs’ corporate governance, influence who works there, and steer their business decisions. Google withdrew from Project Maven in 2018 after thousands of employees signed a protest letter, forcing the company to publish AI principles that prohibit weapons applications. When Anthropic was cast aside by the Pentagon over its objections to “fully autonomous weapons” and domestic surveillance, OpenAI immediately stepped in but suffered several public resignations shortly thereafter. Google’s recent deal with the Pentagon and its relationship with the Israel Defense Forces have led to employee resistance.
State-directed U.S. export promotion thus cannot treat labs as neutral suppliers. The administration can push U.S. AI to allies only to the extent that the labs agree on supporting those partners, on what terms, and with what restrictions.
The United States’ position as the world’s preeminent arms exporter is in jeopardy. The wars in Ukraine and the Middle East are broadcasting the limitations of the U.S. defense base’s decades-long focus on costly, exquisite systems. Facing trade disputes, arms delivery delays, and shifts toward cheaper unmanned systems, an increasing number of nations—including Germany, India, Taiwan, Japan, South Korea, and others—are pushing to build up sovereign capacity away from the United States.
As AI grows increasingly central to military power, partners that are wary of overreliance on the United States will need concrete reassurance that they retain operational freedom and persistent access to frontier capability. Instead of treating foreign rules about where sensitive data can be stored and who can access it as obstacles to U.S. firms, Washington should respect sovereignty concerns and allow allies to keep military data under their own control.
The most durable reassurance is interdependence: partnerships where the United States controls key parts of the AI stack while allies contribute indispensable pieces of the supply chain. Washington should lean into its advantages in coding and cyber rather than try to rapidly reclaim the manufacturing sectors that it lost decades ago. Allied access to U.S. frontier models can boost the software that allies build themselves, such as Ukrainian drone autonomy models and battlefield management systems such as Delta; in return, the United States can rely on partners for what it can’t easily rebuild: drone production, chip fabrication, and precision manufacturing.
Collaborating with AI labs is equally imperative. Export promotion cannot succeed if the companies whose models underwrite it are unwilling partners. Coercion, such as invoking the Defense Production Act, will likely backfire, leading to resignations and reputational damage at top labs, as well as exacerbating middle powers’ concerns about overbearing U.S. influence.
That means building lasting frameworks in place of the ad hoc negotiations that produced the Anthropic-Pentagon standoff. The two sides must agree on which models can be deployed abroad, to whom, and with what auditing and usage restrictions. In exchange, the government can offer what labs cannot secure on their own: classified threat information, expedited security clearances, and a real say in how their tools are deployed.
The actual AI decisions that middle powers face are far more complicated than a choice between the United States and China. Successful export promotion will require a nuanced understanding of the options that each nation has, the factors that will shape their choices, and what independent labs have to say about it. The way that military AI proliferates will determine whether the United States can maintain its position as the anchor of allied defense and reclaim the uncontested military superiority and cohesion that once underpinned Pax Americana.
