Anthropic published evidence in February 2026 detailing how Chinese labs, including DeepSeek, used Claude to generate censorship-safe alternatives and synchronized traffic to train their own models. Tencent researchers also published a June 2026 paper on on-policy distillation techniques, which the author implies are used for cross-model training. Despite distillation, Chinese models require similar GPU inference capacity as US frontier models, and US robots are unlikely to run Chinese models due to embedded biases and backdoors.
Chinese labs can achieve near-frontier model capabilities without equivalent GPU training spend, but inference still requires significant compute, and geopolitical concerns limit US deployment of these models.