Chinese AI companies have long rented data centers in Southeast Asia, Japan, and the Middle East to train models using advanced Nvidia GPUs that cannot be sold directly to China.
Washington is considering restricting remote cloud access; the US House of Representatives has passed the Remote Access Security Act to extend export controls to cloud computing capacity.
Cloud rental prices have also risen sharply: Nebius raised rates by about 30% in June, while Nvidia A100 reached $3.75 per hour in late August.
Southeast Asia could be impacted as regional data center capacity is forecast to grow by 35% annually from 2023 to 2028, compared to 15% globally.
China still relies on Nvidia for advanced model training; ByteDance and Tencent are each reported to have received only about 10,000 H200 chips.
Migrating workloads to domestic chips remains costly, with research estimating that using Huawei Ascend could increase training time and costs by at least 50%.
Meituan stated that LongCat-2.0, with 1.6 trillion parameters, was trained entirely on a cluster of 50,000 domestic computing cards.
Goldman Sachs forecasts that China’s domestic advanced chip shortage could decrease from 92% in 2025 to 34% by 2035.
📌 If the US closes off remote access to Nvidia GPUs via foreign clouds, Chinese AI will lose a crucial lifeline for accessing advanced training capabilities. The impact could ripple across Southeast Asian data centers and Nvidia’s revenues while accelerating China’s transition to Huawei and domestic chips. However, this transition will not be easy: migrating workloads could increase training time and costs by at least 50%, even though domestic advanced chip supply is projected to improve significantly by 2035.
