The New AI War: Five Lessons for NATO Doctrine

Flavien Chervet, July 19, 2025

On July 18, 2025, a historic military "mega-deal" is under consideration between Kyiv and Washington. The deal would have America commit to buying from Ukraine drones designed for the country's battlefield. In exchange, Kyiv would also buy weapons from the United States and, above all, share the experience gained over three years of war against Russia.

Why are the United States so eager for this feedback? To understand it, we have to go back to the start of the war.

In early March 2022, a sixty-four-kilometer Russian column advances toward Kyiv. Satellite images show it as a "caterpillar of death," confident in its mechanical power… until some thirty volunteers from the Aerorozvidka unit, armed with makeshift drones, destroy three vehicles at the head of the convoy; the road is blocked, the offensive bogs down, and the caterpillar of death eventually turns around and heads back to Belarus. Thirty geeks, a few kilos of plastic explosive, and some software: information has just proved that it outweighs sheet metal.

That night tips the war: victory becomes a matter of algorithm rather than tonnage. The following months confirm the shift. Palantir installs its AI-packed system, Gotham, in Kyiv; the tool fuses satellite data, drone footage, and field reports to generate firing options and collect evidence of war crimes. In its wake, Microsoft, Google, Starlink, and Clearview turn Ukraine into an "AI War Lab."

To orchestrate this influx, Kyiv creates Brave1, a platform that now grants the "Test in Ukraine" label to prototypes validated in combat. On July 17, 2025, the government even opens the front to foreign suppliers: drones, countermeasures, and lasers can be tested under real conditions, with near-immediate feedback. The battlefield becomes an incubator.

Miniaturization follows. The Switchblade 300 (three kilos, a grenade as payload, GPS guidance) costs less than a tank shell and deploys from a portable tube. On the ground, the fifteen-kilo BAD-2 robot dogs carry ammunition or demining kits; thirty of them already patrol the Donbass and, a telling detail, lift the morale of the troops, who name them like mascots. Against them, Russia steps up its waves of Shahed: nearly a thousand attacks a month in 2024, more than three thousand five hundred on average over the first five months of 2025. The "swarm" becomes the new unit of fire.

NATO watches and integrates: on April 14, 2025, the Alliance signs a record contract to deploy Palantir's Maven Smart System in its strategic chain of command. The war in Ukraine is no longer a distant event; it is rewriting Western doctrine. Here are five lessons for NATO in the age of AI.

Lesson one: mass yields to agility

Classic doctrine rested on "deterrence capabilities" measured in tanks, artillery tubes, and tons of kerosene. Yet a swarm of cheap micro-drones is now enough to saturate these behemoths. The United States has delivered more than 700 Switchblade 300s: three kilos of explosives, GPS guidance, Starlink link, launched from a frozen trench. On the Russian side, the Shahed suicide drone, at 50,000 € apiece, becomes the target of Ukrainian interceptors mass-printed for ten times less; Moscow nonetheless fires more than a thousand a week in the spring of 2025. The cost-effectiveness ratio, once favorable to heavy tonnage, is reversed: striking "a thousand small blows" ruins a mechanically superior enemy without straining the defender's budget.

Lesson two: the kill chain becomes distributed

Yesterday, the kill chain broke down into three sequences: detect, plan, engage. Today, these links merge in a platoon leader's tablet. Drone video feeds, analyzed by vision models and local LLMs, deliver a trajectory or strike recommendation in seconds. The operator validates with a gesture and the loitering munition sets off. The OODA loop (Observe, Orient, Decide, Act, conceptualized by Colonel John Boyd) contracts: observation and orientation are handled by the algorithm, decision and action execute almost simultaneously. Command centers become synchronization nodes rather than bottlenecks. Any overly hierarchical organization loses the tempo from the outset.

Lesson three: the fog shifts from identification to attribution

Seeing the enemy has never been easier: satellite sensors, open-source data, social networks, and quadcopter drones keep the battlefield under permanent high resolution. The real fog now lodges in the question of who decides and how: does a lethal strike come from an automatic routine? From an allied center? From an isolated operator? How can one prove the presence, or the absence, of a finger on the digital trigger? In this narrative war, the auditability of algorithms becomes a diplomatic weapon.

Lesson four: innovation moves at the pace of software releases

A Leopard tank evolves in ten-year increments; a Ukrainian FPV drone goes from version 1.0 to 1.3 in six weeks. The "Test in Ukraine" label lets a European manufacturer validate a ground drone in three months, with concrete field feedback to show for it. Every offensive tests an update, every retreat triggers a patch. NATO, whose programs still stretch over ten years, must learn to version its arsenals the way one versions code: patches, rollback, and just-in-time updates.

Lesson five: deterrence changes register

Nuclear terror rested on the promise of apocalyptic damage. Yet a software hyperweapon, stealthy, replicable, and cheap, can today inflict energy or economic paralysis without reaching the nuclear threshold. When Palantir adapts Gotham to map Russian infrastructure in real time, it gives Kyiv the ability to thwart, or to threaten, an entire rail network in a few clicks. The credibility of "massive punishment" crumbles: escalation becomes more likely, the red line blurred. From now on we must design an algorithmic deterrence (resting on network resilience, sensor redundancy, and AI transparency) that complements nuclear deterrence without replacing it. The Maven-NATO contract already illustrates this shift: securing the code becomes protecting the territory.

In my book on the geopolitics of AI, Hyperarme (NIV, 2025), I wrote that the world of information had barged into industry the way Tesla had barged into the automobile: by overturning the established order. In Ukraine, the same reversal strikes the military art: steel yields the decisive advantage to the line of code that illuminates, aims, and strikes before the adversary realizes he has already been outmatched. For NATO, the challenge goes beyond buying drones or plugging in new sensors; it is to integrate, iterate, and govern an ecosystem of weapons that learn faster than staffs deliberate. It is this experience, acquired by Ukraine out of sheer necessity, that so interests the United States and pushes them to consider a framework agreement making Ukraine a privileged industrial partner.

The 64-km column did not merely have too little fuel; above all, it was a century behind in software.