Chip Smuggling

2025-12-22 · Source: Papers & Reports | Epoch AI · Field: Finance & Economics — Economic Analysis & Policy, International Business & Trade, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

An analysis estimates that a median of 660,000 Nvidia H100-equivalent AI chips (90% CI: 290,000 to 1.6 million) will have been smuggled into China by the end of 2025, despite stringent US export controls. This figure represents approximately 3% of the global compute stockpile and could account for a quarter of China's legally acquired or domestically produced AI compute. Public allegations and investigative reporting, including indictments against Supermicro employees for diverting \$2.5 billion in Nvidia chips and suspicions surrounding Singaporean cloud provider Megaspeed, indicate nearly 300,000 H100-equivalents have been diverted or are missing. The US Commerce Department implemented export controls on advanced chips in October 2022, tightened them in October 2023, and added licensing requirements for H20 chips in April 2025, intensifying smuggling incentives. The estimation methodology uses Monte Carlo simulations, combining evidence from chip diversion and resale markets within China.

Key takeaway

For policy makers evaluating the efficacy of AI chip export controls, this analysis reveals that current restrictions are being circumvented at a substantial scale. Your strategic assessments must account for an estimated 660,000 H100-equivalent chips smuggled into China by 2025, potentially comprising a quarter of their total AI compute. You should reassess enforcement mechanisms and international cooperation to mitigate this significant leakage and ensure controls achieve their intended national security objectives.

Key insights

US export controls have fueled a significant, complex AI chip smuggling operation into China.

Principles

• Export controls intensify illicit trade incentives.
• Supply chain vulnerabilities enable large-scale diversion.
• Undetected smuggling significantly skews official figures.

Method

Monte Carlo simulations integrate diversion allegations and gray market resale data, adjusting for detection rates and reporting accuracy, to estimate total smuggled compute.

In practice

• Implement robust supply chain monitoring.
• Analyze gray market activity for smuggling trends.
• Employ dual-source modeling for illicit flow estimation.

Topics

• AI Chip Smuggling
• Export Controls
• NVIDIA H100
• Compute Capacity
• Supply Chain Diversion
• Monte Carlo Simulation

Best for: CTO, VP of Engineering/Data, Executive, Policy Maker, Director of AI/ML, Investor

Related on AIssential

• Diversion and resale: estimating compute smuggling to China — US export controls have driven substantial, difficult-to-quantify AI chip smuggling to China, impacting global compute distribution.
• China Plans to Let Top AI Firms Buy Limited Amount of Nvidia H200 Chips — China will allow limited Nvidia H200 chip purchases, balancing AI demand with domestic chipmaker protection.
• “Thank You, Stupid Trump” — Why China Is Quietly Grateful for the Chip Ban — Export controls on chips inadvertently fostered China's domestic AI industry, creating a captive market and accelerating self-sufficiency.
• China boosts 'AI for Science' computing in 2 months without US chips - Business Standard — China rapidly deployed a large, domestically-chipped AI computing cluster for scientific research, reducing reliance on US technology.
• Trump’s China Visit Puts AI Chips and Trade Back in Focus — AI chips are now a strategically vital global resource, akin to oil, influencing geopolitical and economic policies.
• DeepSeek’s AI chip plans — Chinese AI firms are developing in-house chips to reduce foreign reliance and compete domestically amid US export controls.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Papers & Reports | Epoch AI.