Chinese users are resorting to unauthorized repairs of NVIDIA chips due to restrictions limiting official support access.

A quiet but growing business has emerged in China focused on repairing Nvidia's advanced AI chips, despite strict US export controls. According to reports from Reuters, around a dozen small firms, mainly based in Shenzhen, claim they service large numbers of Nvidia's H100 and A100 GPUs, even though these chips were officially banned from sale to China in 2022. The scale of this operation is substantial: one company told Reuters it repairs up to 500 Nvidia AI chips every month, and with roughly 12 similar firms operating year-round, this could amount to tens of thousands of chips annually. Many of these units are worn down from heavy use, having run around the clock for years in AI training workloads. As one co-owner of a Shenzhen firm that moved into AI hardware in late 2024 stated, "There is really significant repair demand."

The repair operations have developed sophisticated infrastructure to handle this demand. One company created a second business purely to handle AI chip repair, operating a facility that can simulate data center conditions with up to 256 servers. Another shop that shifted from GPU rentals to repairs told Reuters it fixes around 200 chips per month, typically charging about 10% of the original purchase price. Repairs can include fan replacement, circuit board fixes, memory diagnostics, and software testing.

Nvidia's position on these operations is unambiguous: the company cannot legally support or replace restricted GPUs within China. An Nvidia spokesperson stated that only the company and approved partners are authorized to offer the necessary service and support, adding that running restricted chips without full infrastructure is not viable long-term. The existence of such a repair sector is fundamentally tied to the widespread smuggling of banned chips into China. While Nvidia has begun offering the H20 GPU in China to comply with export restrictions, many customers there still prefer the banned H100 for training large language models.

The long-term implications of this situation remain uncertain. The potentially high failure rate raises significant concerns about what will happen to tens of thousands of aging A100s and earlier GPUs once they fail beyond repair.