Agentic AI in the US vs. China: How Do They Compare?

National Interest  |  Quisan Adams, Earl A. Carr Jr.

The United States Department of Defense allocated $200 million each to Anthropic, Google, OpenAI, and xAI in July 2025 to build classified agentic AI workflows. However, Anthropic was blacklisted in early 2026 following disputes over autonomous weapons use, prompting the Pentagon to expand contracts to eight alternative vendors by May 2026.

China is rapidly narrowing the technological gap through knowledge distillation and open-weight models like DeepSeek V4, which offers output tokens at $0.87 per million—up to 34 times cheaper than American counterparts like GPT-5.6 Sol. Distillation works. Chinese military researchers have successfully deployed distilled target-recognition models on People's Liberation Army drones and unmanned vessels. However, joint evaluations by the UK Artificial Intelligence Security Institute and the United States Center for AI Standards and Innovation revealed that Chinese models like Kimi K3 struggle with complex multi-step cyberattacks. The competition now centers on global ecosystem adoption through initiatives like the American AI Exports Program.

Comment

The People's Liberation Army's integration of distilled target-recognition models onto tactical drones and unmanned submarines alters tactical autonomy at the forward edge. By compressing large language models to run directly on low-power onboard chips, Chinese field units bypass the bandwidth bottlenecks that traditionally constrain real-time sensor processing. Command nodes no longer require continuous high-throughput data links to execute target recognition during electronic warfare jamming. Autonomous strike packages retain operational continuity even in denied electromagnetic environments.

This local compute autonomy degrades conventional anti-access/area denial countermeasures aimed at severing enemy command networks. Intercepting or jamming signal relay links between field platforms and central nodes fails to neutralise autonomous target engagement cycles. Consequently, electronic countermeasures against People's Liberation Army autonomous swarms yield diminishing returns compared to direct kinetic interception.

Strategic Question for Discussion

Which operational vulnerability becomes more decisive when People's Liberation Army autonomous swarms rely on distilled edge models: the mathematical limits of distilled reasoning or the physical attrition of edge hardware?

The mathematical boundary of distilled models represents the more immediate operational vulnerability for People's Liberation Army units in high-complexity combat. As evaluations of models like Kimi K3 demonstrate, distilled architectures degrade when confronted with non-linear, multi-step tactical anomalies that exceed their training distributions. Consequently, edge-deployed autonomy remains susceptible to algorithmic disruption long before physical hardware attrition takes effect.

Share your assessment in the comments below.

To read in details click here