What Follows as Drones Take Over the Final Attack

This piece follows one change. Onboard AI is taking over the last phase of a drone strike from human pilots, and that shift is happening through procurement tests ahead of any policy decision.

Key points

• Jam-proof tethers and AI terminal homing are reducing the human role to choosing a target before launch¹².
• The US is about to buy this pattern in bulk: the autonomy group's request rises from $226M to as much as $54.6B³.
• At current interception volumes, defense is moving to automated engagement, and NATO's eastern border is adopting the same approach⁴⁵.
• The rules are weakening from both sides. The US directive is being rewritten, and the treaty track may split away from the major drone powers⁶⁷⁸.
• We put the odds at about 70% that by 2028, "human judgment" will in practice mean pre-launch target selection.

The change

Brave1, Ukraine's state defense-tech cluster, reports that successful strikes using AI terminal auto-homing rose tenfold between January and September 2026. Six of seven vendors passed Ministry of Defence tests against moving armor and are recommended for procurement². Most of more than 10,000 Hornet loitering-munition strikes used AI guidance². Electronic warfare is driving the change. Fiber-optic tethers made jamming useless at the front, and each new round of countermeasures pushes both sides toward full onboard autonomy as the next way around it¹. As our signal scan found, test ranges and contracts are settling how much autonomy is acceptable before doctrine does.

What follows

First-order

• Procurement specs become doctrine. The Pentagon's Drone Dominance Program had ordered 20,000 FPV drones from 10 vendors by June through fixed-price "Gauntlet" competitions⁹. The autonomy group is buying $500M of systems tested in combat in Ukraine³. Whatever passes a Ukrainian test is likely to become the American requirement.
• Defense runs at machine speed. Ukrainian interceptors reportedly destroyed 33,000 Russian UAVs in a single month⁴. At that volume, detection and engagement have to be automated, and the human mostly sets the rules for the loop.
• Chokepoints move. Cable spools are now a strategic input¹. China's controls on rare-earth magnets cap Western drone output, while US rules now restrict foreign drone hardware as a whole category¹⁰.

Second-order

• US policy codifies what is already in service. NSPM-11 ordered a rewrite of Directive 3000.09 to allow fast deployment and commercial and open-source AI. As of October, no revised text has been published⁶. We think the rewrite will most plausibly legitimize terminal autonomy already being bought, without setting a new limit.
• Diplomacy splits. Seventy-six states back treaty negotiations, but the CCW decides by consensus, and its Review Conference meets 16 to 20 November⁷. At REAIM, only about 35 of 85 states signed, and the US and China were not among them⁸. We put roughly 60% on the CCW granting no negotiating mandate. In that case the likeliest result is a treaty outside the CCW that the main drone powers never join.
• Model supply becomes weapons supply. Chip export controls are tightening¹¹, but about 70% of new derivative open models are built on Qwen¹², and the PLA has reportedly put DeepSeek into swarm programs¹³. Once targeting runs on open weights, controls on chips cover less of the risk. That is the shared supply chain our convergences piece described.

Third-order

• Escalation risk moves up the stack. A PLA test claims one operator directed more than 200 drones under jamming¹³, which would rule out meaningful human control of each engagement. In simulated crises, frontier language models crossed the nuclear threshold in 95% of games¹⁴. Those are the same commercial models defense users are being encouraged to adopt.
• Attribution blurs at borders. Russia gave a deniable explanation for the drones that entered Poland⁵. Faced with probes like that, automated counter-drone defenses raise the risk of misidentification, and nobody has defined a clear trigger for responding. The US-China AI "incident line" is still at an early stage and depends on trade talks holding up¹⁵.
• The shadow norm. If the pre-launch click counts as human judgment, that becomes the global standard by default². No state will have argued for it. As our contradictions piece noted, control and scale pull against each other, and scale is currently ahead.

Who gains, who loses

• Gains: drone startups that can deliver fast at fixed prices⁹³, and Ukraine's industry, whose test results are becoming other countries' purchase specs².
• Gains: whoever controls inputs that are hard to replace, such as magnets, motors and fiber cable¹¹⁰.
• Loses: electronic-warfare incumbents whose jammers lose value¹, and makers of expensive missile interceptors, which are being pushed out of counter-drone work⁴.
• Loses: middle powers that rely on soft-law norms, which carry less weight without the leading AI states⁸.
• Loses: consensus-based arms control, which the major drone powers can block⁷.

A glimpse ahead

Gauntlet Phase III results, spring 2028. Eleven vendors qualify. The scoring sheet weights terminal-homing hit rate at 40%, cost per unit at 30% and spool-free range at 20%. The last 10% is labeled "operator workload." No line on the sheet asks who decides once the drone is in the air. A footnote cites the revised Directive 3000.09, published 14 months late, as "satisfied by pre-launch target designation." That week in Geneva, 82 states sign a treaty text on autonomous weapons. None of them builds drones at the scale on that scoring sheet.