The $100 Billion Loophole in U.S. AI Export Controls

The core of this loophole is a legal mismatch: U.S. export controls aggressively restrict advanced hardware (semiconductor chips) from reaching China, but they do not equivalently restrict access to AI models themselves — and the Pentagon's blacklist only bars U.S. defense contracting, not commercial software sales. Here is how the gaps break down.

The central loophole: hardware vs. model access

The U.S. Bureau of Industry and Security (BIS) has authority to control advanced semiconductor chips as "dual-use items," but it lacks statutory authority to impose direct controls on AI models or model weights . Under the current framework, BIS can regulate chips used to train AI models but cannot directly restrict the distribution of the trained models themselves . This means a U.S. AI company can legally sell API access or deploy frontier models to a foreign subsidiary — even if that subsidiary is owned by a blacklisted Chinese parent — as long as no controlled chips are physically transferred.

The FT-reported pipeline: Singapore subsidiaries as conduits

A Financial Times investigation (corroborated by multiple industry reports) found that OpenAI and Google have both established applied AI labs in Singapore, and that Singapore has become a neutral hub where U.S. and Chinese tech firms operate side-by-side . Chinese companies including Alibaba affiliates, Tencent, and Ant Financial have been able to access advanced U.S. AI models — including OpenAI's GPT models and Anthropic's Claude — through Singapore-based subsidiaries and intermediary arrangements . In September 2025, Anthropic announced it would halt sales to majority Chinese-owned groups after acknowledging that Chinese entities had gained unauthorized access to Claude through these structures . OpenAI and Google have likewise begun cooperating through the Frontier Model Forum to combat model extraction by Chinese competitors, but the cooperation is voluntary and reactive, not mandated by export law .

The Pentagon blacklist: broad but narrow in practical effect

On June 8, 2026, the Pentagon published its largest-ever expansion of the Section 1260H "Chinese military companies" list, increasing it from roughly 134 to 188 entities . The additions swept in major civilian technology companies: Alibaba, Baidu, BYD, NIO, CXMT, YMTC, Unitree Robotics, WuXi AppTec, and RoboSense .

Critical distinction: The Pentagon list bars these companies from U.S. defense procurement contracts and triggered a ban on June 30, 2026 barring DoD from contracting with any lobbyist who also represents blacklisted firms . It does not impose any restriction on commercial sales of AI software, API access, or cloud-based model inference. A Singapore subsidiary of Alibaba can still legally purchase OpenAI API credits under current law, because the Pentagon list governs government contracting, not export-controlled items.

Alibaba's legal challenge and the list's vulnerability

Listed companies can petition the Pentagon for removal . Alibaba, Baidu, and BYD have all publicly denied being military companies . The Department of Defense has faced prior legal challenges over the 1260H list — it published and then withdrew an earlier expanded version in February 2026 following industry pushback . Alibaba separately filed a lawsuit to be removed from the list, arguing the designation is inaccurate and commercially damaging. The litigation exposes that the blacklist, while politically potent, is legally contested and has limited reach over commercial AI transactions.

The chip loophole (largely closed, but only for hardware)

Until late May 2026, a parallel subsidiary loophole existed for chips: hundreds of thousands of advanced Nvidia AI servers reached Chinese entities through subsidiaries in Singapore, Malaysia, and the UAE . On May 31, 2026, BIS issued guidance confirming that advanced AI chip export licenses are required for any entity whose ultimate parent is headquartered in China, regardless of subsidiary location . The Commerce Department acknowledged the prior enforcement failure . This closure affects hardware only — it does not extend to AI model access.

Where the policy gap remains

What is restricted	What is not restricted
Advanced semiconductor chip exports to China or China-parented subsidiaries	API access or inference of U.S. AI models by China-parented subsidiaries
Defense contracts with Section 1260H-listed firms	Commercial AI software sales to those same firms
Hardware for training frontier models	Deployment of already-trained models via cloud
Chip transfers through Singapore/Malaysia/UAE intermediaries (now closed)	Model access through Singapore-based subsidiaries

Critics argue the net effect is that Washington has built a high wall around compute (chips, data center hardware) while leaving the software gate wide open. Chinese AI developers can still use the best U.S. frontier models through legally registered subsidiaries in Singapore, extracting valuable inference data and model capabilities that feed Beijing's AI development — all while parent companies are officially labeled as Chinese military affiliates .