Autonomous Warfare Now Shares Its Supply Chain With Civilian AI

Autonomous warfare is no longer a self-contained military story. Its pace is now set by three civilian currents: the US-China decoupling of chips and hardware, the spread of commercial and open-weight AI models into defense use, and the manufacturing base that decides who can build drones by the hundred thousand.

Key points

• Drone hardware bans, chip export controls and the humanoid robot ban are one decoupling, and each side now holds a chokepoint it can pull in a crisis.¹²³
• The White House rewrite of the Pentagon's autonomy directive opens the door to commercial and open-source models at the moment Chinese open-weight models carry about 61% of routed tokens and frontier models show measured escalation biases.⁴⁵⁶
• Mass drone procurement in the US is already below its own targets, while Asia installs 74% of the world's industrial robots; production capacity, not software, is the binding limit.⁷⁸⁹
• The likely result is two autonomous-weapons blocs with separate hardware, separate models and no shared oversight, which makes the November CCW decision harder, not easier.¹⁰¹¹

Where the lines cross

Drone warfare and compute decoupling

The drone supply chain is splitting on the same lines as the chip supply chain. The FCC moved in December 2025 from blacklisting DJI and Autel to blocking foreign drone hardware as a category, and in July 2026 proposed banning 'military-grade' foreign drones outright, while China answered with rare-earth magnet controls.¹ In the same months the US closed chip-control loopholes on subsidiaries and cloud access, and DeepSeek ordered 160,000 Huawei Ascend chips for a sovereign training stack.²¹¹ The FCC's ban on foreign humanoids and quadrupeds completes the set.³

What this makes possible is a fully bloc-separated autonomy stack: Chinese drones on Chinese chips running Chinese models, and a Western equivalent that costs more and ships slower. We think this is now the base case, about 70% likely by 2030. The risk is that each bloc holds a lever the other needs. China controls magnets and motors; the US controls advanced lithography. In a Taiwan crisis, either could be pulled before a shot is fired.

Military autonomy policy and commercial AI models

NSPM-11 ordered the Pentagon to rewrite Directive 3000.09 in 90 days to allow commercial and open-source AI in weapons.⁴ The commercial landscape it opens onto has two features defense planners did not choose. First, Chinese open-weight models account for about 61% of tokens on the largest neutral router, and about 70% of new derivative models are built on Qwen.⁶ Second, Payne's pre-print found three frontier models crossed the nuclear threshold in 95% of simulated crises and never chose to accommodate.⁵

Put these together and the fastest path to a 'commercial AI' decision-support tool runs through a model lineage shaped by PRC content rules or through a Western model with a measured bias toward escalation. Neither is disqualifying on its own. The convergence is that the policy that lowers the review bar is arriving before anyone has a test for either problem. We covered the open question of what 'appropriate human judgment' still means in the contradictions piece; this is the mechanism that could settle it by default.

Attritable mass and the factory floor

The Pentagon's Drone Dominance Program wants 340,000 small drones in two years, and its first phase came in under the 30,000-unit target.⁷ DAWG's request jumped from $226 million to up to $54.6 billion.⁸ Money is not the constraint. Ukraine's interceptor output doubled in four months, and that is a wartime economy with 25 domestic models.¹²

The civilian comparison is stark. Asia takes 74% of the 542,000 industrial robots installed each year, and the US installed 38,000.⁹ Amazon's million-robot fleet shows what robot-native production looks like at scale, but it is a logistics design, not a drone line.¹³ The PLA's 200-drone single-operator swarm, if the state-media claim holds, is a software achievement sitting on a hardware base the US does not have.¹⁴ The plausible outcome is that US autonomy doctrine gets written for a fleet size it cannot yet build.

What the overlap opens up

Together these three convergences point to a world where the rules of autonomous warfare are set by procurement tests, model licenses and factory capacity rather than by the November CCW vote, which the major drone powers can block anyway.¹⁰ The opportunity is real: a Western bloc that solved its drone manufacturing gap would gain a deterrent that costs thousands rather than millions per unit. The risk neither field shows alone is a bloc-split AI ecosystem with no common test for what a weapon's model will do under pressure, and two chokepoints that make the first move in a crisis an industrial one. The signal scan showed autonomy outrunning its rules; the convergences show why the rules are losing the race.