The Iran war offered the clearest demonstration yet of what artificial intelligence can do on a battlefield. In the first week of fighting alone, the US military struck more than 3,000 targets; by the time a ceasefire arrived 38 days later, that figure had climbed to roughly 13,000. Behind those numbers sat Palantir’s Maven Smart System and an extraordinary surge in computing demand — the Pentagon’s daily consumption of AI model tokens jumped by more than 4,400 percent, a figure one official described as reflecting an “insatiable appetite” for the technology. It is tempting to read this as proof that AI-powered warfare is about to flood the globe. The reality is more complicated, and considerably slower.
The headline-grabbing capability obscures a stubborn truth: building a functioning AI targeting system is brutally hard, expensive, and dependent on infrastructure that most countries simply do not possess. Right now, only a handful of states — China, Israel, North Korea, Russia, South Korea, Ukraine, the United Kingdom, and the United States — are actively deploying AI for wartime targeting. That short list is not an accident. It reflects the genuine barriers to entry that separate aspiration from operational capability.
The Israeli Blueprint Shows How Hard It Is
No country illustrates the difficulty better than Israel, which spent decades and billions of dollars assembling the pieces. The foundation was data infrastructure built up over years. By 2020, Israel was deploying integrated surveillance across the West Bank and East Jerusalem, including a database called Wolf Pack containing detailed records on Palestinians, facial recognition checkpoints under the Red Wolf program, and a smartphone system called Blue Wolf that matched photographed faces against that database in real time. None of this happened quickly or cheaply.
The compute and storage requirements proved equally demanding. Israel’s Project Nimbus — a $1.2 billion cloud contract signed with Google and Amazon in April 2021 — gave the military access to a full suite of machine-learning tools, including automated image classification, object tracking, and sentiment analysis. When domestic servers ran short of capacity, Unit 8200 signed a separate agreement with Microsoft to host data from millions of intercepted phone calls on Azure infrastructure located in the Netherlands and Ireland. The scale is staggering: by 2024, the IDF was using Microsoft and OpenAI infrastructure roughly 200 times more than the year before, with stored data exceeding 13 petabytes — the rough equivalent of 14 billion printed books.
Only after all of that groundwork could Israel build the software layer that compresses the kill chain. Reporting from +972 Magazine and The Guardian described the Gospel and Lavender targeting platforms as largely built on artificial intelligence, generating targets almost automatically at unprecedented scale — Gospel focused on physical infrastructure, Lavender producing kill lists of individual suspects. Finally, Israel needed precision weapons to act on those recommendations, supplied by a domestic drone and missile industry — Elbit Systems, Israel Aerospace Industries, and Rafael — that took decades to mature. According to SIPRI, Israel’s share of the global arms export market reached 4.4 percent between 2021 and 2025, making it the world’s seventh-largest supplier.
Why the Technology Resists Easy Copying
The hardest part of military AI is not the algorithm — it is the data quality and the continuous intelligence feed that the algorithm depends on. Training a model to reliably identify targets requires enormous volumes of precisely labeled material; a single algorithm may need as many as 10,000 carefully labeled images to function usefully. And models trained in one environment degrade sharply in another. The US military found that systems achieving 70 percent success rates in Afghanistan dropped to 30 percent in the Philippines, where people walked in front of dense green jungle rather than dusty terrain. In Ukraine, targeting algorithms were repeatedly fooled by snow, sleet, and even Russian tanks with damaged turrets.
This brittleness is the crucial difference between AI targeting and the other technology transforming modern war — cheap drones. Drones are proliferating rapidly across advanced militaries, smaller states, and even non-state armed groups, precisely because they are simple, modular, and forgiving of imperfect conditions. Ukraine recorded 819,737 successful drone strikes against Russian forces in 2025 alone, with AI-enhanced first-person-view systems pushing strike accuracy from 30 to 50 percent up to around 80 percent. But a drone with autonomous terminal guidance is a fundamentally different thing from a comprehensive targeting complex that fuses signals intelligence, satellite imagery, and battle management software into a single kill chain. The first spreads easily. The second does not.
The Rising Powers to Watch
That said, a cohort of well-resourced states has both the motivation and the means to eventually field serious military AI — Brazil, India, Pakistan, Saudi Arabia, Singapore, Turkey, and especially the United Arab Emirates. The UAE is the clearest case. Its armed forces, which former US Defense Secretary James Mattis once dubbed “Little Sparta,” have invested heavily in surveillance drones and ISR capacity. Domestically, the Oyoon program integrates facial recognition and CCTV across the country, built with significant Chinese support.
The infrastructure piece is falling into place rapidly. The UAE’s national AI champion, G42, secured a $1.5 billion investment from Microsoft in 2024, and during Trump’s 2025 visit the two countries announced a massive Abu Dhabi data center complex featuring 2.5 million Nvidia chips and five gigawatts of capacity — one of the largest such deals anywhere in the world. On the weapons side, the EDGE Group, formed in 2019 from 25 Emirati companies, has already become one of the world’s top three manufacturers of precision-guided munitions. Palantir launched its first joint venture in the Emirates in 2025, and it is not a stretch to expect Emirati forces to adopt a Maven-style targeting platform within a few years.
The policy implication is uncomfortable but clear. Because AI is a general-purpose technology useful for civilian and military applications alike, stopping its spread outright is not feasible — and with 2025 ranking as the third most violent year of the post-Cold War era, demand will only grow. What regulators can realistically do is raise barriers around the most dangerous high-end capabilities: restricting exports of cutting-edge AI chips, limiting access to the most advanced models, and controlling the spread of military-specific software like Maven itself. Beyond that, normative pressure matters. When figures like the Pope warn that AI systems risk making war more “feasible,” it forces a public reckoning with a subject most people still treat as abstract. Even an imperfect, partial AI capability — Hezbollah has already deployed AI attack drones against Israeli forces with minimal human control — can meaningfully shift the battlefield. The killing machines are coming. They are just not arriving on the timeline the headlines suggest.
Original analysis inspired by Steven Feldstein from the Bulletin of the Atomic Scientists. Additional research and verification conducted through multiple sources.
