In a bold workaround to bypass growing export restrictions on AI hardware and data, a Chinese AI firm reportedly flew 4.8 petabytes of training data to Malaysia—via human couriers.
Four engineers, each carrying 15 high-capacity (80TB) hard drives in their luggage, transported the data to Kuala Lumpur. Upon arrival, they began training large language models (LLMs) using 300 NVIDIA-powered servers rented at a local data center.
Rather than risking surveillance or delays via internet transfer, the team opted for this "sneakernet" method to stay under the radar. The hardware lease was signed via a Singaporean subsidiary—though even Singapore has recently tightened AI oversight.
While inefficient, this physical export method is increasingly seen among AI startups facing chip shortages and cross-border scrutiny. In today’s environment, creativity may be as crucial as compute power.
