China's Tech Giants Take AI Model Training Offshore to Tap Nvidia Chips (www.ft.com)

China’s biggest internet and cloud firms are increasingly moving large-scale model training offshore to countries and jurisdictions that still have access to Nvidia’s top-tier GPUs, after U.S. export controls limited shipments of the newest accelerator chips into mainland China. Rather than wait for domestic silicon to catch up, companies are using foreign data centers, cloud accounts, and overseas subsidiaries to rent clusters of H100/A100-class GPUs (and regional variants) for multi-node training runs, or staging training jobs remotely over high-bandwidth links. Technically, this trend underscores how dependent modern LLM and vision-model scaling is on high-memory, NVLink-connected GPU clusters and low-latency fabrics—resources that are hard to replicate quickly with homegrown chips. The move raises practical trade-offs: higher costs, data-sovereignty and compliance headaches, and extra latency, but it preserves R&D momentum for teams that need thousands of GPU cards for model/data/model-parallel training. It also accelerates parallel responses: more investment in domestic accelerators, and faster uptake of efficiency techniques (quantization, pruning, LoRA, ZeRO-style optimizer/sharding) to reduce GPU-hours. Strategically, this workaround blunts some intended effects of export controls and highlights how compute demand, cloud economics, and distributed training tooling shape geopolitical leverage in AI.