AI Defense Push Collides With Chip Stock Swings and Trade Fears

Chinese military researchers have turned to American artificial intelligence systems to accelerate development of their own battlefield tools, according to a new analysis of academic papers. The practice, known as model distillation, allows teams to train smaller, specialized AI by drawing on the reasoning patterns of larger Western models. While the technique is common in commercial labs, its appearance in defense-related Chinese studies has drawn fresh attention in Washington.

A review by the Jamestown Foundation examined more than sixty papers and found repeated instances in which institutions tied to the People’s Liberation Army used distillation to build capabilities for surveillance, cyber operations, content monitoring and autonomous battlefield functions. Sunny Cheung, the fellow who led the analysis, noted that the papers show researchers attempting to transfer not just answers but the underlying logic from advanced models. That step is technically harder and raises questions about how export controls on AI hardware and software are holding up.

The findings arrive as the United States continues to tighten restrictions on advanced chip exports to China. Officials argue that distillation can extract useful performance even when direct access to the newest hardware is blocked. Chinese authorities have dismissed the concerns, saying Washington is simply trying to preserve its lead. The papers themselves do not reveal classified data, but they illustrate a systematic effort to close capability gaps across several military domains.

At the same time, American defense startups are racing in the opposite direction, embedding AI directly into new weapons platforms. Anduril, founded by Palmer Luckey, is among the most prominent. Its chief executive, Brian Schimpf, has argued that fully autonomous weapons without human oversight are neither desirable nor under discussion inside the military. Existing systems such as certain torpedoes and anti-radiation missiles already make targeting decisions without continuous human input, he noted in a recent interview. The goal, Schimpf said, is to give operators better information and faster options rather than to remove judgment from the loop.

Schimpf framed the alternative as accepting “dumber weapons,” which he called ethically inferior. He stressed that Anduril’s systems are designed to keep humans in meaningful control while shifting routine analysis to algorithms. The company’s approach reflects a broader push by newer defense firms to shorten development cycles and incorporate commercial AI advances that legacy contractors have been slower to adopt.

The contrast between the two trends is stark. One side is repurposing foreign models through academic channels to build specialized military tools; the other is racing to integrate original AI into platforms that could be fielded within years rather than decades. Both developments underscore how quickly the technology is moving and how difficult it has become to wall off foundational capabilities. For policymakers, the immediate challenge is updating export rules and verification methods so that distillation does not simply become a workaround. For technologists on both sides, the race is already underway to determine whose systems will set the standard for the next generation of military applications.