Diversion and Resale: Estimating Compute Smuggling to China

Key Takeaways

• Estimated 660,000 H100‑equivalent chips smuggled to China by 2025
• Smuggling could represent up to 3% of global AI compute stockpile
• Evidence includes DOJ indictments, NYT, WSJ, and Bloomberg investigations
• Upper bound suggests majority of China’s AI compute may be illicit
• Undetected smuggling likely exceeds reported 300,000‑chip figure

Pulse Analysis

Export controls on advanced AI chips have tightened dramatically since 2022, yet a shadow market has flourished. Early rounds targeted semiconductor equipment and specific entities, but the October 2023 rules broadened the scope to all high‑performance GPUs. This regulatory tightening created powerful incentives for illicit channels, as highlighted by investigative reports from the New York Times, Wall Street Journal and Bloomberg that documented a growing network of shell companies, cloud providers and logistics firms funneling banned hardware into China. The financial stakes are high, with over $1 billion in chips reportedly moved in a single quarter, underscoring the scale of the problem.

Epoch AI’s methodology blends two independent data streams: diversion evidence that tracks chips exiting legitimate supply chains, and resale evidence that maps the gray‑market ecosystem within China. By applying Monte Carlo simulations and adjusting for reporting bias, the firm arrives at a 90 percent confidence interval of 290,000‑1.6 million H100‑equivalents, centering on a median of 660,000 chips. This figure translates to roughly three percent of the global AI compute pool, comparable to the entire compute capacity of leading U.S. labs such as xAI. The upper bound even hints that most of China’s AI compute could be illicit, a stark contrast to earlier CNAS estimates that projected modest smuggling volumes.

The implications are profound for policymakers and industry leaders. Persistent smuggling erodes the deterrent effect of export controls, potentially accelerating China’s AI capabilities while exposing U.S. firms to legal and reputational risk. Strengthening end‑use verification, expanding cooperation with allied customs agencies, and tightening oversight of cloud‑service supply chains could curb the flow. For investors, the report signals heightened geopolitical risk in the AI hardware sector and may influence valuation models for companies dependent on high‑performance GPUs. Understanding the hidden dimensions of chip diversion is now essential for any strategic decision‑making around AI technology and national security.