Repair demand for Nvidia’s high-end AI GPUs notably the H100 and A100, whose exports have been banned to China since September 2022 under U.S. restrictions is undergoing a sharp surge across Shenzhen and beyond
- At least a dozen boutique firms in Shenzhen now specialize in servicing these forbidden chips, handling smuggled stock inside local repair centers
- A leading workshop reports repairing up to 500 AI chips per month, with diagnostic, memory/PCB repair, fan fixes, and full testing (including environments simulating data center conditions
Repair economics and demand dynamics
- Repair fees typically range from ¥10,000–¥20,000 (~US $1,400–$2,800) depending on damage complexity
- One firm charges roughly 10% of the original selling price, servicing around 200 GPUs/month, while others operate larger scale operations with capacity to host 256 servers for testing
- This booming repair market stems from the aging installed base of H100/A100 GPUs many already 2–5 years old and operating under continuous heavy use
Smuggling routes adapt, new GPU generations enter the fray
- The spike in repair work underscores the continued prevalence of smuggled Nvidia chipsets in China, despite export controls
- Demand has shifted toward newer, higher-tier chips such as the B200 and forthcoming B300, with full eight‑GPU server racks commanding prices over ¥3 million (~US $420,000–$490,000) in the informal market
Legal ambiguity and emerging legislative concerns
- While buying, selling, and repairing these Nvidia GPUs is legal under Chinese law, official support or servicing is impossible Nvidia does not support restricted models in-country
- To address unauthorized use, both Republican and Democratic lawmakers in the U.S. have proposed legislation to track the location of chipsets post‑sale, seeking accountability if banned hardware enters unauthorized channels.
Why H20 chips aren’t cutting it
- Nvidia has been permitted to resume sales of H20 GPUs to China under revised and compliant export licensing, targeting inference applications rather than high-end training workloads
- However, H20-based servers (often housing eight GPUs) can cost over ¥1 million (~US $139,000) and the chip’s performance is significantly inferior to that of banned models like H100, making it less attractive for training large language models or advanced AI systems..
Outlook
- The illegal but persistent flow of advanced Nvidia GPUs into China, combined with robust repair demand, reveals both beijing’s continuing reliance on U.S. AI hardware and the limitations of export controls.
- Unless major policy shifts or enforcement capabilities emerge, the underground market ranging from smuggling to covert repairs looks poised to remain central to China’s AI training infrastructure.
- Meanwhile, U.S. lawmakers are pressing for new mechanisms to trace GPU destinations, and regulators may soon face pressure to clamp down on transshipment routes through Southeast Asia.